The Big AI Secret · ongoing research

One agency's AI rule: leave precise creative work alone

This month's agency evidence points towards selective automation, not a race to remove people from every task. The useful distinction is between human work that protects quality and human work that merely keeps the process moving.

9 October 2026 · 6 minute read

An agency's AI rules tell staff not to use it when creative direction needs to be precise. That is a useful counterweight to accounts of AI producing complete deliverables. A process can become more capable without becoming suitable for every brief. For agency directors, the question is not simply whether AI can produce the work. It is whether the agency knows when to use it, where to stop it and what happens next.

This month's published article examined the gap between generating a finished deck and moving it through the business. Read alongside the wider call evidence, it suggests a practical guide to agency progress: define the boundary, inspect the handoff, support repeatable use and test the claimed return. These are accounts of particular workflows, not proof of how the whole sector operates.

Set the boundary before expanding the use

The broad Index figures provide context, not an agency benchmark. The calls offer something more specific: examples of where agencies place limits. In a creative workflow, a rule against AI can be a sign of deliberate selection rather than reluctance. If the brief requires precise control, generating alternatives may create work rather than remove it.

“And that includes things like if you want finite direction on a piece of creative output, don't use AI.”Agency representative, agency and marketing, 25 to 100 staff

Elsewhere, the boundary comes after generation. AI produces campaign versions, but an account manager decides whether they can leave the business. Completion and permission are separate. A complete set of assets is not necessarily an approved set.

“It makes the whole set, but nothing leaves the building until an account manager has been through it and pressed approve.”Agency representative, agency and marketing, 25 to 100 staff

These accounts should not be collapsed into a single model of agency practice. They describe different choices. Together, they show why the amount AI produces is an incomplete measure of progress. A director needs to know which briefs suit generation, which require tighter creative control and who can reject the result. The practical test is whether those choices are explicit before work begins, rather than negotiated after an unsuitable output arrives.

Separate carrying the work from judging it

The month's handoff article made an important distinction. Downloading a finished deck, uploading it elsewhere and notifying colleagues are not the same activity as checking its claims. Both involve a person. Only some of that involvement necessarily requires judgement. Treating every manual step as a quality safeguard can hide avoidable administration.

“I bring it out of, once it's been refined and I'm happy with the state that it's in, it comes out because I have to share it with clients in their environments. So it has to go and live somewhere else.”Owner, agency and marketing, under 25 staff

This account also adds a caution. A transfer may exist because the client works somewhere else, not because the agency has overlooked an obvious connection. Removing it may require agreement about access, ownership and where the approved version lives. The call identifies the boundary but does not establish whether automating that transfer would be worthwhile.

“We have gates whereby, or milestones, should I say, in this particular workflow, whereby the information is delivered, it's then checked, there is human intervention, and then it is interpreted or it is redirected to do some further work.”Agency representative, agency and marketing, 25 to 100 staff

Here, the person does more than pass information along. They interpret it and may send it back for further work. For a useful process review, follow a deliverable from brief to client. Mark each place where someone moves it, changes it or authorises it. Then examine the carrying separately from the checking. That preserves the reason for a review while making the cost of the handoff visible.

Permission to experiment is not capacity to adopt

The calls describe considerable freedom to try AI. Staff can test ideas directly in some agencies; elsewhere, a specification or risk assessment comes first. Neither arrangement, by itself, tells us whether a useful experiment becomes normal practice. Trying an approach and maintaining it are different demands on the business.

“So it's a bit of a mixture, but equally lots of stuff stalls.”Agency representative, agency and marketing, under 25 staff

The accompanying account describes people trying ideas themselves or involving a more technical colleague. Time, capacity and priorities still hold work back. This is not evidence that experimentation fails. It is evidence that permission alone does not provide a route from an interesting result to a process colleagues can depend on.

“And we have an AI playbook, as we call it, which is a collection of our global policies and guidelines for AI use. And that is there to really be strict with our team about what AI is for and what it's not for.”Agency representative, agency and marketing, 25 to 100 staff

A playbook can define the permitted work. It does not, on its own, demonstrate adoption or compliance. The director's next check is operational: who maintains a useful workflow, who teaches colleagues to use it and who handles failures? An experiment needs an owner after the demonstration, not just an enthusiastic sponsor before it.

Measure the work that remains

The evidence for productivity is uneven. Some accounts describe tracked workflows. Others report a strong impression of improvement without a measured comparison. Those statements can identify promising work, but they should not carry the same weight in a board discussion.

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

The accompanying answer says there was no before-and-after productivity measurement or specific target. That limits the conclusion. It does not establish that there was no benefit. It means the benefit remains unverified in that account. Elsewhere, a speaker explains that returns are difficult to estimate because AI redistributes work while creating new tasks and changing expectations.

For agencies, measurement should therefore follow the deliverable beyond generation. Include preparation, checking, revision, transfer and approval. Record changes in quality and scope alongside effort. If clients receive more versions or deeper analysis, that is a different outcome from producing the same work with less input. Both may matter commercially. Neither should disappear into a general claim that the team feels faster.

Find your stage

The AI Integration Index assessment shows your stage through the way AI operates in your business: who uses it, what it can complete, how work moves, how people access it, what governs it and whether its effects are measured. That is a more useful starting point than a list of tools or a general impression of enthusiasm.

Take the assessment using the work your agency does now, including the checks and handoffs that happen after AI has produced an output. Use the result to identify what needs closer examination. The next step may be a better connection, clearer ownership or stronger evidence. It may also be an explicit decision to keep a person in control.