On the 30th of September I'm doing three sessions on the show floor at CHS Manchester. By the end of the day I'll have built a piece of working software, live, in front of whoever's in the room.

Here's the part that keeps me up: I don't get to choose what it is. You do.

If you want to be part of that, the single most useful thing you can do before the day is fill in The Event AI Behaviour Study. It takes a few minutes and it feeds directly into session one.

How it works

Session one. You tell me what's broken.

Not in a workshop-y way. There's a QR code, you scan it, and you type in the thing that slows you down. The task you rebuild from scratch every week. The system that doesn't talk to the other system. The bit of Tuesday afternoon you'd pay someone else to do.

It goes up on screen, anonymously, and groups itself into themes as it comes in. Then the room votes.

Somewhere in the middle of that I'll share early findings from The Event AI Behaviour Study, which we've been running to work out how event professionals are actually using AI rather than how conference panels say they are. Some of it is not what I expected.

Whatever wins the vote, I start building in the break.

Session two. I come back with something rough.

It won't be finished and it won't be pretty. That's deliberate. Everyone shows you the polished end result and nobody shows you the middle, which is where all the useful information is.

You get to pull it apart. What's missing, what's wrong, whether it solves the thing you actually asked for.

Then I'll give you the method. There's a prompt I use with clients that I'm giving away in that session, because the questions were never the valuable part. It flips the usual thing on its head: instead of you interrogating AI, it interviews you. About your week, your repeated work, the parts of the job you avoid. Fifteen minutes later you've got a map of where your time is genuinely going, which is almost never where you think.

Session three. The finished thing, and how it was made.

I'll build a piece of it again from scratch, on screen, so you can see there's no sleight of hand. Plain English in, working software out. No code, no setup, nothing you can't open tonight.

Then we're launching the CHS Innovation Hub with the team at CHS. A free place for the industry to get practical AI tools, learn to build their own, and put forward the problems worth solving next. What gets built on it starts with what the room decides on the day.

Why I'm doing it this way

I've sat through a lot of AI talks in this industry. Most of them are the same talk. Someone shows you what's theoretically possible, everyone nods, and precisely nothing changes on Monday.

The gap isn't knowledge. Almost everyone in events has used ChatGPT by now. The gap is that people don't know which of their problems is worth solving, and they don't know how to describe it once they've found it.

That's it. That's the whole bottleneck.

The tools have got extraordinarily good at building things. They're still completely dependent on someone being able to say clearly what's wrong. Which is why session one is people typing problems into their phones rather than me talking at them, and why the prompt I'm giving away is about finding problems rather than solving them.

Nobody in that room needs to become a developer. They need to get better at describing what's broken.

Before the day

Two things that would genuinely help.

• Fill in the study. It takes a few minutes and the more responses we get before the 30th, the more I've actually got to say. It's at research.zoby.ai.
• Come with something. Not a polished problem statement. Just the thing that irritates you at your desk. Bring the frustration and we'll find the problem underneath it.

And if you can only make one session, make it the first. That's where the room decides what gets built, and it's the one that makes the other two make sense.

The details

CHS Manchester, AO Arena, 30 September 2026. Three sessions, twenty minutes each, on the show floor. Times to be confirmed, check the CHS programme on the day.

I'm Ed Richards. I run Zoby, where we work out where event businesses are losing time and then build the things that give it back. Before that I spent years in the industry doing the repetitive work I now automate, which is the only qualification that's ever mattered.

Come and give me a problem.

Voice went from novelty to infrastructure this year

Back in May, OpenAI launched three new Realtime API voice models in one go, including one built for reasoning-level conversation and another handling translation across more than seventy languages. The same month, ElevenLabs closed a $500m Series D at an $11bn valuation to expand its enterprise voice-agent platform. Neither of those was a demo. Both were infrastructure moves, the kind vendors make when they expect the thing to be used constantly, at scale, by businesses that are not tech companies.

Events is one of those businesses. A venue enquiry line, an agency's out-of-hours delegate helpline, a supplier's order desk, all of them are phone-shaped problems that a convincing voice agent can now plausibly take on. That is worth being specific about, because "AI can answer the phone now" covers a genuinely useful capability and a genuinely dangerous one, and they look identical from the outside.

Half one: what a voice agent is actually good for

An enquiry line or a delegate helpline spends most of its time on a small number of repeatable jobs: taking down the basics of an enquiry, checking availability against a known calendar, answering the same handful of FAQs, and booking a callback with a human when the question is bigger than that. A voice agent that sounds natural is genuinely good at exactly that list.

• Capture. Name, date, headcount, budget range, the shape of the enquiry, taken down accurately so a human is not starting cold on the callback.
• Qualification. A first pass at whether the enquiry fits what you actually do, so the human callback is a conversation with someone worth having a conversation with.
• FAQs. Opening hours, parking, accessibility, what is and is not included, the questions your team answers identically forty times a week.
• Translation. At a multilingual conference or exhibition, a voice agent that can hold a competent conversation in a delegate's own language, at 11pm when nobody bilingual is on shift, is a real improvement on a hold message.

All four of those share a property: getting them slightly wrong is annoying, not dangerous. A misheard headcount gets corrected on the callback. A slightly clunky translation still gets the delegate to the right answer eventually.

What never to hand to a voice agent

Two categories are worth ruling out entirely, not softening, ruling out.

• Anything with money. Taking a payment, confirming a refund, changing an invoice, agreeing a price outside a published rate card. A voice agent should capture the request and route it to a human, never complete it.
• A distressed delegate. Someone who is lost, unwell, anxious, or dealing with an emergency at your event needs a person on the line within a small number of seconds, not a competent-sounding agent working through a script. Build the escalation trigger for distress in early and test it, do not assume the agent will recognise it on its own.

Half two: the same technology, pointed the other way

Back in January, a Swiss businessman was reported to have transferred several million Swiss francs after a series of calls from what he believed was a trusted business partner. It was an AI voice clone. Not a single suspicious call, a series of them, convincing enough across multiple conversations to move real money.

That story is not really about Switzerland or about millionaires. It is about the fact that voice cloning good enough to sustain a conversation, not just a ten second clip, is now ordinary infrastructure, available to the same degree to anyone with bad intentions as it is to a legitimate business building a helpline. Agencies that move large supplier payments fast, which is most of them during peak event season, are a natural target. "The MD, on the phone, asking finance to pay the new supplier account today, it's urgent, I'm in a meeting and can't email" is no longer a hypothetical. It is a phone call a cloned voice can now make convincingly.

An events agency in the south east already treats this as a live risk rather than a theoretical one. Its finance lead has a standing rule that no supplier detail changes on a phone call alone, full stop, regardless of how senior the caller sounds or how urgent the request feels. That rule predates most of the voice AI news above. It is simply more obviously necessary now than it was a year ago.

The callback-and-codeword rule

The fix does not need new software. It needs a rule that survives urgency, because urgency is the entire attack.

• Any request to pay a new supplier, change existing bank details, or move a payment outside the normal approval chain, made by phone, gets a callback to a known number before it happens, never a number given during the call itself.
• Agree a codeword or a simple verification question in advance with anyone senior enough to plausibly make an urgent payment request by phone, something a voice clone working from public information would not know.
• Make the rule apply to everyone, including the MD, especially the MD. The whole point of the attack is impersonating authority to bypass a process. A rule with an exception for authority has no rule in it.

A prompt to build the escalation rules

If you are speccing a voice agent for an enquiry line or helpline, this is a useful first pass at the escalation logic before you brief a vendor. Paste it into your AI assistant with a description of your line's current volume and typical calls, and expect back a draft set of rules for what the agent handles, what it escalates immediately, and what phrasing it should never use to imply it is human.

I am briefing an AI voice agent for our events business phone line (describe: venue enquiry line, agency delegate helpline, supplier order desk). Help me draft the escalation rules before I brief a vendor.

Here is what the line currently handles:
[describe typical call volume, the most common enquiry types, and roughly how many calls a human currently takes out of hours]

Draft me:
1. A list of call types the agent can handle end to end (capture, qualify, answer FAQ, book callback).
2. A list of call types that must escalate to a human immediately, including anything involving payment, account changes, or a caller who sounds distressed or confused.
3. Wording the agent should use at the start of the call to make clear it is automated, in plain language, not buried in a disclaimer.
4. A short brief for the vendor on what "escalate immediately" should trigger technically, phrased so a non-technical person could check the vendor has actually built it.

Keep this to one page, practical, no jargon.

This week

Two separate jobs, both small. If your enquiry line already runs or is about to run a voice agent, check the four things above are what it actually does, and that anything involving money or a distressed caller routes to a person. Separately, whether or not you touch voice AI at all, put the callback-and-codeword rule in front of your finance team this week. It costs nothing and it closes a door that is now easier to walk through than it used to be. I will be talking through more of this at CHS Manchester later this month, but the rule itself does not need to wait until then.

Two numbers worth sitting with

Deloitte's State of AI in the Enterprise report landed last month, and one split in it is worth pinning to a wall. Organisations fall into roughly three groups: thirty seven per cent are using AI at the surface, drafting and summarising and light assistance. Thirty per cent have gone further and redesigned a process around it. Thirty four per cent have made deeper business model changes. Sixty six per cent report real productivity gains, whichever layer they're on.

Put that next to the ONS figure from this summer: twenty nine per cent of UK businesses now use at least one AI technology, up from twenty one per cent a year earlier, rising to forty nine per cent among firms with two hundred and fifty or more staff.

Read those together and the picture for events businesses is fairly clear. Most agencies, venues and suppliers are small, which the ONS data suggests puts you well below that forty nine per cent figure. And of the businesses using AI at all, the Deloitte split says you're most likely sitting in that thirty seven per cent surface layer group.

None of that should feel alarming. It's simply where the average small business sits right now, events or otherwise. What matters is what you do with that information, not how it makes you feel reading it over coffee.

That is a completely reasonable place to be

I want to say this plainly, because it's easy to read numbers like these and feel like you're failing at something. You're not. Drafting proposal copy faster, summarising a client call, tidying an agenda: that's a legitimate, useful first layer of AI use, and it's where almost everyone starts.

The problem isn't being at the surface. The problem is not knowing you're at the surface, and assuming that's as far as this goes.

The step that actually matters doesn't need new AI

The move from surface use to process redesign, the thirty per cent group, is the one that changes what a business actually looks like day to day. And here's the part that surprises people: it usually doesn't require a new tool.

It requires the process getting written down. Most events businesses don't have their quoting process, their handover from sales to delivery, or their post event follow up documented anywhere. It lives in people's heads, slightly differently in each head, and gets reconstructed from memory every single time.

You cannot redesign a process that has never been written down in the first place. AI can speed up a documented process enormously. It cannot speed up a process that doesn't exist yet as anything other than habit.

This is the part that gets skipped, because writing down a process is unglamorous work, and it doesn't feel like progress the way installing a new tool does. But it's the actual bottleneck for most of the businesses I meet. The AI conversation gets all the attention. The documentation conversation is where the real gains sit waiting, mostly untouched.

Which layer are we actually on

A quick, honest self-test, three questions, no scoring needed, you'll know the answer as soon as you ask them.

• Is AI use in our business individual, something people picked up themselves, or is it built into how we actually run a job from brief to delivery.
• If our best person left tomorrow, would the way we do things leave with them, or is it written down somewhere anyone could follow.
• Have we changed a process because of what AI makes possible, or are we just doing the same process slightly faster.

If the honest answers are individual, it would leave with them, and slightly faster, you're at the surface. That's not a criticism. It's a starting point.

What moving up a layer actually looks like

For a venue sales team, surface use looks like drafting proposal responses faster. Process redesign looks like a documented, consistent way of qualifying an enquiry, checking capacity, and handing a confirmed booking to operations, with AI doing the drafting inside that structure rather than instead of it.

For a small agency, surface use looks like summarising client calls. Process redesign looks like every brief following the same format regardless of who took the call, so a producer picking up a project midway through isn't reconstructing context from three different people's memories.

Neither example needed a new AI subscription. Both needed someone to sit down and write the process out properly, once.

That's usually a week of focused work, not a quarter long project. It's also the kind of work that's easy to keep putting off indefinitely, because the business keeps running perfectly well without it, right up until the day it doesn't.

Where this goes next

This is exactly the ground we'll be covering at CHS Manchester at the end of the month, one sentence on it here and the rest on stage: what actually separates surface use from structural change in this industry, and why most of the gap has nothing to do with which AI tool you bought.

If you want to work out which layer your own business is genuinely on before then, that's the first conversation in a Discovery Lab.

A question I keep hearing answered the same way

Ask anyone in events whether they use AI and the answer is yes, almost every time now. Ask them how, and you get a shrug and "you know, emails, the odd proposal".

I believe them. I also think "the odd proposal" is doing a lot of work in that sentence.

Every conference panel this year has had a view on how event professionals are using AI. Nobody has had numbers. Not survey numbers, where people describe the version of themselves they would like to be, but numbers from watching people actually do a job with an assistant sitting next to them. That bothered me for about a year before I did anything about it.

So we built something instead of asking another question

You cannot ask someone to describe how they work with AI and trust the answer. Not because people lie, but because nobody remembers how they phrased a request on a Thursday afternoon with three other things open. The moments that stick are the interesting ones. The ordinary ones, which are the ones that matter, never register.

So we did not build a survey. We built a small planning job, the kind of thing you would hand an AI assistant in a normal week, and made it something you can do in about three minutes on your phone.

You talk to it, or type if you would rather, work through the job together the way you would on a real day, and at the end you get a card. An archetype, one of five, collectible, top trumps style, the sort of thing you look at and immediately want to know which one your colleague got.

It is genuinely fun. That was not decoration bolted onto something serious. It is just a better way to get an honest three minutes out of someone than a form ever was.

What it actually is, for anyone who likes to know before they click

• It is anonymous.
• It takes about three minutes.
• It works on your phone, standing up, on a break, wherever you are when you read this.
• You do a short planning job with an AI assistant, the way you normally would, and you get your card at the end.

That is the whole of it. The interesting part is not the mechanics. The interesting part is what we learn once enough people have played it honestly, and that is the bit I will share later, properly, with the numbers behind it rather than a hunch.

What the card says about you

The five archetypes are about how you work with an assistant. How much you hand over and how much you keep. How you ask for things. Whether you treat it like a colleague, a tool, or a slightly unreliable intern. None of them is the right answer, and each has a stat or two the others would quite like to have.

People are already arguing about which one their team lead would get. That is roughly the reaction I was hoping for.

The one instruction that actually matters

Play it the way you would actually work, not the way you would want to be seen working. Ask for things the way you ask for them. Hand over what you would normally hand over. Do not perform anything for a card.

That is the whole design decision behind making it short and playful: honesty in, honesty out. Anything else and the card is decorative and the findings mean nothing.

Where this goes next

Early findings get their first public airing at CHS Manchester on 30 September, where I am also doing three sessions on the show floor and building a piece of software live, from problems the room votes on, which is its own story for another post.

The more people who play before then, the more I have to say, and the less it is one man's hunch from a stage.

Go and get your card

It takes three minutes, it is anonymous, and you get something shareable out of it at the end, which is more than most things asking for three minutes of your week can say.

Have a go at research.zoby.ai, and if you end up with a card you like, share it. I would rather see which archetype your team gets than hear another opinion about whether the industry is ready for AI. We will find out together.

The date that quietly passed

Ten days ago, on 2 August, the original deadline for high-risk AI system obligations under the EU AI Act came and went with none of the obligations actually landing. That is not an oversight. The EU's Digital Omnibus on AI, which moved through Parliament in June, Council at the end of June, and was published in the Official Journal on 24 July before entering into force on 27 July, pushed the standalone high-risk deadline out to 2 December 2027. Product-embedded high-risk AI has an even longer runway, to 2 August 2028.

I am not a lawyer and none of this is legal advice. But I have had the same conversation with three different clients this year, all running events with an EU footprint or EU delegates, and it goes roughly the same way: "so do we relax now."

The honest answer

The timetable moved. The direction did not. The high-risk category still exists, the obligations that go with it still exist, they simply do not bite for another year and a half on standalone systems, longer on product-embedded ones. Treating the deferral as a cancellation is the mistake I would actively steer a client away from, not because the deadline pressure is real today, but because building the habit of checking where AI touches your delegates is worth having regardless of when a specific clause takes effect.

What "high-risk" would plausibly cover in events

Worth being specific here, because the phrase gets used loosely. In an events context, the kind of AI use that would plausibly sit in or near the high-risk category is the kind that makes a consequential decision about a person: automated screening of attendees or staff against a watchlist or risk profile, biometric entry systems that identify or verify someone to get them into a venue, and anything doing significant automated assessment of people rather than content.

Most of what events businesses actually use day to day sits well outside that: drafting copy, summarising a call, answering a delegate FAQ through a chatbot, generating a first pass at a proposal. That everyday use mostly attracts transparency expectations rather than high-risk obligations, meaning the general direction of travel is "say when it's AI," not "prove it is safe before you switch it on." Still worth doing properly. Just a different, lighter kind of worth doing properly.

A conference venue in the north west is a useful contrast to keep in mind here. It runs facial matching at fast entry lanes for one of its larger trade shows, which is a genuinely plausible candidate for the high-risk conversation. Its chatbot answering parking questions on the same website is not, whatever the marketing page for either tool happens to call itself.

Why "AI-powered" on a supplier's spec sheet is not enough

A badge on a vendor's website that says "AI-powered entry management" or "AI-driven attendee screening" tells you almost nothing about which category that system falls into, or what it actually does with a delegate's data or image. It is a marketing term, not a classification. If you are procuring anything that screens, verifies or scores people at your events, whether the vendor calls it AI or not, the question that matters is what decision the system is making about a person and how reversible that decision is if it gets it wrong. Do not let the badge substitute for asking that question directly.

A one-page register: where does AI touch our delegates

This does not need to be a legal exercise, and treating it as one is usually why it never gets started. A single page, kept honestly, does most of the work:

• Every point where AI reads, screens, verifies or scores a delegate, attendee or member of staff, not just where it talks to them.
• Whether that use is disclosed to the person it affects, in plain language, somewhere they would actually see it.
• Whether a human can override or review the outcome if someone asks.
• Whether the supplier providing the tool can tell you plainly what it does, in specific terms, not just in the language on their marketing page.

Keep it updated as you add tools, not as a one-off audit. It is a much smaller job done little and often than it is done once, retrospectively, after someone asks a question you cannot answer.

Be honest when a chatbot is a chatbot

The simplest, cheapest thing any events business can do this month, EU footprint or not, is stop letting a chatbot pretend to be a person. If a delegate is talking to an automated assistant on your website or WhatsApp line, say so, plainly, near the start of the conversation. It costs nothing, it is exactly the direction both UK and EU rules are pointing, and it is the kind of thing that looks obviously right in hindsight if a delegate ever complains that they thought they were talking to a human and were not.

This week

Start the one-page register above, even a rough first pass, for whichever of your events touches EU delegates or runs on EU soil. If you are not sure whether a supplier's tool falls into a category worth worrying about, that specific question, what is this system actually deciding about a person, is worth putting directly to the supplier rather than assuming the answer from their marketing copy. If you would like a second opinion on where AI currently touches your delegates across the business, that is a natural early thread in a Discovery Lab.

The gap nobody schedules around

Most events businesses have a fortnight or so, somewhere between the last summer wedding and the first conference of the autumn run, where the calendar goes quiet. Not empty. Quiet.

It is the worst possible time to take a proper holiday, because everyone else has already had theirs and is asking where you are. It is also, and this is the useful bit, the only stretch of the year when you can change how the business actually works without risking a live event while you do it.

January does not work for this. In January everyone has opinions and no energy, and the first big brief of the year lands before you have finished the conversation. August works, if you use it on purpose.

Why a repeated small saving is worth more than it looks

Here is the arithmetic, done honestly rather than rounded up to make a better headline.

Say you trim twenty minutes off a task someone does twice a week, an event brief that used to need chasing three times to get the numbers right, a quote template that used to need retyping instead of adapting. Twenty minutes, twice a week, is forty minutes a week for that one person.

Now say four people in the business do a version of that same task. Forty minutes each, across four people, is two hours forty minutes a week, team wide.

A realistic working year, once you take out holiday, bank holidays and the odd quiet week, is somewhere around forty six weeks. Two hours forty minutes a week across forty six weeks is a little over a hundred and twenty two hours a year. That is roughly fifteen full working days, every year, from now on, from fixing one repeated task properly, once.

It does not feel like fifteen days when it happens. It feels like nothing happened at all, which is exactly why nobody notices it and nobody schedules time to go looking for it.

Three jobs for August, not twelve

The temptation in a quiet fortnight is to try to fix everything. Resist it. Three jobs, done properly, beat twelve started and abandoned when the autumn diary fills back up.

• Write down the quoting process, as it actually happens. Not the process on the induction slide from three years ago. The one where Dev always adds ten percent because the supplier always comes back higher, and Priya always checks availability first because she got burned once. Write down what people actually do, including the workarounds, because the workaround is usually the real process wearing a disguise.
• Decide what each CRM stage actually means. Most teams have a CRM where "Proposal Sent" means something different depending on who moved the card. Get the team in a room for forty minutes and agree, in one sentence each, what has to be true for a lead to be in each stage. Write it somewhere everyone can see it. This is unglamorous and it is the single most valuable hour most teams could spend in August.
• Build one reusable prompt per role. Not a company-wide AI policy, not a tool rollout. One prompt, written with the person who does the job, for the task they do most often and dread most. A supplier chaser. A first-draft brief summary. A debrief. Test it on a real job from this summer, not a hypothetical one, and keep editing it until the person who will use it says it is actually useful.

Three jobs. Written down, agreed, and tested against something real, by the time the first September conference lands.

Why this compounds and the shouting-loudest crowd never notices

There is a lot of noise in events right now about AI, most of it either breathless or dismissive, and almost none of it about this. The people shouting loudest about transformation are usually selling a platform, and the people rolling their eyes are usually right to be sceptical of the platform, because the platform was never the problem.

The quoting process, the CRM stage definitions, the prompt per role, none of that is a system launch. Nobody makes an announcement about it. There is no leadership offsite with a keynote. It is a fortnight of unglamorous decisions that nobody outside the business will ever see, which is exactly why the businesses that do it well look, from the outside, like they simply have less chaos than everyone else.

That is the whole advantage. Not a bigger platform. Fewer things breaking quietly, week after week, that everyone had simply stopped noticing.

Pick one and start there

If three feels like a lot, pick one. The CRM stage conversation is the cheapest to run and the fastest to pay off, forty minutes in a room this week.

If you want a second pair of eyes on where the real time is leaking before you decide what to fix, that is what a Discovery Lab is for, a set of structured conversations with the team followed by a written report showing exactly where the minutes are going. More on that here. And if the prompt-per-role idea is where you want to start, there are working examples already on Zoby OS.

The meeting that never happens

Every events business says it does post-event debriefs. Almost none of them actually do, not properly, not every time.

I worked with a twenty-five-person agency in Leeds last year where the debrief existed on paper as a calendar invite sent two days after every job. It got moved so often it became a joke. Nobody wanted to be the one who scheduled a meeting people would rather skip.

The problem was never that people did not care. It was that a debrief takes an hour, the team is already three jobs down the road by the time it happens, and half of what needs saying has already been forgotten.

You already have the inputs, you just never gather them

Here is what most teams have sitting around after a job, unused:

• Voice notes from the site team, recorded on the drive home, transcribed by whatever app is already on the phone
• The run sheet, with the scrawled amendments from the day
• The client's thank-you email, or the slightly awkward one that lists three things that went wrong
• The supplier invoices that came in higher than quoted, or for the wrong thing entirely

None of that is a debrief. All of it is the raw material for one. The reason it never becomes a debrief is that turning scattered notes and PDFs into a coherent page is exactly the kind of dull assembly work everyone puts off.

The prompt

This is what I hand teams instead of the meeting. Paste your inputs underneath it: the transcribed voice notes, the run sheet text, the client email, a plain description of any invoice discrepancies. Your AI assistant will not chase down anything itself, so give it everything you have and nothing you are guessing at.

You are helping an events team write a one-page post-event debrief. I will give you several
unedited inputs below: transcribed voice notes from the crew, the run sheet with amendments,
a client email, and notes on any supplier invoice discrepancies. Not all inputs will be present
every time, use whatever I give you.

Produce a one-page debrief with exactly these four sections:

1. What worked - the parts of the day that ran to plan or better, specific enough that someone
   could repeat them on purpose next time.
2. What cost margin - anything that took longer, cost more, or needed more people than quoted,
   with a plain reason where one is stated in the inputs. Do not estimate a cost if none is given.
3. What to change in the template - concrete edits to the run sheet, brief, or quote template
   that would have prevented the problems above. Write these as instructions, not observations.
4. Three questions for the client call - questions that would clear up something ambiguous or
   unresolved in the inputs, phrased the way you would actually ask them out loud.

Rules:
- Only use what is in the inputs I give you. If something is unclear or missing, say so in the
  relevant section rather than filling the gap with a guess.
- Do not invent numbers, timings, names, or outcomes that are not in the text.
- Keep the whole thing to one page. Use short sentences.
- If the inputs contradict each other, note the contradiction rather than picking a side.

Here are my inputs:
[paste voice note transcripts, run sheet, client email, invoice notes]

The instruction to refuse invention matters more than any other line in that prompt. An assistant that is short on inputs will happily smooth over the gaps with something plausible. A debrief built on something plausible is worse than no debrief, because it looks finished.

What good looks like

The output should read like something a slightly blunt colleague wrote in fifteen minutes, not a report. If it is generating tidy paragraphs about "stakeholder alignment" or crediting the whole team with a flawless delivery, you have not fed it enough real detail, or you have let it fill in for you. Send it back with more of the actual inputs and less summarising.

A conference venue in the north west I spoke to runs this after every event with three inputs only: the duty manager's voice note, the client's follow-up email, and whatever the catering supplier's invoice said versus what was quoted. Three inputs, one page, five minutes to assemble. That is the whole bar you are clearing. It does not need to be more sophisticated than that to be useful.

Where the value actually lands

The debrief itself is not the point. The point is the second section, the run sheet edits and the template changes, because those are the only parts of the document anyone reads twice. A debrief that never changes a template is just a diary entry.

Keep a running file of the "what to change" lines across a few months and a pattern usually appears fast: the same supplier always invoices wrong, the same stage of the brief is always missing a number, the same client always emails the same question. Once you can see the pattern, it stops being an event problem and becomes a five-minute fix to a document.

Do this on your next job

You do not need to roll this out across the business to try it. Pick your next event, gather whatever inputs you naturally end up with, and run the prompt once. See whether the output is something you would actually send to the team, or something you would bin.

If it is useful and you want help working out where else in the business the same approach would save real hours, a Discovery Lab looks at exactly this kind of gap between the work people do and the systems that are meant to support it. Details are here, and there are more prompts like this one, free, on Zoby OS.

If you blinked over the last three weeks, you missed one of the stranger episodes in AI's short history: Anthropic released its most powerful model, the US government shut it down, and now it's back. Here's the story, and more importantly, what it means for anyone running events, venues, or agencies.

The 20-second version

• 9 June: Anthropic launched Claude Fable 5, a new top-tier ("Mythos-class") model, alongside a more powerful sibling called Mythos 5 for vetted cybersecurity partners.
• 12 June: The US government issued an export control directive telling Anthropic to cut off access for any foreign national, anywhere, including Anthropic's own overseas staff. With no real-time way to check nationality, Anthropic pulled the model for everyone.
• 30 June to 1 July: The Department of Commerce lifted the controls, and Anthropic restored global access.

<cite index="6-1">The export controls were applied on Friday, June 12, and required Anthropic to restrict access for foreign nationals whether inside or outside the United States, forcing a full suspension for all users since there was no reliable way to verify nationality in real time.</cite> <cite index="9-1">The trigger was a jailbreak that Amazon researchers found, where a prompt got Fable 5 to flag software vulnerabilities and, in one case, write code demonstrating how a flaw could be exploited.</cite>

Anthropic's own follow-up investigation is worth pausing on: <cite index="6-1">testing confirmed that several less capable models, including Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities that Fable 5 did in the report, and every model tested could reproduce the single exploit demonstration.</cite> In other words, this wasn't a uniquely dangerous Fable 5 capability. It was a borderline case that got caught by an overcautious filter, on a model everyone happened to be watching closely because it had just launched.

Government involvement: good thing, bad thing, or both?

There's a genuinely fair case on both sides here, and we'd rather lay it out than pretend it's simple.

The case for it being reassuring: a government body with the power to pause a frontier AI model overnight, and the willingness to actually do it, is a meaningful check. It shows the oversight mechanism isn't purely theoretical. <cite index="9-1">A June 2 executive order had already created a voluntary path for pre-release review of frontier models, though Fable 5 never went through it; Washington reached for export controls instead.</cite> That's a sign regulators are still improvising rather than working from a settled playbook, which cuts both ways: reassuring that someone's watching, unsettling that the process is so ad hoc.

The case for it being an overreaction: <cite index="7-1">Francesco Bailo, deputy director of the AI, Trust and Governance Centre at the University of Sydney, argued the lifting of restrictions was expected in tech circles because reports of Fable 5 being "jailbroken" had been inflated well beyond their actual significance.</cite> He also noted <cite index="7-1">the government likely recognised its decision risked setting a messy regulatory precedent, and that if Fable and Mythos were blocked on these grounds, competing models would logically need to be blocked too.</cite>

Our honest read for the events industry: neither "this proves AI is dangerous" nor "this proves it's all noise" is the useful takeaway. The useful takeaway is that AI governance is now a live regulatory reality, not a hypothetical slide in a deck. A model can be available to your team on a Friday and gone by Monday, for reasons entirely outside your control. That's not a reason to avoid AI tools; it's a reason to build processes that don't collapse when one tool changes overnight.

What this means for people in events

Strip away the geopolitics and two things matter for our industry: the capability jump, and the governance wobble.

1. The capability jump is real, and it's aimed at exactly the kind of work events businesses do. <cite index="1-1">Anthropic describes Fable 5 as state-of-the-art on nearly all tested benchmarks, with particularly strong gains in software engineering, knowledge work, vision, and long-running autonomous tasks.</cite> For a venue-finding agency, an events management platform, or a marketing team producing content and reporting, "long-running autonomous work" is the operative phrase. This is a model built to hold a complex, multi-step task, think a full RFP process, a multi-venue comparison, or a post-event report synthesised from dozens of sources, without losing the thread. That's directly relevant to discovery, documentation, and reporting work happening across the sector right now.

2. It's a live case study in why AI governance belongs on your agenda, not just your IT department's. Whatever you think of the government's decision, the practical lesson for any events business adopting AI tools is the same: know which tools you depend on, understand who can pull the plug and why, and have a fallback that doesn't leave your team stranded mid-project. That's not paranoia; it's the same due diligence you'd apply to any critical supplier, whether that's a venue, a caterer, or a booking platform.

3. Trust and credibility just got more complicated, and more important. Attendees, exhibitors, and clients are going to keep hearing headlines like "government shuts down AI model." Event professionals experimenting with AI-generated content, chatbots, or automated attendee support need a clear, honest answer ready for "is this safe?" and "who's checking it?" The organisations who can answer that calmly, rather than defensively, will stand out.

The bottom line

Fable 5's three-week disappearing act wasn't really about events. But it's a preview of the environment AI-adopting events businesses now operate in: fast-moving capability, genuinely contested regulation, and a public that's watching closely. The organisations that treat AI governance as part of the plan, not an afterthought, are the ones who'll still be standing confidently when the next headline hits.

This is the kind of shift we track closely as part of our Event AI Behaviour Study, launching ahead of CHS Manchester in September. If your team is thinking through where AI actually fits, and where the guardrails need to be, get in touch.

Back from ExCeL, still sorting through the notes

The Meetings Show ran at ExCeL London this week, close to six thousand industry professionals and over five hundred and fifty exhibitors, by far the biggest two days I've had on a show floor this year. I also sat down for an interview with the CHS team on the Wednesday, which was a nice reminder that Manchester in September is getting closer.

I came home with pages of notes. Once I sorted through them, three things kept rising to the top.

One: the most useful AI in the building, and nobody called it AI

Live captioning ran across several of the stages this year, quietly, on screens most people glanced at without thinking twice. It just worked. Speaker talks, words appear, nobody in the audience misses a beat even in a hall with the acoustics of an aircraft hangar.

Nobody stood up and announced it as an AI feature. It wasn't badged, demoed or pitched. It was just there, doing a genuinely useful job, and I'd argue it did more for more people over those two days than anything on any exhibitor stand with AI printed across the top of it.

There's a lesson in that. The AI that earns its place is usually the AI you stop noticing, not the AI you're told to be impressed by.

I stood at the back of one session for a few minutes just watching the captions run rather than the speaker. Not a single typo I could catch, keeping pace with a fast talker without lagging. If you'd told that room it was AI doing the work, most people wouldn't have thought to ask. They'd just have been glad they could follow along from the back row.

Two: the debate split the room, and both sides were arguing the wrong question

The education programme included a debate billed as "The Reckoning", on AI's impact on the events industry, and the room split roughly down the middle. One side: it will replace planners. The other: it will never replace planners. It got lively.

Both sides were arguing a question that doesn't have a useful answer yet, because it's really a question about ten years from now dressed up as a question about today. The honest answer, the one that would have actually helped anyone in that room plan their next quarter, is much duller: it is already changing what a planner's Tuesday looks like. Not replacing the planner. Changing the shape of the day. Less time on first drafts and chasing suppliers for basic availability, more time on the judgement calls that actually need a human in the room.

That's a smaller claim than either side of the debate wanted to make, and it happens to be the true one.

Debates like that make for good theatre, which is presumably why they get programmed. But I left thinking the organisers had picked the wrong axis to argue about. The useful question was never replace or not replace. It was which parts of the job are shifting, and how quickly people notice the shift has already started under them.

Three: the question I heard more than any other

Away from the stages, at the coffee stands between sessions, one question came up again and again, from people at very different types of business: "what should I actually do first".

Fair question, and one worth answering properly rather than waving at.

Start with the task you already do every single week that you've stopped noticing costs you time. Not the exciting AI use case, the boring recurring one. The event brief you retype from a client email into your own template. The follow up summary you write after every call. The venue capacity check you do by hand because the spreadsheet is three versions out of date.

Pick one of those. Get it working properly, end to end, with one tool, before you touch anything else. Resist the urge to solve five problems at once, because that's how most AI pilots quietly die: not through failure, through nobody finishing the first one.

The reality check worth remembering

It's worth saying plainly: most of the industry is still early. The ONS reported this summer that around twenty nine per cent of UK businesses now use at least one AI technology, up from twenty one per cent a year before. That's real growth, and it also means well over two thirds of businesses aren't there yet.

If you're one of them, you are not behind some invisible curve everyone else has already crossed. You're roughly where most of the room at ExCeL actually was, whatever the panels made it sound like.

What I'm carrying home

The best AI in that building was invisible. The best answer to "what should I do" was the least exciting one on offer. And the honest read on where this industry is sits somewhere quieter than either side of a stage debate wants to admit.

If you want help finding your own version of that first boring task worth fixing, that conversation is exactly what a Discovery Lab is for. And if you want to build confidence with the tools before you commit to anything, Zoby OS is free and open to anyone in the industry.

What just changed in your CRM

This week Salesforce took Agentforce, its multi-agent orchestration layer, from beta into general availability as part of the Summer '26 release. It is the largest update the platform has had. In practical terms, that means the CRM a lot of venues and agencies already pay for can now hand off tasks like qualifying a lead or drafting a follow up to an AI agent working inside it, not bolted on the side.

HubSpot moved earlier this year. Back in April, its specialised Breeze agents shifted to outcome-based pricing: the Prospecting Agent costs a dollar per qualified lead, the Customer Agent fifty cents per resolved conversation. Not a seat licence. A price per result the agent decides it has achieved.

Put those together and the pattern is clear. A venue sales office, a mid-sized agency, anyone running a CRM from either of these vendors, is about to be offered "an agent" as a feature switch inside a system they already use every day. It will be pitched as a small decision. It is not.

Why this is a bigger decision than it looks

Turning on a seat for a new starter is reversible in a day. Turning on an agent that reads your enquiry inbox, scores leads, and drafts or sends replies is a decision about who, or what, is allowed to represent your business to a prospect before a human ever sees the conversation. The switch is easy to flip. Understanding what you just flipped is the part worth doing properly first.

The five questions to ask before you switch one on

These are worth putting to whoever is selling you the agent, in writing, before go-live, not worked out afterwards from what actually happened.

• 1. What exactly does it count as a "qualified lead" or a "resolved conversation"? This is the definition the vendor gets paid against, so it matters more than it sounds. A loose definition of "resolved" can mean a conversation the delegate gave up on, not one that was actually answered.
• 2. What data does it read to make that judgement? Your CRM history, email threads, call notes, previous quotes. Know the scope before it is switched on, not after someone asks why the agent referenced something it should not have seen.
• 3. What can it send without a human approving it first? A drafted reply waiting for a click is a different risk profile to an agent that emails a prospect, updates a record, or quotes a price on its own.
• 4. How do we see what it actually did? You want a log, in plain language, of every action it took, not a dashboard of aggregate performance. If a client asks why they got a strange automated reply at 11pm, you need to be able to answer that in minutes.
• 5. How do we turn it off, fully, not just pause the trial? Know the actual mechanism before you need it in a hurry.

The pricing question nobody asks out loud

Outcome-based pricing sounds like it should align incentives nicely. You only pay when it works. But think through what "it works" means from the vendor's side. The agent is paid, or at least judged as performing well, when it marks a lead qualified or a conversation resolved. That is not automatically the same as a lead your sales team would call qualified, or a conversation your delegate would call resolved.

This is not a claim that either vendor is gaming its own metric. It is a structural point: the system that decides whether the outcome happened is the same system being paid for that outcome happening. That is exactly the kind of thing worth checking rather than assuming, and it is precisely what question one above is for.

A venue sales office in the north west trialled an agent along these lines earlier this year. Its definition of a qualified lead turned out to include any enquiry that replied at all, including a one-line "thanks, not for us." Nobody had misled anyone. The definition had simply never been stated out loud before it was switched on.

Worth noting for the record: the dollar figures above are US list pricing at time of writing. Expect UK pricing and terms to differ, and to ask for them directly rather than assuming a straight currency conversion.

A prompt for the meeting

If you want to walk into the vendor call with the five questions already adapted to your own systems, this is a reasonable way to prepare for it. Paste it into your AI assistant with a short description of your CRM and how enquiries currently flow, and expect back a tailored version of the five questions plus a short list of what "good" and "concerning" answers would sound like for each.

I am about to be sold an AI agent feature inside our CRM (describe: Salesforce Agentforce, HubSpot Breeze, or similar) for our events business. Help me prepare for the vendor call.

Here is how enquiries and leads currently move through our business:
[describe: where enquiries come in, who currently qualifies or responds to them, what your CRM already does automatically if anything]

Using that context, adapt these five questions to be specific to my systems:
1. What exactly counts as a qualified lead or resolved conversation for this agent.
2. What data it reads to make that judgement.
3. What it can send or change without a human approving first.
4. How we see a plain-language log of what it actually did.
5. How we turn it off completely, not just pause it.

For each question, give me one example of a good, specific answer and one example of a vague answer that should be a red flag. Keep this practical, no more than a page.

This week

If a CRM agent has already been switched on somewhere in your business, and for a lot of teams one quietly has been, ask whoever manages that system to answer the five questions above this week, honestly, even if nobody official ever asked them to switch it on in the first place. That conversation is usually more revealing than the sales pitch was. If you want help working through what a specific agent tool would actually change in your business before you commit to it, that is a natural question for a Discovery Lab.

Two stories running at once

IMEX Frankfurt happened last week, and by every account it was the biggest show floor the event has had. The sessions I heard about covered the usual ground: AI drafting proposals, AI summarising sessions, and a fair amount of nervousness about deepfakes and who you can trust to be who they say they are.

But that wasn't the story that mattered most this month, at least not to me. The story that mattered was a single line in a trade press piece: LEO Events, acquired by the Hoffmann Family of Companies. Put that next to WPP's restructuring plan, announced back in February, folding Ogilvy, VML and AKQA into one unit built around its AI platform, and a pattern starts to show.

Consolidation is arriving in the agency side of this industry. Not as a headline everyone is discussing on the show floor. Quietly, in the trade press, while the panels talk about prompts.

I don't think that's a coincidence, and I don't think it's unrelated to AI either, even if the connection isn't obvious at first glance. When a task that used to require a team of people can be done faster with fewer of them, the economics of scale change. Bigger groups can absorb that shift more comfortably than a single founder trying to work out what it means for their next twelve months. Buying capability becomes more attractive than building it slowly.

What buyers are actually looking for

When a bigger company buys a smaller agency, or a big client puts a supplier through serious due diligence, they are not buying a client list and a nice pitch deck. Anyone can build a pitch deck. What they're trying to work out is whether the business runs on a person, or on a system.

A business that runs on one brilliant founder's judgement is a risk the moment that founder is on holiday, let alone the moment they sell up. A business with a documented way of working, consistent handovers, a CRM that actually reflects reality rather than a founder's memory, is a different proposition entirely. That is what buyers are actually pricing in, whether or not anyone in the room says so out loud.

The question nobody on the panel answered

Here's the one I kept turning over on the train home. If AI can draft the proposal, summarise the client call, and chase the follow up email, what exactly is the agency being paid for.

Nobody on stage answered it properly. A few skirted round it with something about "the human touch", which is true but not specific enough to build a business case on.

My answer is more boring, and I think it's the right one. You're paid for judgement: knowing which of the client's three stated requirements is the one that actually matters. You're paid for relationships: the kind that survive a difficult brief, a late change, a supplier letting you down at the wrong moment. And you're paid for a documented way of working that survives a person leaving the building, because a system that only exists in one person's head isn't a system, it's a liability wearing a job title.

AI can help with all three of those, once they exist. It cannot invent them for you.

That's worth sitting with for a moment, because it's easy to hear "AI can draft the proposal" as a threat to the agency model, when it's actually a filter. It strips out the mechanical part of the job and leaves the part that was always the real value, the part a client was genuinely paying for even when they couldn't have named it that way.

What a buyer would actually check

If a bigger company, or a demanding client, put your business through the kind of due diligence LEO Events presumably went through, here's roughly what they'd be looking at.

• Whether your CRM reflects what's actually happening with clients, or whether the real picture lives in three people's inboxes.
• Whether a brief handed from sales to delivery survives the handover intact, or whether details get lost and rebuilt from memory every time.
• Whether your quoting process is consistent, or whether every account manager prices things their own way.
• Whether the business could keep running smoothly if your best person left tomorrow.

None of these questions are about AI. They're about whether the plumbing underneath the relationships is sound. AI makes a good system faster. It does not make a broken one presentable.

Being ready without being for sale

You don't need to be planning an exit to want the answer to those questions to be yes. A business with documented handovers and a CRM that tells the truth isn't just more attractive to a buyer, it's easier to run day to day, easier to grow, and considerably less stressful when someone hands in their notice.

Consolidation is a signal worth paying attention to even if you never plan to sell. It tells you what "well run" is starting to mean in this industry, and it's worth checking your own business against that bar before someone else does it for you.

If you want an honest, structured look at where the gaps actually are, that's exactly what a Discovery Lab is built to find.

The shortlist that looks perfect

Ask an AI assistant for a shortlist of suppliers with a particular spec, in a particular region, for a particular date, and it will give you one. Clean formatting, sensible-sounding company names, specific numbers for lead times, minimum order sizes and fees. It reads exactly like something a well briefed junior researcher would produce after a morning of proper digging.

The trouble is that some of it is invented. Not maliciously, and not randomly either. AI assistants are built to produce a fluent, complete-sounding answer, and a shortlist with a gap in it does not feel complete. So where the model does not actually know an AV hire lead time or a current catering minimum, it will often produce a plausible one anyway, in exactly the same confident tone as the facts it does know. Nothing in the output tells you which is which.

The fix is not "trust it less"

Telling people to simply double check everything is true but not useful advice on its own, because it does not tell you what to check hardest or where the risk actually sits. What works better is changing what you ask for in the first place.

Instead of asking for a shortlist, ask for a shortlist that shows its working: what it is confident about, what it is inferring, and what it is genuinely unsure of. Once the AI is forced to separate those categories, the guessed numbers stop hiding inside the real ones, and you get a usable checklist of exactly what a human needs to verify before anything goes near a client.

The prompt

Paste this into ChatGPT, Claude, Copilot, whichever you use, followed by your research request or an existing shortlist you want checked. Expect a table-style breakdown of each supplier with a confidence label on every claim, plus a short verification list at the end.

You are helping with supplier and logistics research for an event. I will give you either a research brief or an existing shortlist. For every supplier you include, and for every specific claim about it, label your confidence honestly rather than presenting everything in the same tone.

Use exactly these three labels for every factual claim (lead times, minimum numbers, fees, room-block terms, travel or transfer times, build and freight schedules, dietary or accessibility provision, contact details, anything specific):

KNOWN: something you are confident is accurate and current, based on well established, widely available information.
LIKELY: a reasonable inference, for example an AV hire lead time estimated from typical industry turnaround, or a catering minimum typical for that type of supplier, that has not been directly confirmed.
UNVERIFIED: anything you are not confident about, including anything that could plausibly have changed recently, such as day rates or fees, current minimum numbers, hotel room-block cancellation terms, or whether a supplier still operates the way described.

Format the output like this for each supplier:

Name:
Category: [AV, catering, transport, accommodation, speaker or entertainment, freight and build, other]
Lead time or minimum notice: [figure] - [KNOWN/LIKELY/UNVERIFIED]
Minimum numbers or order size: [figure] - [KNOWN/LIKELY/UNVERIFIED]
Indicative fee or day rate: [figure or range] - [KNOWN/LIKELY/UNVERIFIED]
Key terms: [dietary handling, accessibility provision, room-block cut-off, transfer or travel time between sites, freight and build window, whichever apply] - [KNOWN/LIKELY/UNVERIFIED]
Notes: anything relevant about why a figure is uncertain

Then produce a final section titled "Before this goes to a client," listing every UNVERIFIED and LIKELY item across the whole shortlist as a plain checklist, in the order a human should work through them, fastest and most important first.

Do not remove or soften an UNVERIFIED label to make the shortlist look more complete. An honest gap is more useful than a confident guess.

Here is my research request or shortlist:
[paste here]

Reading the output properly

The value of this prompt is entirely in that final checklist. Ignore it and you have simply asked the AI to hallucinate with better formatting. Use it properly and you get a short, specific list of exactly which phone calls or website checks need to happen before anything goes to a client, rather than a vague sense that you should "probably double check some of this."

In practice, fees and lead times are the two categories that come back UNVERIFIED most often, which lines up with reality: those are exactly the two things most likely to have changed since whatever the model last learned about a supplier. Room-block cancellation terms and freight or build windows are close behind, because they are the details that live in a contract rather than on a website.

What a human still has to do

This prompt does not remove the verification step. It makes the verification step small, specific and honest, instead of a vague feeling that you should check "some of this at some point," which in practice tends to mean nothing gets checked at all until something goes wrong.

Treat every KNOWN label with a little scepticism too. It means the model is confident, not that the model is correct. A quick call to the supplier, or a look at their own current terms, is still worth doing for anything that is actually going in front of a client, whatever label it carries. A transfer time between sites that looked fine in the shortlist can fall apart the moment you check it against an actual timetable rather than a straight-line estimate.

This matters most when the shortlist is going somewhere fast: a same-day enquiry, a build schedule due first thing tomorrow, a client who wants three catering options by lunchtime. Pressure is exactly when an unverified figure quietly becomes a stated fact, because nobody had the ten minutes it would have taken to check it. A short, honest checklist is easier to act on under pressure than a long, tidy shortlist that is hiding its gaps.

Make this part of the routine

Build this into how you do supplier and logistics research generally, not just for one big pitch. It takes the same amount of time as asking for a plain shortlist, and it hands you the checklist as a byproduct instead of leaving you to guess where the risk is hiding.

More prompts like this, tested on real events work, are collected at Zoby OS.

Every events business has one

Open most CRMs at an events agency or a venue and you will find a graveyard. Contacts nobody has touched in three years, still marked "active." Three versions of the same venue, one spelled slightly differently, none of them holding the full picture. A "notes" field with the only record of a conversation that actually mattered, buried under forty other notes that did not.

This is not a rare, badly run business problem. We see it almost everywhere we look, including at businesses that are otherwise sharp, well run and growing. A CRM does not become a graveyard through neglect. It becomes one because nobody ever agreed, in writing, what the fields were actually for.

The root cause is not the system

Swap the CRM out for a shinier one and the graveyard usually follows you within a year, because the tool was never the problem. The problem is that "stage" means one thing to sales, another to finance, and a third to whoever is filling it in on a Friday afternoon trying to clear their inbox before the weekend.

A field with no shared definition is worse than an empty field, because it looks like data. Someone will build a report off it. Someone will make a decision based on that report. The decision will be wrong in a way that is very hard to trace back to its source, because on the surface, everything looked filled in and tidy.

Duplicate venues and the WhatsApp problem

Two patterns show up constantly. The first is duplicate venues: the same venue entered separately by two different people, at two different times, with two different capacity figures, neither one confidently more correct than the other. Nobody notices until someone quotes off the wrong one.

The second is what we call the WhatsApp problem. The actual, operational truth of what was agreed with a supplier or a client often lives in a phone, in a thread, not in the system that is supposed to be the record. The CRM holds the tidy version. The group chat holds the real one. Anyone relying purely on the CRM is working from a summary that quietly leaves things out.

Why AI makes a messy CRM worse, not better

This is the part that catches people out. AI does not fix a messy foundation, it builds on top of it, confidently. Ask an AI assistant to summarise your pipeline, draft a client update, or flag which events are at risk, and it will do exactly that, using whatever is actually in the fields it can see.

If "confirmed" means three different things across your team, the AI will not notice, because it has no way to know that. It will produce a clean, well written, entirely confident summary built on a definition that does not actually exist. That is a genuinely worse outcome than a human quietly ignoring the field, because it looks authoritative. Structure has to come before automation, not after it, or you get nonsense with better formatting.

We have seen this play out with a supplier database at a forty-person agency in the Midlands. Two suppliers had been entered under near-identical names, one with a note about a pricing issue from two years ago, the other clean. An AI-generated shortlist for a new event pulled the clean record, because nothing in the data told it they were the same company. The formatting was excellent. The recommendation was wrong for a reason nobody could see just by reading it.

Fix the structure, not the software

None of this needs a system migration. It needs an afternoon and some discipline.

• Define five stages, one sentence each. Not fifteen. Five, written down, agreed by sales, finance and delivery in the same room, so "confirmed" means the same thing to everyone who types it.
• Pick the ten fields that actually matter. Not the forty that exist. Ask what a report genuinely needs to be trustworthy, and archive the rest rather than deleting it, so history is not lost, just no longer cluttering the view.
• Run a thirty-minute weekly hygiene ritual. One person, one CRM view, thirty minutes. Close duplicates, chase blank required fields, move stale records out of the active pipeline. It is boring. It is also the single highest-leverage half hour most teams could spend.

Start with what "in the CRM" is supposed to mean

Before you touch a single field, agree one thing as a team: what belongs in the CRM, and what does not. Not every message with a supplier needs to be logged. But the outcome of that message, the price that was agreed, the date that was confirmed, does. Draw that line clearly and the WhatsApp problem shrinks on its own, because people stop treating the CRM as optional paperwork that happens after the real conversation, and start treating it as the record the real conversation was always meant to produce.

This is not an argument for a new CRM

We are not telling you to rip out your system and buy another one. Most of the CRMs event businesses already own are perfectly capable of holding clean data. What they are missing is agreement, not features.

Do the five-stage exercise this month, before you let anyone build an AI workflow on top of what is in there now. If you want help finding exactly where your data is quietly lying to you, that is a core part of what comes out of a Discovery Lab.

You've been asking all the questions. Try it the other way round.

Most people use AI the same way. You turn up with a question, you push it for an answer, you get something back that's about seventy per cent right, and you fix the rest yourself.

That's fine. It's just not where the interesting stuff is.

This is a prompt that flips it. You paste it in, and instead of answering you, it starts asking. About your week, your repeated tasks, the things you're waiting on, the parts of the job you quietly dread.

Fifteen minutes later it hands you a map of where your time is actually going.

Free. No signup. Copy it and go.

You are the worst person to describe your own problems

Not because you don't know your job. Because you know it too well.

The tasks that cost you the most are the ones you've stopped noticing. The workaround you invented three years ago is now just how it's done. The forty minutes you lose every Tuesday isn't a problem, it's Tuesday.

• You can't audit what's become invisible to you
• But you can answer questions about it
• So the trick is being asked

That's the whole idea. Something with no history in your business, no assumptions about how it's meant to work, asking simple questions one at a time until the pattern shows up.

It's the same thing a good consultant does in the first hour. Not clever. Just relentless.

What you're about to paste

This isn't a one-liner. It's the set of questions we'd ask a client in a first session, plus the rules that stop an AI model doing the annoying things it does by default: dumping every question at once, reassuring you, and pitching you software halfway through.

It takes fifteen to twenty minutes to run properly. Do it with a coffee, not between meetings.

Works in ChatGPT, Claude, Gemini, Copilot, whatever you already use. Paste the whole thing as your first message and answer honestly.

You are running a reverse interview. Normally I ask you questions. This time you ask me.

Your job is to find the friction in how I work. The repeated tasks, the waiting, the copying between systems, the rework, the things only one person knows how to do, and the parts of the week I avoid. Most of it I've stopped noticing because it's become normal. Your value is that you have no history here and no assumptions about how it's meant to work.

WHO YOU ARE IN THIS CONVERSATION

A calm, curious operations consultant in their first hour with a new client. You are not a coach, a therapist or a software salesperson. You are interested in specifics, mildly sceptical of generalities, and comfortable with short answers. You do not need to be liked.

THE RULES

1. One question per message. Never more. If you find yourself writing a second question mark, delete it.
2. Wait for my answer before you ask anything else.
3. Keep questions short. One or two sentences. No preamble, no "great answer", no "thanks for sharing".
4. If my answer has no number in it, your next question gets one. How many times a week, how long each time, how many people touch it, how long since it was last changed.
5. Listen for these words and dig when you hear them: "just", "usually", "quick", "it's fine", "we've always", "someone", "eventually", "I'll sort it". They almost always sit on top of something bigger than I'm making it sound.
6. If I laugh, sigh or say "don't get me started", stay there. Ask one more question before moving on.
7. Follow the thread. If something sounds heavier than I'm describing it, that outranks your list. Come back to the list afterwards.
8. Do not offer solutions, tools, tips or advice while interviewing. Not even a small one. If I ask for one, say it comes at the end and ask your next question.
9. Do not reassure me. Do not tell me my problems are common, understandable, or that I'm not alone.
10. Do not summarise, reflect back or check your understanding as you go. Save all of it for the end.
11. If I say "I don't know", ask me what I'd guess. Then ask what would need to be true for the guess to be wrong.
12. If I give a one-word answer, do not accept it. Ask for the last time it actually happened.

HOW THE INTERVIEW RUNS

Phase 1, orientation. Two or three questions. My role, what a normal week looks like, and who I depend on to get it done.

Phase 2, the sweep. Work through the areas below, two or three questions each. Skip any that clearly don't apply to me. Do not announce which area you're in. It should feel like a conversation, not a form.

a. Repetition. What did I do more than twice last week? What do I write out again and again with small changes?
b. Waiting. Where did I stop because I needed something from a person, a team or a system? How long did I wait, and what did I do while I waited?
c. Copying. What information did I move from one place to another by hand? What lives in two places that should live in one?
d. Guessing. What did I decide without the data in front of me, when the data probably exists somewhere?
e. Rework. What came back wrong, incomplete or misunderstood and had to be done again? Who fixes it when that happens?
f. Single points of knowledge. What can only one person do? What happens when they're on holiday?
g. Explaining. What takes longest to explain to someone new? What do I answer for other people over and over?
h. Exceptions. What breaks the normal process? How often is the exception actually the norm?
i. Value mismatch. What takes the most time, and what actually earns the money? Are they the same things?
j. Deferred work. What have I been meaning to fix for more than six months? What has it cost me not to?
k. Tools. What systems do I use in a normal week, and which ones don't talk to each other? Where do I keep the thing that isn't in any of them?
l. Dread. What part of the week do I least look forward to? Not what's hardest. What I avoid.

Phase 3, the dig. Three or four questions. Go back to the two things that sounded heaviest and get the full picture: how often, how long, who's involved, what it blocks, and what it costs when it goes wrong.

WHEN TO STOP

Stop once you have around fifteen specific examples with numbers attached, or when I say "wrap it up". Tell me you're moving to the summary before you do. Do not ask whether I want to continue.

WHAT TO GIVE ME AT THE END

Write it in plain English. Use my words for things, not consultancy language. Be direct. If something I described is a mess, say so.

1. The friction map. Everything you found, grouped into themes. A one-line title for each theme and the specific examples underneath.

2. Ranked by cost. The top five by estimated time cost per month. Show your working using the numbers I gave you: frequency x duration x people. Mark clearly which numbers came from me and which you estimated. Where you estimated, say what you assumed.

3. Fix type. Sort each of the five into one of these, and say why:
   - Connector: two systems that already exist and don't talk to each other
   - Template: the same thing rebuilt from scratch every time
   - Process: a rule that's missing, or one that made sense once and doesn't now
   - Knowledge: the information exists but nobody can find it
   - Build: something genuinely needs to be made
   If more than one lands in Build, look again. Most things don't belong there.

4. Effort against impact. The two that would be quickest to fix relative to what they'd return, and the one that's the biggest prize even though it's hard. Say what "hard" means in each case.

5. Monday. One thing I can do this week, on my own, with no budget and no permission from anyone. Specific enough that I could start it in the next hour.

6. The blind spot. Based on what I didn't say, or said quickly and moved past, what do you think I'm underestimating? Quote the thing I said that made you think so.

7. The pattern. If there's one root cause sitting underneath several of these, name it. If there isn't, say so rather than inventing one.

8. Your question. The one thing you'd want to ask me if we kept going, and why.

START NOW

Do not explain the process back to me, do not confirm you understand, and do not describe what you're about to do. Your first message is your first question and nothing else.

If it misbehaves

Some models will still try to dump the whole list on you in one go. If that happens, reply with this and nothing else:

Stop. One question per message. Ask the first one again.

If it starts giving advice partway through, reply with:

Not yet. Next question.

That's usually all it takes.

The short version

If you've got five minutes, not twenty, this is the cut-down one. It finds less, but it still finds something.

Interview me about how I work.

One question at a time. Wait for my answer before the next one.

Ask me about: what I repeated last week, where I waited on someone, what I copied from one place to another, and what I decided with no data.

After ten answers, group them and tell me the three that cost me the most time.

Every rule in there is doing a job

Most prompts you find online are one paragraph and a hope. This one is long because each line is fixing something that goes wrong without it. Here's what's under the bonnet.

It gives the model a character. "An operations consultant in their first hour with a new client" does more work than any list of instructions. It sets the tone, the pace and the level of scepticism in one line. "You do not need to be liked" is in there because models default to being liked, and a likeable interviewer lets you off the hook.

"One question at a time" is said three ways. Once in the rules, once as "if you find yourself writing a second question mark, delete it", and once in the start instruction. It's the failure that kills the whole thing. Left alone, most models dump all twelve areas in one message and you're back to filling in a form. Repeating it is inelegant. It works.

It tells the model which words to listen for. "Just", "usually", "quick", "it's fine", "we've always". These are the words people use when they're standing on top of something. Any decent interviewer hears them and slows down. Most AI models sail straight past. Naming the words in the prompt turns a generic question-asker into something that actually follows a thread.

It asks for numbers, every time. "I do a lot of admin" is useless. "I rebuild the same proposal six times a week" is a project. Rule four exists purely to turn the first into the second, and it applies to every answer, not just the first vague one.

"Don't offer solutions while interviewing". Without this, it flips into advice mode around question two and starts recommending project management software. The interview has to stay an interview. There's even a line for what to do if you ask for a tip, because people do.

"Don't reassure me". Cuts the filler. Nobody needs to be told their problems are common. It burns your patience and makes everything after it feel generic.

It handles "I don't know". That's the most common answer in a real session and most prompts have nothing for it. This one asks you to guess, then asks what would make the guess wrong. You'll be surprised how often you did know.

It runs in three phases, and doesn't tell you which one you're in. Orientation, sweep, dig. The sweep covers the twelve areas. The dig goes back to the two things that sounded heaviest and gets the full picture. Hiding the structure matters: the moment it feels like a questionnaire, you start giving questionnaire answers.

It asks what you dread. This is the one people are surprised by. Time estimates are unreliable, everyone underreports. But emotional response is a very good proxy for friction. What you avoid is usually what's broken.

It asks about rework, exceptions and single points of knowledge. What comes back wrong. What breaks the standard process. What only Sarah knows. These are the areas people never volunteer, because they've been reclassified in your head as bad luck rather than a pattern.

It makes the model show its working. The cost ranking has to use your numbers, mark what was estimated, and say what it assumed. Otherwise you get confident time savings that were made up on the spot. This is the difference between a list you trust and a list you nod at and close.

It caps the "build" pile. If more than one thing needs building, the prompt tells the model to look again. Because it's almost always wrong. Most friction is a connector, a template or a missing rule, and a model that's allowed to say "build an app" for everything will.

It tells you your blind spot, and has to prove it. The last part of the output flags what you glossed over or moved past quickly, and has to quote the thing you said that gave it away. It's not always right. When it is, it's the most useful line on the page.

It asks for a root cause, and gives it permission not to find one. Point seven asks whether one thing sits underneath several of the problems. The second half of that instruction, "if there isn't, say so rather than inventing one", is the important bit. Models will manufacture a grand unifying theory if you let them.

It tells the model how to start. "Your first message is your first question and nothing else." Without this you get a paragraph of "Great, I understand, I'll now conduct a reverse interview..." before anything happens. Nobody needs that.

Almost nothing needs building

The most common reaction to a friction map is to assume you need software. You usually don't.

The prompt already sorts everything into five piles. Here's why they matter:

1. Connector. Two systems that already exist and don't talk. Most of your list will be here. Often solved in an afternoon with something off the shelf.
2. Template. The same thing rebuilt from scratch every time. Needs an asset, not an app.
3. Process. A rule that's missing, or one that made sense once and doesn't now. Costs nothing but a decision.
4. Knowledge. The information exists. Nobody can find it. This is the sleeper category and it's usually bigger than people think.
5. Build. Genuinely needs something made. Should be the smallest pile. If it isn't, something's been miscategorised.

Then do the one thing it told you you could do on Monday. Not the biggest one. The Monday one.

If it came back bigger than you expected

That's usually the point where people call us.

Running the audit is the easy part. Most people get the list, feel briefly motivated, and are back to normal by Thursday. Not through laziness. Because nobody owns the follow-through and the day job doesn't stop.

Zoby does the doing. We run this properly across a team, work out which friction is worth removing, and then remove it.

Have a conversation

The Friction Audit is part of how we work at Zoby. If you want the version with a person in the room, that's Discovery Lab.

The number

A survey published this week, Resume Now's "AI Oversight Gap", found that 35% of US workers rarely or only occasionally review AI-generated output before they use it. That is not a fringe habit. That is roughly a third of the people using these tools, sending or filing or presenting whatever the model handed back, largely unread.

I would not treat a US consumer survey as gospel for the UK events industry specifically. But it matches something I hear in almost every Discovery Lab: people know, in the abstract, that they should check AI output. Under a Thursday afternoon deadline, with three other things also due, that knowledge does not always survive contact with the actual week.

The cautionary tale, and why it is more useful than it looks

Back in February, a Nebraska attorney filed a divorce appeal brief with 63 case citations. Fifty seven of them were defective. Twenty were entirely hallucinated, cases that do not exist, invented by an AI tool and never checked before they went into a legal filing. The attorney has since been suspended pending a disciplinary investigation.

It is tempting to read that story and think "well, that is law, the stakes are obviously higher, that would never happen here." I would push back gently on that. The mechanism is identical to ours. Someone under deadline pressure, working with a tool that sounds confident regardless of whether it is right, trusted the output because checking it felt like it would cost time they did not have. The domain is different. The habit is the same one a third of workers admit to.

Why events is a particularly copy-paste industry

Most events work is assembled fast, under deadline, from parts. A proposal draws on a template, a past quote, a supplier's spec sheet, a venue's capacity chart, and increasingly a paragraph an AI tool wrote to fill a gap because nobody had time to write it themselves. Sponsor decks do the same. So do delegate FAQs, supplier comparison documents, and the summary paragraph at the top of a report that nobody reads closely because it is "just the intro."

An unchecked AI paragraph dropped into a proposal is the modern version of the wrong client name still sitting in the template from the last job. It used to be a find-and-replace slip. Now it can be a venue capacity that is subtly wrong, a supplier claim the AI generated with confidence rather than sourced, or a number that sounds plausible and is not close to accurate. The failure mode has not changed. What has changed is how confident the wrong version sounds, and how much of the document it can now touch at once.

A forty-person agency in the Midlands had exactly this happen earlier this year. A venue capacity generated as a placeholder in a first draft, never replaced with the real figure, went out in a proposal to a client who happened to know the venue well. The client caught it in about four seconds. The agency spent considerably longer explaining it.

The three things worth actually checking

Nobody has time to re-verify an entire document line by line under deadline, and telling people to do that is how the advice gets ignored. A shorter, honest list survives contact with a Thursday afternoon better than a long one does.

• Numbers. Capacities, prices, dates, headcounts, percentages. Anything with a digit in it should be traceable to a source you can point to, not just "that sounded about right."
• Names. Client names, venue names, supplier names, job titles. These are the easiest things for an AI tool to blend or misremember from context, and the fastest thing a client will notice if wrong.
• Anything the client could check in ten seconds. A claim about their own industry, their own previous event, or a fact about their own sector. If they could disprove it with one search, assume they might.

That is a three-line habit, not a process. It fits on a sticky note. It is also, not coincidentally, close to what a careful proofreader would have checked before AI tools existed at all. The tool changed how fast a paragraph gets written. It did not change what is worth checking in it.

Who signs off what

The other gap I see is not the checking habit itself, it is that nobody owns it. If everyone is "supposed to check," in practice the person under the most deadline pressure is the one least likely to, and the document goes out anyway because someone assumed someone else already looked at it.

A cheap fix: name who signs off numbers and who signs off names, per document type, so it is not a diffuse responsibility that quietly belongs to nobody. A proposal does not need a committee. It needs one named person whose job it is to read the numbers before it leaves the building, even when, especially when, they did not write the paragraph themselves.

The sentence to watch for

"We always check that" is the sentence people say right before they describe the one time they didn't. It is not said dishonestly. It is said because checking used to be the default, back when writing was slow enough that reading it back happened almost by accident. AI tools removed that accidental read-through. If checking is going to keep happening, it now has to be a deliberate step, not a side effect of how slowly the document used to get written.

This week

Pick one document type your business sends out often, a standard proposal template is a good place to start, and write down, in one line, who checks the numbers and who checks the names before it goes out. Not a policy document. One line, pinned somewhere the team will actually see it. If you want help finding where else in the business this kind of unchecked step is already sitting, that is the sort of thing a Discovery Lab is built to find.

A month later, still thinking about the floor at ExCeL

I was at International Confex at the end of February, two days on the show floor at ExCeL, over three hundred suppliers, ten content theatres running back to back. It has taken me a few weeks to work out what actually stuck, rather than what I noticed in the moment.

What stuck wasn't a product. It was a pattern.

Every second stand had AI on it. Most of them meant the same thing

Walk the aisles and you'd think the whole industry had transformed overnight. AI powered this, AI enabled that, a chatbot avatar on more screens than I could count. Scratch the surface on most of them and the story was thinner than the banner: they had added a chatbot to an existing product, pointed it at their own FAQ, and called it AI.

That's not nothing. A well built chatbot saves someone a phone call. But it's a feature, not a transformation, and the stand graphics didn't always know the difference.

The interesting conversations weren't about AI, they were about data

The stands that actually held my attention were the ones where the sales conversation kept sliding sideways into something else: where does the data live, who owns it, what happens to it if we leave.

I heard a version of "where does our data go" more times in two days than I've heard it in the previous two years on a show floor. That's a genuinely new question for this industry to be asking suppliers out loud, and it's a better question than "does it have AI in it".

A couple of years ago the standard question at a stand was about price, or integration, or whether it played nicely with the CRM everyone already had. This year, for the first time, data handling sat right alongside those, asked early rather than as an afterthought once the deal was already half agreed.

Buyers are starting to understand, even if they can't always articulate it yet, that the AI feature is the easy part. The data behind it, who can see it, where it sits, what it's trained on, is the part that actually determines whether the tool is safe to use with a client's information.

I watched more than one sales conversation stall right there. The rep would run through the AI feature confidently, then hesitate the moment someone asked what happens to the transcript afterwards. Not because the answer was necessarily bad, but because it clearly wasn't the question they'd rehearsed for.

The theatre sessions were still stuck one step behind

I sat in on a handful of the content theatre sessions across both days. Good speakers, decent attendance, but the framing was almost always "what is possible with AI now" rather than "here is what we actually changed and what happened when we did".

That's not a criticism of the speakers, it's a fair reflection of where most of the industry genuinely is. Possibility is still the interesting story for a lot of people. But it means the room is having last year's conversation while the buyers wandering the floor outside were already asking this year's question.

What I see inside businesses tells a different story to what's on the stands

This is the bit that struck me hardest, because it's the opposite of what the floor would suggest. In every Discovery Lab I run, adoption inside a real events business looks nothing like a stand demo. It's individual. It's invisible. Nobody signed off on it and nobody is managing it.

One person on the team found a tool that works for them and quietly kept using it. Another has never opened one. There's no rollout, no training, no shared standard, just a scatter of personal habits that the leadership team usually can't see and rarely gets asked about.

The suppliers on the floor are selling a version of AI adoption that is planned, branded and demoable. The version actually happening inside most agencies, venues and supplier businesses is none of those things. That gap is where the real opportunity sits, and it has almost nothing to do with which chatbot you buy.

It's an odd thing to notice on a show floor built entirely around selling you the next tool: the tool was rarely the constraint. The constraint was that nobody had sat down and looked honestly at what the team was already doing, good and bad, before reaching for something new to bolt on top of it.

Three questions for the next stand with AI on the banner

Next time you're on a show floor and a stand has AI on the banner, skip the demo for thirty seconds and ask these instead.

• Where does our data actually go once it's in your system, and who else can see it.
• What changes for our team's day to day work, specifically, not generally.
• What happens to our data and our workflow if we stop using you in a year.

The answers will tell you more in two minutes than the demo will in twenty. A supplier who answers all three without flinching has probably thought about this properly. A supplier who steers you back to the demo hasn't, and that's worth knowing before you sign anything, not after.

If you want the same kind of honest look at where your own business actually is, not where the market thinks it should be, that's what a Discovery Lab is for.

The brief that looks finished isn't

Most client briefs read as complete. They have a date, a headcount, a vague sense of "premium" or "relaxed," a budget range that might be real or might be aspirational, and a list of things the client mentioned once in a call that never made it into writing.

You cost it anyway, because the alternative is going back with a list of questions before you have even sent a proposal, which can feel like you are already asking for homework. So the gaps get filled in with assumptions. Most of the time those assumptions are fine. Occasionally one of them is the exact assumption that erodes your margin three weeks before the event, when the headcount turns out to have meant something different to the client than it did to you.

What this prompt actually does

This is not a prompt that writes your proposal for you. It is a prompt that reads the brief the way your most sceptical colleague would, before anyone commits a number to a client. It looks for four things: gaps, contradictions, assumptions you are about to make without noticing, and anything that smells like scope creep dressed up as a small ask.

It then does something more useful than a list of red flags. It groups the questions by how much risk they carry to your margin, so you know which ones you genuinely cannot proceed without answering, and which are nice to know. And it drafts the actual message back to the client, written in their tone, so asking for clarity does not read as friction.

The prompt

Paste your client brief below the instructions, in full, into ChatGPT, Claude, Copilot, whichever you use. Expect back a structured breakdown: gaps grouped by risk, a list of assumptions it has spotted you making, a scope creep flag if there is one, and a client-ready message at the end.

You are helping an events professional stress test a client brief before it gets costed and quoted. I will paste a brief below. Do not write a proposal or a quote. Your job is to interrogate the brief.

Work through it in this order:

1. MISSING FACTS
List every piece of information a costing exercise would need that is not clearly stated in the brief: exact date and any flexibility, confirmed headcount and whether it is a hard number or an estimate, budget range or ceiling, the decision maker and who else needs to sign off, the venue or format if not specified, and any deadline for the proposal itself.

2. VAGUE LANGUAGE THAT NEEDS DEFINING
Quote every word or phrase in the brief that is doing a lot of work without being defined: "premium," "relaxed," "a few extras," "the usual," "nothing too corporate," and similar. For each one, note what it could reasonably mean at the cheap end and the expensive end, so the gap is visible.

3. CONTRADICTIONS
Flag anything in the brief that does not sit together logically: a stated budget that does not match a stated headcount and stated ambition level, a date that conflicts with something else mentioned, or requirements that pull in different directions.

4. ASSUMPTIONS I AM LIKELY TO MAKE WITHOUT NOTICING
List the assumptions someone experienced would probably make by default when filling this brief's gaps, and state plainly why each one is a guess, not a fact.

5. SCOPE CREEP RISK
Identify anything phrased as a small extra, a "quick add," or "while we're at it," that would actually require meaningful extra time, budget or supplier coordination. Explain why in one line each.

6. QUESTIONS TO SEND BACK, GROUPED BY RISK TO MARGIN
Produce two groups: "must answer before quoting" and "useful to know but not blocking." Keep each question short and specific, not open ended.

7. A DRAFT MESSAGE TO THE CLIENT
Using the tone of the original brief (formal, casual, warm, terse, whichever it is), write a short message asking the "must answer" questions in a way that reads as helpful and thorough, not as pushback. Do not apologise for asking.

Here is the brief:
[paste brief here]

What good output looks like

Run this on a real brief and the most useful part is rarely the list of missing facts, most people can spot those. It is the contradiction and assumption sections. A brief that mentions "up to 200 guests" and a budget that only really works at 120 is the kind of thing that is easy to miss when you are reading quickly and keen to get a proposal moving.

A conference venue in the north west started running this on inbound enquiries before costing them, and the most common flag it raised was not a missing number, it was a phrase like "similar to last year but bigger," where nobody had actually confirmed what last year cost or what bigger was supposed to mean.

Before you paste anything in

Strip client names, company names, and anything commercially sensitive before you paste a brief into a public AI tool. Replace names with a placeholder like "the client" and remove anything you would not want appearing in a training set or a screenshot. The interrogation works exactly as well on an anonymised brief. There is no reason to send more than the prompt actually needs.

Where this fits in the week

The best time to use this is the moment a brief lands, before anyone starts building a proposal around it. Five minutes here, before the costing starts, is cheaper than the conversation you have to have with the client three weeks before the event when the assumption turns out to have been wrong.

If you want a wider set of tested prompts like this one for events work, they live at Zoby OS.

The prevailing narrative surrounding Artificial Intelligence often focuses on the 'arms race' for adoption. In the events and hospitality industry, the pressure to modernise operations is intense. However, a critical oversight is emerging: the belief that AI regulation is a future problem rather than a present reality.

For leadership teams, understanding the intersection of digital infrastructure and global compliance is now a strategic necessity.

The Myth of the Regulatory Horizon

Many organisations operate under the impression that AI is currently a 'wild west' and that rules will be codified in several years. This is a misunderstanding of the legal landscape.

For any business operating in the UK or Europe, the framework for AI governance is already established. The EU AI Act is active, impacting any organisation that provides or uses AI systems within the EU market. Simultaneously, the UK GDPR remains the standard for anyone processing personal data, regardless of whether that processing is done by a human or an algorithm.

What many designate as 'experimentation' is, in the eyes of regulators, 'processing'.

High-Risk Exposure Points

Regulatory exposure often hides in standard operational workflows. In our industry, three areas are particularly vulnerable:

**Internal Strategy and Data Leakage**
Using public LLMs (Large Language Models) to summarise confidential client contracts or draft sensitive internal strategies often inadvertently places proprietary data into the public domain, violating both privacy laws and client NDAs.

**Automated Human Resources**
Using AI to screen CVs for event staffing or to automate hiring decisions is now subject to strict transparency requirements. If a candidate cannot be told exactly how an algorithm assessed them, the employer is at risk.

**Behavioural Profiling**
Analysing attendee data to predict future booking patterns is a staple of modern hospitality. However, when AI performs this profiling without a clear 'human-in-the-loop' or explicit data consent, it may breach current protection standards.

Speed vs Infrastructure

The competitive advantage in 2026 will not belong to the companies that adopted AI the fastest. It will belong to the companies that built the sturdiest guardrails.

Modernisation requires more than just new tools; it requires a sophisticated digital infrastructure that accounts for data sovereignty and ethical AI use. Leadership teams must shift their focus from 'what can this tool do' to 'how does this tool fit into our compliance framework'.

Practical Steps for Governance

To mitigate risk, directors should consider the following:

1. Conduct an audit of all 'shadow AI' currently used by staff.
2. Implement formal policies on the input of sensitive data into third-party tools.
3. Consult the UK Information Commissioner’s Office (ICO) guidelines on AI and data protection to ensure current workflows are compliant.

Innovation is essential for growth, but it must be sustainable. In the current regulatory environment, the most confident move a leader can make is to pause, assess, and build with purpose.

What actually happened on 5 February

Earlier this month, new sections of UK GDPR came into force under the Data (Use and Access) Act 2025. They replace the old Article 22, which was built around a fairly blunt idea: a decision made "solely" by automated means, with legal or similarly significant effect on someone, was restricted by default.

The new version is more permissive. It gives organisations more room to actually use automated decision making, provided certain safeguards are in place. This is not a crackdown. If anything it is the opposite: the door has opened a bit wider. What has changed is what you owe the person on the other end of the decision.

I am not a lawyer and this is not legal advice. The point of this post is not the fine print, it is the fact that most events businesses are already doing the thing the law is now talking about, and have not necessarily noticed.

Where automated decisions already live in an events business

Nobody in events tends to say "we use automated decision making." But look at what is actually running:

• A CRM that scores inbound leads and quietly deprioritises the ones under a threshold, so a sales person never sees them.
• A sponsorship or exhibitor application form that auto-declines anyone who does not meet a stated criterion, with no human looking at the borderline cases.
• Dynamic pricing on ticketing that raises or drops a price band based on sales velocity, with nobody reviewing individual outcomes.
• Registration screening that flags or blocks a delegate signup based on a rule set, sometimes with an AI layer added on top to catch duplicates or fraud.
• A chatbot on the website that can say no to a refund request without a human ever reading the message.

None of these were built with Article 22 in mind. They were built to save someone an afternoon. That is exactly the pattern the new rules are aimed at.

A conference venue in the north west found this out the practical way. Its exhibitor application form had quietly auto-declined anyone who left one field blank, since the launch of the form two years earlier. Nobody had decided that on purpose. Someone had ticked "required field" in a form builder and moved on. The decisions it made were real, even though nobody had ever called them decisions.

What the safeguards actually mean in practice

Stripped of the legal language, the shift is roughly this: you get more freedom to automate, in exchange for three things people can rely on.

• Being told it happened. Someone whose lead score, application, price or refund was shaped by an automated process should be able to find that out, not discover it by accident.
• Being able to ask for a human. There needs to be a real route to a person, not a contact form that also gets triaged by the same system.
• Being able to contest it. If the decision was wrong, there is a way to say so and have it looked at again by someone with the authority to change it.

A dynamic ticket price is a low-stakes example. An exhibitor application that gets auto-declined the week before a show, for a company that would have been a good fit, is not.

Why this is worth an hour now rather than later

Most of the automated decisions above were never signed off as automated decisions. They were a CRM feature someone switched on, a form logic someone built in an afternoon, a chatbot script someone wrote to stop the same refund question landing in the inbox forty times a week. Nobody sat in a meeting and decided "we are now making automated decisions with legal effect on people."

That is normal, and it is also exactly the gap the new rules are pointing at. The businesses that will find this easy are the ones who already know where their automated decisions sit. The ones who will find it uncomfortable are the ones who find out when someone asks, in writing, why they were declined.

A one-page exercise: where are our automated decisions

This does not need a compliance project. It needs a list. Paste the prompt below into your AI assistant along with a rough description of your CRM, your booking or ticketing system, your application forms and any chatbot or auto-reply tool, and ask it to help you build the first draft of the register. Expect back a simple table you can hand to whoever owns each system.

I run an events business (agency, venue or supplier, describe which). Help me build a one-page register of where we might already have automated decision making running, in the sense used by UK GDPR (a decision made by a system, with no meaningful human review, that affects a person's opportunities, price, access or outcome).

I will describe our systems below. For each one, tell me:
1. Whether it plausibly counts as an automated decision under this definition, and why.
2. What the likely effect on a person is (an opportunity lost, a price set, an access decision, a refusal).
3. Whether a human currently reviews individual outcomes, or only monitors the system in aggregate.
4. One practical question I should be able to answer if someone challenged this decision.

Do not give legal advice or quote specific article numbers. Keep this practical and specific to what I describe.

Our systems:
[describe your CRM lead scoring, application forms, ticketing or pricing tool, registration screening, and any chatbot or auto-reply system]

What to do with the answer

Once you have the list, the fix is rarely technical. Most of the time it is a line added to a form ("this application is assessed automatically against these criteria, you can ask for a human review"), a genuine route to a person for the chatbot, and someone named as the person who reviews a contested decision. None of that requires new software.

If you want a second pair of eyes on where this sits in your systems specifically, that is the kind of thing a Discovery Lab tends to surface early, usually in the first few conversations, well before anyone gets to the written report.

What people say when the door is closed

A Discovery Lab is a set of structured 1:1 conversations with everyone on a team, followed by a leadership session and a written report. I have now sat through enough of these inside event agencies to notice the same handful of things coming up, almost word for word, whoever the client is.

What follows is a composite. No client names, no identifying detail, because the value of these conversations depends on people trusting that what they say stays anonymous. But the patterns are real, and they repeat.

"We don't really use AI yet"

This is what leadership tells me at the start of almost every engagement. It is usually said with a slight apology, as if the business is behind and needs to catch up.

Then I talk to the team, one at a time, and the picture changes completely. People are using AI daily. Drafting supplier emails. Summarising long client threads before a call. Getting a first pass at a run sheet done in ten minutes instead of an hour. They are just doing it on personal accounts, on their own phones, quietly, because nobody told them they could not, and nobody told them they could either.

The gap is not adoption. The gap is visibility. Leadership is often the last to know what is actually happening on the ground, which is exactly why a survey sent from the top rarely captures it.

The person who "just knows"

Almost every agency has one. Usually not the most senior person in the room. The one who understands why the quoting spreadsheet has that odd column in it, why a particular supplier always gets a manual price adjustment, why a certain client's briefs need translating before anyone else can work from them.

Nobody wrote this down. It lives in one person's head, and everyone else has quietly learned to route around it rather than ask, because asking feels like admitting the process is broken. It is a serious risk hiding behind a running joke: "oh, just ask Sarah."

The forty-minute Tuesday

This is the line that comes up more than any other, in one form or another. Someone describes a task, dismissively, as just something they do. Not a problem. Just Tuesday.

Except when you sit with it, it turns out to be forty minutes of manually cross-checking a spreadsheet against an inbox, every single week, because two systems do not talk to each other and nobody has ever been given the time to fix that. The person doing it has stopped seeing it as a problem, because they invented the workaround themselves, years ago, and a workaround you invented yourself does not feel like a broken process. It feels like competence.

Nobody agrees what "confirmed" means

Ask five people in the same agency what stage an event needs to reach before it counts as confirmed in the CRM, and you will often get five different answers. Sales counts it as confirmed once a verbal yes has been given. Finance counts it once a deposit has landed. Delivery counts it once a signed contract exists.

None of them are wrong. But a shared system built on an unshared definition produces reports that everyone privately distrusts and nobody wants to be the one to say so out loud.

The best ideas come from the person nobody asked

Leadership sessions tend to generate leadership ideas: strategic, structural, sensible. The best practical fix I have heard in a Discovery Lab so far came from a junior account exec, about twenty minutes into an unremarkable 1:1, offered almost as an aside. It solved a handover problem that had been quietly costing the delivery team hours every month, and it had never once come up in a management meeting, because nobody had ever asked that person directly.

Why a conversation beats a form

A survey gets you what people are willing to write down, with their name attached, knowing their manager might read it. A structured, anonymous 1:1 gets you what is actually happening.

That difference is the entire point. You cannot fix a workaround you do not know exists, and people will not tell you it exists in a form with a submit button on it.

When I started running these, I expected the interesting findings to be about software: which CRM people hated, which spreadsheet was one bad update away from collapse. Some of that is real, and it matters. But it is not what stays with me afterwards. What stays with me is how often someone says a version of "nobody's ever asked me that before" about a job they have done for years. Not asked in an appraisal, where the question is really about them. Asked about the work itself: what actually happens, in order, when this lands on your desk. That tends to open people up in a way that catches them a little off guard.

It also explains why the write-up at the end of a Discovery Lab rarely feels like a surprise to the team once they read it. They already knew most of it. What they did not have was a reason to say it out loud in one place, gathered together, where leadership could finally see the whole shape at once.

If you want to know what your team would say in a room where the boss genuinely is not listening, that is what a Discovery Lab is built to find out.

The government thinks it's moving. Ask what your team is already doing

This month the government put out its one year progress check on the AI Opportunities Action Plan. Thirty eight of fifty recommendations met, on their own scorecard. Whatever you make of that number, the direction of travel is clear: this is now infrastructure policy, not a side project.

Most events businesses I sit down with have no policy at all. Not a bad one. None. And in that gap, people have quietly made their own decisions.

That gap isn't unique to events, but it shows up in a particular way here, because so much of what this industry handles is sensitive by nature. Client budgets. Attendee lists. Unreleased programme details. Supplier pricing that would cause a headache if a competitor saw it. When shadow AI use meets that kind of material, the stakes are higher than a rewritten email, even if the rewritten email is where it starts.

The account manager, the producer, the sales exec, the designer

Here is what I actually find when I run a Discovery Lab and start asking people what they do on a normal Tuesday.

• The account manager rewrites client emails through ChatGPT on their own personal login, because the tone comes out better and faster than anything the brand template gives them.
• The producer drops a recorded client call into a free note taking tool to get a summary, because writing up notes after a two hour briefing used to eat their evening.
• The venue sales exec pastes a draft contract into an assistant "just to check it reads properly", not really thinking about what else is on that page.
• The designer generates mood board images because it is quicker than three hours on stock sites, and the client never asks where they came from.

None of this was approved. None of it was hidden, either. Nobody asked, so nobody told.

Why this isn't the telling off you think it is

The instinct, when you find this, is to tighten up. Send a memo. Remind everyone about the acceptable use policy that doesn't exist yet. I'd stop you there.

People didn't go looking for a shortcut because they're lazy. They went looking because the tools you gave them officially are slower than the ones they found themselves. That is not a people problem. That is a systems problem wearing a people costume.

If your approved software is worse than what's free on someone's phone, shadow AI isn't a rebellion. It's a rational choice made forty times a day by people trying to get home for dinner.

Blame doesn't fix that. A better default does. If the sanctioned way of working is genuinely the easiest way of working, most people will take it without being asked twice. Make it the harder option and you'll be fighting this quietly forever, one personal login at a time.

The one hour amnesty conversation

Before you write a single rule, have one conversation. Not an audit, not a disciplinary, an amnesty. Tell the team plainly: nobody is in trouble, I want to know what you're actually using and why, because I'd rather know than guess.

Run it as a straight question round the table or one to one, whichever suits your team. You will hear things you didn't expect. You will also hear the same three or four tools coming up again and again, which is useful, because it tells you where the real appetite is.

Two columns: fine, and not with client data

Once you know what's being used, sort it fast. Two columns on a whiteboard is enough.

• Fine. Drafting internal copy. Brainstorming session titles. Tidying up a rough agenda. Nothing that names a client or carries their numbers.
• Not with client data. Contracts. Attendee lists. Anything with a supplier's pricing on it. Anything the client would be unhappy to learn left your building through a tool you don't control.

That split alone removes most of the actual risk without removing any of the actual usefulness.

Pick one paid team account

Then do the boring, unglamorous thing that fixes the root cause: pay for one proper team account, on one tool, for everyone. It doesn't need to be exotic. It needs to be sanctioned, secure, and no worse to use than what people already found on their own.

This is usually the cheapest decision in the whole business. A single seat licence across a team costs less than one bad week caused by a client contract that went through the wrong window.

Three sentences people can actually remember

Skip the twelve page policy document nobody will read. Write three sentences and put them somewhere people will actually see them.

Something like: use the team account, not your own login. Nothing with a client's name or numbers goes into any AI tool unless it's on the approved list. If you're not sure, ask, you won't be told off for asking.

That's a policy your team can hold in their head on a Tuesday afternoon, which is the only kind that works.

If you want a proper look at what's actually happening inside your team, not what you assume is happening, that's the whole point of a Discovery Lab. And if you want to build a team's confidence with these tools before you write a word of policy, Zoby OS is free.

The January reflex

Every January, event businesses do the same thing. Someone comes back from the Christmas break with a bit of energy, looks at the year ahead, and decides this is the year things get modern. A CRM gets shortlisted. A new proposal tool gets a free trial. Someone asks whether AI can "help with the admin."

None of that is wrong. But it is usually early.

Buying a tool before you understand the process it sits on top of is how a forty-person agency in the Midlands ends up with three CRMs running at once, each one championed by a different department, none of them trusted enough to be the single source of truth.

The tool is not the problem

Almost every event business I have sat down with has a version of the same shape underneath. A brief comes in. Someone costs it. A quote goes out. It gets won. It gets handed to delivery. The event happens. Someone reconciles the numbers afterwards, usually late, usually from memory as much as from records.

Each one of those handoffs is a place where information gets rekeyed, guessed at, or lost. New software rarely removes a handoff. It just gives the same handoff a nicer interface. If nobody has written down what actually happens at each step, the new system inherits the same gaps as the old one within about six weeks.

This is the bit that gets missed: a workflow nobody has documented is a workflow nobody can fix. You cannot improve what you have never actually looked at end to end.

A four-step audit you can run this week

This takes half a day, not a project plan. You need one job, one person who does it weekly, and a notepad.

• Pick one job that happens every week. Not the whole business. One thing: brief intake, quoting, supplier confirmation, invoice reconciliation. Pick whichever one makes people sigh when it comes up.
• Walk it end to end with the person who actually does it. Not the manager's version of how it works. The real version, including the workaround nobody put in the process doc because there is no process doc.
• Count the handoffs and the re-keying. Every time information moves from one person, system or spreadsheet to another, that is a handoff. Every time someone retypes something that already existed somewhere else, that is re-keying. Write down the number.
• Put a rough time cost on it. Not a scientific one. Ask the person doing it: how long does this take you, on a normal week versus a bad week. Multiply by how often it happens. You will have a number nobody in the leadership meeting has ever said out loud.

What this actually finds

Run this on a quoting process and you will usually find the same things. A spreadsheet that only one person really understands. A capacity figure for a venue that exists in two places and disagrees with itself. A "final" version of a proposal that gets edited after it was marked final. None of these are software problems. They are structure problems wearing a software costume.

A conference venue in the north west ran this audit on their event handover process last year and found that the same booking details were typed into four different places by three different people, and that the version delivery actually used was the one with the most typos, because it was the one that got updated last.

That is not a people problem. Nobody there is careless. It is what happens when a process grows organically for a decade with no one owning the whole shape of it.

Where AI actually fits

Once you can see the workflow clearly, AI becomes genuinely useful, because you can point it at a specific, bounded step rather than the vague hope that it will "help with the admin." Summarising a brief. Drafting the first pass of a quote from a template. Flagging when a capacity figure looks inconsistent across documents. These are small, unglamorous wins, but they are real ones, and they only work once the underlying process is clear enough to automate a piece of it without automating the chaos along with it.

Skip the audit and you get an AI tool confidently doing the wrong version of the process, faster.

This is the mistake I see most often with in-house event teams especially. Someone senior asks the whole department to "start using AI more." Without a clear picture of where the actual bottleneck sits, everyone reaches for the same generic assistant and applies it to whatever is in front of them that day. A few people find something genuinely useful. Most drift back to the old way within a month, because the tool was never aimed at the step that was actually costing time.

Fix the plumbing before you buy the tap

None of this needs a consultant, a project code, or a slide deck. It needs half a day, one honest conversation with the person who does the job, and a willingness to write down what you find even when it is unflattering.

Do that before you sign anything new this year. The tool you eventually buy, if you still need one, will be a much better fit once you know exactly what it needs to fix.

If you want a structured version of this across a whole team rather than one process, that is what a Discovery Lab is for.