Where AI quietly changed our work

A little over a year ago we decided as a team to start putting AI into more of how we work.

Not as a product feature. Not as a vision statement to put on a deck. Just as a tool we'd lean on instead of doing certain things by hand. The decision was less dramatic than the writing about AI at the time made it sound like it ought to be. We tried a small set of experiments. Some worked. Some didn't. We kept the ones that worked and stopped the ones that didn't, the way you would with any other operational change.

A year in, the shape of our workdays has noticeably changed.

Some of it went the way I expected. Most of it didn't. Here is what I actually saw, told without the marketing layer.

The quiet wins

The parts that worked were much quieter than I expected they would be.

We used to spend roughly a day per week, collectively, summarizing meeting notes. Stand-ups, customer calls, post-mortems, design reviews, the lot. We now spend something closer to an hour. Nobody made a big deal of this. It happened gradually. Someone tried it on a Friday. It worked. Other people copied. By month three, summarizing-by-hand had quietly stopped being a thing anyone did. There was no announcement.

The same thing happened with first drafts. Specs, internal memos, the technical part of a customer email, the first pass at a release note. Whoever was assigned the draft would generate something, edit aggressively, send it. Total time spent went down. Quality went up slightly in the cases where the bottleneck had been "getting started" rather than "knowing what to say". It went down slightly in the cases where the bottleneck had been the thinking itself. We learned which was which by getting it wrong a few times and noticing.

PR reviews got faster at the edges. AI is good at catching small mistakes and less good at catching the subtle ones, so we tuned the workflow accordingly. AI does a first pass. A human reviewer does the second and is responsible for the final call. The number of "did anyone actually read this?" PRs dropped to roughly zero.

None of these wins are dramatic on their own. Together, the rhythm of a week is different.

The social part

The parts I didn't expect were mostly social.

The two heaviest AI users on our team today are people who, twelve months ago, were the most skeptical. One of them had a strong privacy concern. The other thought it was overhyped. Both of them came around when they hit a specific problem where the tool just worked, and now they're the ones telling other people which workflows to migrate.

Conversely, two people who were extremely enthusiastic in week one quietly stopped using anything by month four. I asked one of them about it later, over coffee. He said the outputs felt close-but-wrong often enough that he stopped trusting them, and the trust never came back. That made sense to me. I'd seen the same pattern in myself in one specific area. I'd tried using AI to write a particular kind of strategy memo for our board. The result was eighty percent there but always twenty percent wrong in subtle ways, and the cost of fixing the wrong twenty percent was higher than the cost of just writing the memo from scratch. After a few cycles of that, I went back to writing those memos by hand. I haven't tried again.

The thing I keep coming back to is that people decided, individually, where AI helped them think better and where it didn't. There was no top-down policy that produced this. The actual map of where AI lives on our team now was drawn by everyone making small local decisions about what they personally trusted.

I find that interesting. We had spent a fair amount of time worrying about how to roll out AI tooling fairly across the team. In practice the team rolled it out for themselves, by trying things, talking about it at lunch, and quietly converging.

The new failure mode

The hardest thing this year wasn't getting AI to work in any specific place. It was the new failure mode it introduced everywhere else.

The tool will give you a confident draft of whatever you ask it for. If you can't tell whether that draft is right, you've quietly shifted the work from doing to judging.

For a while we didn't notice this. Our weekly output went up. People felt productive. Then we hit a couple of small issues that had been hiding inside confident-looking AI-generated drafts that nobody had quite looked at hard enough. The fixes weren't expensive. The pattern was unsettling. We'd built more reliance on judgment than we'd realized, and judgment is harder than doing because judgment requires you to already know what good looks like.

We started talking about this internally as the "judgment tax".

When you replace doing with AI-assisted doing, you don't actually remove the work. You move it. The work moves from typing and clicking to deciding whether what's on the screen is right. For a senior engineer reviewing a generated function in a domain she has spent five years in, that move is a win. The judging part is fast and confident. For someone newer to a domain, that move can be a loss. The bad output looks more authoritative than their handwritten version would have, so they ship it, and the bug shows up in production a week later.

If your team is generally good at the kind of judging the work requires, AI compounds beautifully. If you have anyone on the team who isn't yet, the cost of their AI-assisted work can actually be higher than the cost of their hand work used to be, because mediocre output now arrives faster and with more confidence.

That has changed how I think about onboarding. The new hires we bring in over the next year are going to need to be taught what good looks like before they reach for the tool that's happy to produce a confident draft of anything. I don't yet know how to do that part well.

The thing I'm still sitting with

There is one piece of this year that I don't have a position on yet.

The things we used to do because they were necessary, like reading documentation slowly, writing from a blank page, sitting with a problem long enough that we knew it from the inside, are now things we do because we want to.

These are still very valuable. I am almost certain of that. They are also now optional in a way they were not before, and the temptation to skip them is much bigger than I expected. Some of our newest hires reach for the AI before they reach for the documentation. They get the answer. They get it faster than I would have at their level. They don't internalize the area in the same way I had to. I don't yet know what that compounds to in five years.

The optimistic version is that this is a generational shift, and that the new generation is going to operate at a higher level of abstraction the way every previous generation has, without having to know the layers underneath. That could be fine. It has happened before.

The pessimistic version is that the layers underneath matter more than we think, and that a generation of engineers who never really learn them will produce a generation of products that look fine and quietly aren't. That has also happened before.

I genuinely don't know which version is closer to right. I'll probably write about it again when I have a better sense of it. The honest answer right now is that we're early.

A year in, the headline I'd give myself is this: the wins were quieter and more useful than I expected, the failure modes were stranger and more social than I expected, and the part I'm most curious about is the longest-term one, which I haven't lived through yet.

More when I've sat with it a while longer.