Things I believe
Mostly about feedback loops
Three things I've had to argue for more than once. They're really all the same argument: most of what goes wrong on a team is a feedback loop that's too slow, isn't closed, or has accidentally started measuring the wrong thing. I've tried to bring the receipts, since I think an argument like this is a lot less interesting without them.
Ship the valuable part first, then decide if you need the rest
Our customers wanted a way to get messages signed off before they went out, and the project came to us forecast as a big one. We built the obvious part first, and it turned out to cover 90% of customers for roughly 10% of the forecast effort. I don't think that was clever analysis so much as ordinary sequencing: go after the easy value early, get it in front of people, and then make the rest argue for itself against everything else you could be doing. Mostly, it didn't win that argument.
The bottleneck usually isn't the model, it's you being in the loop
A lot of people are finding these tools disappointing, and I think that's usually a workflow problem rather than a model problem: drop AI into a process designed around the old costs and you tend to move the work rather than remove it. Step back and look at the whole thing, and the problem often dissolves. It also changes what's worth attempting at all. Reverse-engineering iCloud's sync protocol was never worth months of my time, but with AI in the loop it was worth a few weekends, and at work I led four AI features into production on the same thinking, the first cutting journey-building time by 83%. The biggest accelerator I've found is giving the model some way to check its own answer, because Goldratt's The Goal applies here as much as it does on a factory floor: time saved anywhere but the constraint isn't really saved, and if the model has to come back to me to find out whether it got it right, the constraint is me.
A number that moves when you start looking isn't measuring anything
We had an overnight check that counted errors in our data, and every so often we'd think of a new kind of error worth looking for. The count would jump, and it always looked alarming. But those errors had been there the whole time; we simply hadn't been looking for them. The metric was meant to tell us when things were getting worse, and mostly it told us when we'd gotten more observant, the same way screening more people for a disease reliably finds more of it without anybody having gotten sicker. I argued we should only count the error types that existed when we set the metric, and I lost that one, which I think about more often than I'd like.