The Big AI Secret · ongoing research

Agencies adopted fastest and hit the wall first

Marketing and creative agencies were ahead of every other sector in our research. They are also the clearest illustration of why being ahead on adoption does not mean being ahead.

9 September 2026 · 8 minute read

They moved first, and they moved properly

Agencies integrated AI into content workflows faster than anyone. Research, first drafts, optimisation, performance analysis — across the whole development cycle rather than one corner of it.

The commercial pressure explains most of it. An agency billing by the hour or the project has an immediate, obvious reason to make production faster. The feedback loop is short and the incentive is direct.

So agencies did what the rest of the market is still working up to. They adopted broadly, quickly, and with real intent. And then they found the ceiling.

Then the tools multiplied

One digital agency in our research had deployed twelve separate AI applications across its departments. Their operations lead described the result:

“We're producing more content than ever, but it takes longer to manage the process than it did before AI.”Operations lead, digital agency

That sentence is the whole chapter. Output went up. The work of coordinating the output went up faster.

It is worth being precise about why, because the instinct is to read it as a discipline problem. It is not. Every one of those twelve tools was adopted for a good reason, by someone solving a real problem, and each one worked. The failure was not in any of them individually. It was that nothing connected them, and the cost of that only becomes visible once there are enough of them for the connecting to become somebody's job.

The five costs nobody puts in a spreadsheet

The pattern repeated across sectors, and agencies hit it first because they got there first. Five costs showed up consistently, and none of them appear on a licence invoice.

Duplicated effort. Teams recreate assets, prompts and templates that already exist somewhere else in the business. Nobody knows they exist because there is nowhere they all live.

Insight that does not compound. Something learned in one tool never reaches another. The second campaign is no better informed than the first, which removes the main reason to use AI on either.

Compliance surface. Every integration point is another question about where data sits and who can see it. Twelve tools is twelve conversations, and most organisations have had none of them.

Context switching. Staff move between logins, interfaces and mental models all day. The switching cost is invisible individually and substantial in aggregate.

No single version of the truth. Leadership cannot say what any of it is producing, because the answer is spread across twelve dashboards that do not agree.

And then governance drifts

Disconnection does not only slow things down. It quietly moves work outside the rules. One global agency found that two-thirds of its AI activity was happening outside approved systems. Their CIO was clear about why, and the sentence has stayed with us:

“It wasn't malicious. It was impatience. People were trying to connect what the stack didn't.”CIO, global agency network

Shadow AI is not usually a discipline problem. It is a plumbing problem wearing a discipline problem's clothes. Which matters for what you do about it. If you read two-thirds of activity happening outside approved systems as a compliance failure, you write a stricter policy and the number goes up, because the workaround is still faster than the approved route. If you read it as a plumbing failure, you fix the route.

The audit: what are you actually running?

Before consolidating anything, count. Most organisations we spoke to could not name their own tool estate, and the count is usually higher than leadership expects by a factor of two or three. This takes an afternoon and it does not need a consultant.

1 · List every AI tool anyone is paying for. Start with the company card statement and the software subscriptions. Then ask each team lead what they use. Then — and this is the one that finds the rest — ask what people use that the business does not pay for.

2 · For each one, write down what it does and who owns it. If two tools do substantially the same thing, mark them. If nobody owns one, mark it. Both are consolidation candidates before you have looked at anything else.

3 · Count the handoffs. For your three most common workflows, count how many times a person moves output from one tool into another by hand. That number is the coordination cost, and it is what the eighteen per cent was made of.

4 · Ask one question of each tool: does anything leave it automatically? If the answer is no for all of them, you do not have an AI stack. You have twelve applications, and the work of joining them is being done by people.

5 · Then ask what would break if you turned each one off tomorrow. In most estates, several answers are "nothing" or "one person would be annoyed". Those are the easy ones and they are worth removing before anything harder is attempted.

What the ones who solved it did

The sequence they followed is worth copying, because the order matters more than the destination.

They picked the workflow, not the tool. The question was never "which platform" but "which piece of work do we want to run properly". The platform decision followed from the workflow rather than preceding it.

They moved one workflow first, completely. Not twelve tools migrated at ten per cent each. One workflow, end to end, working, before the second was started.

They kept the shared library from day one. Prompts, templates and assets in one place, from the first workflow rather than added later. This is the part that makes the second workflow faster than the first.

They wrote the rules while they were building, not after. What data can go where, who can approve what, what stays human. Written during, which is when the answers are obvious, rather than retrofitted.

They measured the first one before starting the second. Which is what funded the second, and the eight after it.

The lesson for everyone behind them

Agencies are not a special case. They are an early one. Every sector adopting AI now is walking toward the same wall, and the agencies hit it first only because they set off first.

The Big AI Secret is an ongoing research programme into how AI is really being used inside British businesses. Edition 1 draws on over 1,500 in-depth interviews with business leaders across sectors. The method, including its limitations, is published in full at thebigaisecret.com/method.