The Big AI Secret · ongoing research

The agency has AI rules. Its speaker cannot explain them.

This month’s evidence separates widespread AI use from the ability to manage it consistently. For agency directors, the gaps concern shared rules, delivery capacity and proof of productivity, not simply access.

7 October 2026 · 6 minute read

A policy can exist without being understood. Permission to experiment can exist without capacity to build. Work can feel faster without anyone being able to prove a productivity gain. Those distinctions run through this month’s material. This guide brings the agency findings together, with evidence from retail showing where similar tensions appear outside client services.

The calls describe particular businesses, not a representative sector survey. The live Integration Index provides wider context, but its figures cover businesses across sectors. They are not agency-only results or a count of this month’s participants. Keeping those boundaries visible matters: a vivid account can expose a useful question without establishing how common the problem is.

A written rule is not yet a shared rule

“Yes, but I couldn't really tell you. It wasn't written by me.”Interview participant, agency and marketing, 100+ staff

This answer concerns written rules about what AI may touch. It does not establish that the agency lacks controls, or that staff ignore them. It establishes something narrower: the speaker cannot explain the rules. For a director, that is a reason to test understanding rather than treat the existence of a document as sufficient evidence of governance.

“We do have a firm rule about nothing that's been created by AI or built with the help of AI should be shared with a client before it has been checked and edited by a human.”Interview participant, agency and marketing, 100+ staff

The contrast is useful. Here, the speaker can state the boundary and the required action. That still does not prove compliance, but it gives the business something observable to check. A general statement about responsible use is harder to test than a requirement attached to the moment work reaches a client.

The practical check is not whether colleagues remember policy wording. Ask them what they would do with confidential material, an unfamiliar tool or a draft ready for publication. Then compare their answers with the written rules. Differences reveal where guidance needs to become clearer, easier to find or better connected to ordinary work.

Widespread use does not establish consistent capability

“It's nearly everybody, but I think each individual will use it to varying degrees based on their own comfort.”Interview participant, agency and marketing, 25-100 staff

Breadth is a recurring strength in the agency calls. Yet this answer contains its own qualification. Being a user says little about the work someone can complete, the support they need or the consistency of the result. Confidence and experience can differ even where access is widespread.

“Everyone, every single person in the team, every day.”Interview participant, retail and e-commerce, 100+ staff

The retail account makes the distinction especially clear. Despite daily use across the team, the accompanying evidence describes AI producing drafts while people complete the remaining work. The agency article on pitch decks showed a different capability, a finished deck, but still found manual transfer afterwards. Neither frequency nor output quality alone describes the full operating process.

These Index shares are context, not an explanation of the interview accounts. For an agency director, the useful move is to look beyond who uses AI. Examine a recurring task from brief to accepted output. Identify what is repeatable, what depends on personal skill and what a colleague could reproduce without the original author beside them.

Permission is not the same as delivery capacity

“So if somebody's got an idea, don't come to me with an idea, go and figure it out, test it, and then come and show me the shiny stuff once you've got something to look at.”Interview participant, agency and marketing, under-25 staff

This is a clear invitation to experiment. It removes the need to win approval before demonstrating something useful. But permission answers only the question of whether someone may start. It does not tell us who will test reliability, maintain the process or make it available to colleagues once the demonstration works.

“He builds all of it. He's good and he's fast, but he is one person and he's also running the studio.”Interview participant, agency and marketing, 25-100 staff

In this account, ideas depend on the head of studio operations, alongside the demands of running the studio. The constraint is not a lack of enthusiasm or technical ability. It is competing responsibility. Encouraging more ideas will not necessarily improve delivery if those ideas all reach the same crowded desk.

The wider material includes stalled experiments and staff using learning time to develop approaches. These are different arrangements, not a universal pattern. They suggest a useful management check: name who can test an idea, who can put it into routine use and who supports it afterwards. If those responsibilities are unclear, apparent freedom can conceal a queue.

A faster feeling is not productivity evidence

“It feels faster than it was, and everyone says it's faster, but I could not tell you by how much. If you asked me to prove we'd saved anything, I couldn't.”Interview participant, retail and e-commerce, 100+ staff

This is not evidence that AI has failed. It is an unusually clear statement of the evidence limit. Positive experience can justify further testing. It cannot, on its own, support a reliable claim about the scale of productivity improvement or its effect on business performance.

“I'd say it's more anecdotal than empirical, but, but you know, it's certainly something that...”Interview participant, agency and marketing, 25-100 staff

The agency evidence also contains stronger measurement practices. Campaign versioning is checked through a recurring job profitability report in one account. Elsewhere, pitch production was tracked initially, then tracking stopped. These examples matter because they distinguish an ongoing management measure from an early test. Neither establishes the return from every other AI-supported activity in the business.

For directors, the next step is a transparent record of a defined task. Include preparation, editing, checking and rework, not just generation. Record whether the accepted output has changed in quality or scope. Otherwise, a quicker draft and a more demanding final deliverable can become mixed together in a claim that nobody can properly assess.

Find your stage

The Integration Index assessment shows your current stage across Explorer, Implementer, Integrator and Transformer. Its questions examine breadth of use, autonomy, flow, access, governance and measurement. That makes it a way to look beyond the most impressive demonstration and consider how AI actually operates across the business.

Take the assessment using what happens in practice, not what the policy or plan says should happen. Where an answer rests on an impression, note that uncertainty. Your stage gives you a starting point for deciding what to examine next: shared understanding, delivery capacity, repeatable work or evidence of productivity.