The AI was the easy part: A field report on automating an insights function – and where the real work moved

The hard part of automating an insights function isn’t the AI. I found that out by building it.

Over the past year I rebuilt my in-house insights function around AI – first as a pilot, now firmly as the operating model. The machine now does much of the production that used to define the job: synthesis, segmentation, moderation, deck generation. I’ll walk through a project end to end – a concept-testing platform built and run by a team of one, with the rigor of a top-agency engagement and the form of a living tool, not a static report.

Automating all of that turned out to be the easy part. The surprise came after. Once you can automate almost anything, the human work doesn’t get smaller; it becomes the entire job. The work splits in two. Half is programming the machine – building the skills and agents that do the production. The other half is the part the machine can’t touch: deciding what’s true, what matters and what ships. That second half is where the job actually gets difficult, but where the function proves its value.

This is a field report for operators who don’t buy either easy answer – “automate everything” or “AI is just a tool.” Wherever AI lands, automation will be table stakes. The edge will be the half it can’t touch.

And I’m handing you the starter pack to do it yourself. You’ll walk out able to set up the same automation at your own company: the framework and the skills I’m running in production right now.

Key takeaways:

- A decision framework for what to hand to the machine and what to keep in human hands.
- A starter pack of the skills I run in production, with the sequence to stand them up.
- The operating model for the harder half: how to run the function once the machine owns production.