The interesting number in this year's CMO spend research is not the one that got quoted.

The quoted number is the allocation, and allocations make for clean headlines. The number that should change how event teams plan is the one sitting next to it, describing how few organisations can actually operate what they have already paid for.

That gap is where budgets get quietly destroyed, and it is largely invisible until the invoice arrives.

The allocation already happened

Gartner's 2026 CMO Spend Survey, covering 401 CMOs across January to March 2026, found that CMOs now allocate 15.3% of marketing budget to AI while only 30% are ready to scale AI capabilities.

Take the first figure seriously for a moment. Fifteen percent of a marketing budget is not an experiment. It is not a pilot line or an innovation fund. It is a major channel, comparable in scale to what many organisations spend on events themselves.

That money is already committed. Whatever debate an organisation thinks it is still having about whether to invest in AI, the survey says the investment happened, and it happened at a scale that displaced something else.

For anyone running a show, this is the more useful framing. Your event sits inside a marketing budget where roughly one pound in seven now goes somewhere it did not go two years ago. It did not come from nowhere.

The gap between allocating and scaling

Only 30% ready to scale is the finding that should govern planning, and the wording repays attention.

Ready to scale is not the same as using. Organisations in the other 70% are using AI. They have the licences, the assistants, the pilots that went well. What they do not have is the ability to take a capability that worked in one place and run it reliably across the business.

That is a specific and recognisable condition. Someone on the team built something genuinely useful. It works. It cannot be extended to the other four processes it would obviously help, because it depends on that person, on manual steps nobody documented, and on data that happens to be clean in one system and not the others.

The organisation is now paying for a capability at full price while accessing a fraction of it. And because the pilot succeeded, the internal narrative is that AI is working, which suppresses exactly the questions that would surface the problem.

Seventy percent of organisations are in that position. It is the majority condition, not an edge case, and it is worth assuming you are in it until you can demonstrate otherwise.

Why agent costs do not behave like software costs

Here is the part most budget processes are structurally unprepared for.

MarTech's analysis of the arithmetic behind an AI workforce sets out where the money actually goes, and it is not where procurement is looking. Agent costs sit outside standard pricing models. API volume fees. Middleware. Integration hours. Background token burn.

Every item on that list behaves differently from a software licence, and the difference matters.

A licence is a fixed, annual, predictable number. You negotiate it once, you know what it costs, and using the product more does not change the bill. Every budgeting process in marketing was built around that shape.

API volume fees are consumption-based, which means cost rises with use. Middleware is the connective software nobody scoped because it is not the thing being bought. Integration hours are labour, usually unbudgeted, frequently absorbed by someone's existing role. And background token burn is the one that surprises people most: agents that run on a schedule, checking, monitoring and processing, consuming budget while nobody is watching them work.

The last one has no analogue in traditional software at all. Nothing in a marketing stack has previously cost money while sitting idle in the way an always-on agent does.

What this looks like on an event calendar

Event organisations have a spending pattern that interacts badly with consumption pricing, and it is worth being explicit about why.

Event work is spiky. Registration opens and volume surges. The six weeks before a show carry more communications, more data processing and more exhibitor activity than the six months preceding them. Then the show happens, follow-up runs hot for a fortnight, and activity drops to a fraction of peak.

Under licence pricing, that pattern is free. You pay the same in the quiet month as in the peak month, and the peak is effectively subsidised by the trough.

Under consumption pricing, the spike is the bill. Your highest-cost period is also your highest-stress period, and it arrives at the same time as venue payments, contractor invoices and everything else that clusters around a show date.

Teams that modelled their AI costs on an average month will find the peak month does not resemble the average at all. And the peak is not an anomaly to be smoothed. It is the entire operational point of the business.

The planning implication is that AI cost for an event organisation should be modelled per event cycle, not per month. Anyone budgeting this as a flat monthly line is going to be wrong in a specific direction, twice a year, in the weeks they can least afford a surprise.

The headcount that paid for it

There is an uncomfortable question underneath the 15.3%, which is what it displaced.

Part of the answer appears in the same period's labour data. A Stanford study, discussed by Social Media Examiner, found roughly a 20% reduction in headcount for sales and marketing roles held by 22 to 25 year olds.

The trade is fairly legible. Budget moved from junior capacity into tooling, on the reasonable expectation that the tooling would cover the work.

Where that reasoning breaks is the 70%. The trade only pays if the capability actually scales. In an organisation that cannot extend a working pilot beyond the person who built it, the junior capacity is gone and the tooling covers one process instead of five. The saving was banked. The coverage did not arrive.

This is the most expensive version of the readiness gap, because it is not recoverable inside a budget cycle. You cannot re-hire the coordinator in November when the show is in January. The market-level version of this trade is set out in why the AI-native agency is emerging now.

What a show actually pays for

Translating this to an operating question, there are three costs an event organisation should be able to state and most currently cannot.

Cost per event cycle rather than per month. What did AI actually cost across the eight weeks around your last show, including the consumption spike. If the answer is a monthly average multiplied by two, it is not an answer.

Cost of the integration that has not happened. The hours currently spent moving data between systems because the connection was never built. This is a real cost, paid in salary rather than invoice, and it is the cost the 15.3% was supposed to eliminate.

Cost of the capability that did not scale. The pilot that works in one place. What would it be worth running across every process it applies to, and what specifically is preventing that. Usually the blocker is documentation or data consistency rather than money, which makes it cheap to fix and easy to never get around to. That distinction between buying a capability and running it is the core of what an AI-native agency is.

The third one is where most of the recoverable value sits, and it is almost never on a budget line because it is not a purchase.

Budgeting for the 70%

If most organisations are on the wrong side of the readiness split, the useful planning posture follows from assuming you are too.

Budget integration hours explicitly, as labour. Not as a project with a business case, which will lose to anything with a deadline attached. As a standing allocation of someone's time. The organisations that never fund this are not the ones that decided against it. They are the ones for whom it was never a line.

Model consumption against your peak, not your average. For events that means the eight weeks around a show. Ask any vendor pricing on consumption what a peak cycle costs, with your volumes, and treat a vague answer as information.

Instrument the background work. Scheduled agents consuming tokens while nobody watches is the cost item most likely to drift. Anything running on a schedule should have someone who can say what it costs per month and what it produced.

Judge the investment on scaling, not on pilots. A successful pilot is the normal condition of the 70%. The question that separates the groups is whether anything moved from one process to five, and it is worth asking at review time rather than accepting the pilot as evidence.

The allocation decision has been made across the industry. Fifteen percent of marketing budget is going into this whether or not any individual organisation feels ready. What remains open is whether a given team ends up in the 30% that can operate what it bought, or the 70% paying full price for a fraction of it. That outcome is decided by unglamorous work on integration and documentation rather than by which tools were selected, which is why so few organisations are doing it and why the ones that do will look inexplicably efficient by next year.

If you want to work out what your AI spend actually costs across an event cycle, book a call with TalkValue.