We spent a year pushing generative tooling into every corner of the studio to find out where it did not belong. Four things came back to humans and have stayed there.
One: the first client conversation
We tried pre-briefing: a summary of the client's site, market and likely problems, generated before the call. It was accurate and it made the calls worse. The team arrived with a hypothesis and spent the hour confirming it instead of listening.
The first conversation is for finding out that the stated problem is not the problem. That requires arriving empty.
Two: estimating
Generated estimates were consistently plausible and consistently wrong in the same direction — optimistic on integration, blind to the client's internal approval cycles. The information that makes an estimate accurate is mostly political and lives in nobody's training data.
The model estimated the work. It could not estimate the client, and that is where the weeks go.
Three: the cut list
We automated the "what should we descope" analysis and got defensible recommendations that nobody defended. The cut list only works as a shared artefact if the people who have to live with it made it. Handed down, it reads as a constraint; argued out, it reads as a plan.
Four: writing the case study
Generated case studies are fluent, structurally correct and completely interchangeable. The detail that makes one worth reading — the operations lead who wanted to stop being a gatekeeper three years ago — is the detail nobody wrote down, which means it has to come out of someone's memory in a conversation.
What stayed automated
Plenty, and it is the boring half: drafting variations, transcript summarisation, asset resizing, changelog generation, first-pass accessibility audits, test data. Volume work where being wrong is cheap and a human checks the output anyway.
The pattern, in hindsight, is simple. Automate the tasks where the output is checkable. Keep the tasks where the value is in the checking.
