Last time I told you about the first wall; the assistant that fell four minutes behind a fake ServiceNow workshop, and the lesson it taught me.
My initial approach was wrong. I was stuffing the entire workshop down the AI's throat every pass, everything captured so far, and expecting it to think. To its credit it didn't choke; it just took longer and longer to chew every mouthful. Once I saw that, tinkering stopped being enough. I needed a plan. It turned into three.
Track one: prove the method end to end. The pretend client gets the full treatment: the workshops, the data modelling, the workbooks, the reviews, the documentation; every deliverable an engagement would need, produced start to finish with synthesised data. If it can't survive my imaginary friends and their awkward data, it has no business anywhere else. Nothing here goes near a client; that's the point of the rig.
Track two: turn my own method into a product. My templates, my lexicons, my checks, the scar tissue from a decade of implementations; the stuff in my head that can't be downloaded; turned into capabilities an AI can execute with my standards attached. AI is the mechanism. The knowledge is the product.
Track three: build the factory. If AI is doing the building, I become the bottleneck; how fast I can safely feed it work, and how carefully I can check what comes back. So the third track is the production line around exactly that: I queue the work in the evening, it runs overnight, and nothing is finished until it's been through my hands the next morning. Build the machine that builds the machine (I did warn you about the geekiness).
And the part I promised myself I'd say out loud: there's a good chance this live-in-the-session stuff might not work at all. Current models may simply not be fast enough for a live workshop, and no amount of enthusiasm changes that. I planned for it two ways. Even if live doesn't work, that's OK; the offline path (the workbooks, the reviews, the documentation) already makes delivery faster and better, so the work pays for itself either way. And everything is built provider-agnostic; if today's models aren't fast enough, tomorrow's will be, and swapping one for the other is a plumbing job.
The signals so far are promising; that forty-five second heartbeat from last time is the first hard evidence. But promising might not translate into working, and I'd rather show you the misses as they happen than claim victory from the sofa.
Six experiments, each one raising the stakes, each one documented honestly. That's the plan. Next time; what the file timestamps say actually happened when the plan collided with reality.
If you were betting: is live AI in workshops possible with today's models, or is this future model sorcery? Call it in the comments.
First published on LinkedIn, 21 July 2026. The conversation lives there; the writing lives here.