I built an AI assistant for ServiceNow discovery workshops. It started forty-five seconds behind the room and ended four minutes behind it; and nobody in the room was even real.
Last time I told you about my imaginary friends; a fake workshop cast with deliberately flawed participants, all the mess of a real session. It earned its keep immediately, because it found the first real wall.
The assistant's job is to listen and build a picture of the workshop as it happens; the systems named, who owns what, decisions made, questions still open. The obvious way to do that live is a simple loop; every half-minute or so, send the AI everything heard so far and ask back an updated picture. Call each pass a cycle. And it worked; first full replay, ten cycles end to end, no errors, everything captured. One problem. Every cycle took longer than the last, because the picture kept growing; by the end of one fake hour-long workshop the assistant was answering questions the room had moved past. Watching it fall behind a conversation that wasn't even real was a deflating evening.
It took five rounds of rework, plus owning up to a design mistake. I'd been running everything through one big brain doing every job in sequence; capture, then suggestions, then checks. The fix was two moves. Stop sending the whole picture; remember it locally, ask only for what changed. And split the work between role-specific agents; a scribe capturing structure, and a second agent watching for gaps and quietly suggesting the next question, each with its own prompt, model and latency budget. After that, the last cycle of a session ran no slower than the first.
Worth being honest about what that does and doesn't prove. "No slower" means it stopped falling further behind the room; it doesn't mean real time. The transcript keeps up as the conversation happens. The thinking layer (coverage, suggestions, decisions, actions) works on a forty-five second heartbeat. Is that heartbeat fast enough for a live workshop? Fast enough that the right question arrives while it still matters? That's what this experiment has to answer. I owe you an honest verdict either way, and it's coming later in this series.
That wall was the first of many, and they all taught the same lesson. The model was never the limit. It extracted, summarised, structured, drafted. What broke, every time, was the machinery around it; what it was sent, what it remembered, how its output was structured, how anything got checked. The hard problems weren't intelligence. They were memory, context, structure and governance.
Once you see that, the question changes. It stops being "is the AI good enough?" and becomes "what do you have to build around it?" Answering that properly needed a plan. Three of them, as it turned out.
Next up, those plans.
If you've been running your own experiments and hit different walls, give me a shout in the comments; I'm collecting other people's bruises as well as my own.
First published on LinkedIn, 17 July 2026. The conversation lives there; the writing lives here.