The Big AI Secret · ongoing research

AI finishes the deck. Someone still carries it by hand.

Agency accounts reveal a gap between generating finished work and moving it through the business. They also show why manual transfer and human review should not be treated as the same problem.

7 October 2026 · 4 minute read

The striking detail is not that AI needs help writing the deck. In this account, it produces the whole thing. The remaining work is downloading and uploading it, sharing a project link and updating the job so client services know it is ready. The content is complete, but the process still needs someone to carry it.

“I mean, it does the hard part brilliantly. Then somebody carries it across the road by hand.”Agency and marketing, 25-100 staff

That distinction is the focus of this week's notebook. The agency call evidence separates what AI can produce from what happens next. A finished output can still sit outside the process that gets it approved, delivered and used. Calling the task automated would miss that remaining work. Calling it wholly manual would miss what has changed.

Finished output is not finished work

“At the moment, we don't have, it's not things actually don't, they just live inside the systems and we go there to look at it and refresh it and take bits out.”Agency and marketing, under-25 staff

Here, the person is still responsible for visiting the system, refreshing the output and extracting what is needed. That is a different arrangement from work arriving where the next colleague can act on it. It may be perfectly workable. But any assessment of productivity needs to include those actions, not stop when AI has generated a result.

Another account makes the remaining human work unusually clear. Generating creative options is only part of the task. Someone must set the brief, decide what is suitable and handle the output. Those activities sit together in the description, although they serve different purposes.

“In that particular workflow, the human elements are defining what you want it to do, what the subject of the image is going to be, and then validating whether you like some of the outputs and then handling the download.”Agency and marketing, 25-100 staff

Defining the brief and judging the result require a decision. Handling a download moves an asset. The evidence does not tell us how burdensome that transfer is, or whether changing it would be worthwhile. It does show why a broad description such as human involvement is too blunt for examining how the work actually happens.

Keep the judgement. Examine the carrying.

“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.”Agency and marketing, 100+ staff

This is not evidence of an unfinished automation project. It is a stated control over client work. A person checks and edits the output before release. Removing that step would change the agency's quality process, not merely make information move more smoothly.

The useful distinction is between a person deciding that work is ready and a person transporting work already judged ready. A review can catch errors or reject unsuitable material. A transfer gets material to its destination. Both may be necessary in the current setup, but they should not be evaluated as though they provide the same value.

“Both, depending which way it's going out. Social carries on by itself once it's approved, paid in CRM, somebody moves it.”Agency and marketing, 25-100 staff

This account provides a useful counterpoint. Approval and automatic movement can coexist. Approved social output carries on, while other output is moved by a person. The boundary differs by destination, even within the work being described. A business-wide label cannot capture that detail.

What the evidence can support

That figure describes the measured businesses across sectors. It is not an agency percentage, and it does not establish how common manual handoffs are. The selected agency call extracts supply examples of how work moves, not a representative count of these practices. They do not establish the cost of carrying work between systems or the return from changing it.

For a director examining a process, the practical next step is to follow an output beyond generation. Record where it lands, who checks it, who moves it and how the next person knows it is ready. Keep decision-making separate from transport. Then assess the whole route, including checking, rework and delivery, rather than the speed of generation alone.

The finding is narrower than a claim that agencies need more automation. These accounts show that capable AI can sit inside a process still held together by people. Some of that human work protects quality. Some carries files. Knowing which is which is a better starting point than assuming every human touch is either a safeguard or a delay.