Back to school

The last post ended with a gap: every person at Findev has an AI assistant, the company itself has none, and we did not know what closes that gap. Stumbling into the answer by trial and error felt expensive. So we did two things instead:

• We went looking for people who had already built something like what we had in mind, to bring them onto the team.
• We went back to school, ourselves and a group of colleagues with us, to hear from practitioners where the real frontier of this technology runs today.

About the school, an honest word first. I expected to nod politely. Instead, we sat there feeling behind. What practitioners actually do inside companies turned out to be much further ahead of the public conversation than we assumed. And none of it was slideware: the things people showed us run every day and carry real work. I remember the feeling of those days very precisely: a train was leaving. The world was changing that fast right at that moment, not someday, and we were not keeping up with it.

The most useful thing school did was name the main wall for us, before we had to discover it the hard way. The models are brilliant, but they do not know enough about your company. They do not know who your clients are, how your invoices are structured, what was agreed on last Tuesday's call, or why one of our entities runs payroll differently from the others. We had assumed the hard part of using AI would be the AI. It is not. The hard part is context: collecting it, structuring it, and feeding the right slice of it to the model at the right moment. Which is also the honest answer to why the personal assistant works and the company assistant does not exist: the personal one borrows your head for free.

There is a trap on the other side of that wall too. Companies of our size, and every size above ours, accumulate a lot of context, and a model is as easy to overfeed as to starve. The practitioners consider this a real problem; we suspect we will not run into it right away. More on that when we do.

Another pattern we saw more than once: AI installed where a deterministic script belongs. People sign themselves up for a permanent token bill on work that, by design, should cost nothing to run. We wrote that observation down.

School also pointed us to tools we had somehow walked past. Granola now sits in our meetings and writes them down, so a call no longer evaporates the moment it ends. Wispr Flow lets us talk instead of type: a good part of this series is dictated on walks. Neither is revolutionary on its own. What struck us was the sum: while we were watching model releases, a whole layer of plumbing had quietly grown up around the models, and most of it does one job, catching context before it disappears.

One thing the stories did not answer, though. Practitioner after practitioner described how they rebuilt their company around AI, and almost none of them said how much they had risked breaking along the way. For us that is the question. Findev works, and we have no intention of breaking what works. It may well be that for some of the storytellers the risk was perfectly rational: if the business was not working that well to begin with, bold surgery costs little. A company that is profitable, and trusted by its clients, does not get that discount. So one more wall of our experiment quietly set itself: whatever we change, the machine keeps running while we change it.

The main conclusion was a sobering one: it is too early for us to build agents.