The other side of five o'clock

Everything on this blog happens after five o'clock. That is the arrangement I have with myself. The apps, the graveyard full of dead ones, the posts about failing fast and cheap, all of it is the version of me who gets to build his own things at night. In that life the AI is pure upside. It dropped the price of being wrong from a quarter of my year to a weekend and a few dollars of tokens, and that alone turned building back into something I love.

Then Monday comes, I go to the day job, and the same tool means something else entirely.

I am a project manager at an outsourcing company. I don't own it, I don't set the prices, and I don't decide the strategy. My job sits in the middle: between the client who wants the thing built, the developers who build it, and the numbers that have to close. From that seat you see something the night version of me never has to look at. The same AI that makes my hobby cheap is making my whole industry cheaper, and nobody in the room has agreed on who gets to keep the savings.

I felt it most plainly a few weeks ago, putting together an estimate. Near the end I did the thing everyone now does: I took the honest, all-human number and cut it, because we would be building with AI and it would genuinely be faster. The client half expected the cut before I sent it. That is the part that stuck with me. The discount is no longer a surprise I offer, it is a rebate the buyer already assumes. And once a discount is assumed, the interesting question is not how big it is. It is whose money it was in the first place.

Here is the trap underneath all of it. Outsourcing, the way most shops still price it, sells time. Man-days, headcount, a rate per hour. When your unit is time and a tool makes the same work take less of it, you have quietly promised to charge less for the same result. The client sees the tool, does the math, and asks for the hours back. So the productivity AI hands you does not land in your pocket. It flows straight through you to the buyer, because the thing on the invoice was never the outcome, it was the clock.

You would think the fix is obvious: stop selling time, sell the result. That is the right answer on a slide. In a real shop, with real contracts and a client who has bought man-days from you for five years, changing the unit of the sale is slow, and it happens right while the price is already sliding under you. You are repricing the plane mid-flight, with the same passengers on board.

The other half of the story is that the savings were never as large as the slide promised. AI is cheap on paper and expensive in the room. In a demo it writes a feature while you watch. Getting a floor of developers to actually use it, every day, on real client work, is a different project. The client's security team will not let the code leave their network, so half the good tooling is off the table before you start. The codebase is ten years old and does not fit in any context window. And the tax nobody puts in the estimate: someone senior has to read what the machine wrote, carefully, because the client is paying us precisely so they do not have to trust it blind. A lot of the time saved in the writing comes back in the reading. The developers feel this from both ends. Some are quietly afraid the tool is here to replace them, and the same person, on the same day, does not trust its output enough to ship it without checking every line. Both feelings are correct. That is what makes it hard.

Which points at the thing clients are actually paying an outsourcing company for, and always were. It was never the code. Code is the artifact. What they buy is someone to hand the risk to. Someone who answers the phone at two in the morning when production is down, who carries the blame when the release is wrong, who is still on the hook a year later. AI got very good at producing the artifact and did nothing for the part where a named human absorbs the consequences. So pricing the work off typing speed is pricing the wrong thing. We discount the part the tool sped up and give away for free the part it cannot touch.

And the ground keeps moving. The clients who used to need us for the whole build can now hold the tool themselves. A capable person on their side can sit with the same AI and get the easy eighty percent, which means the part they still come to us for is narrower and harder: the domain we know that they do not, the integration into systems that already exist, the operating of the thing after launch, and again, the accountability. The moat is still there. It is just thinner, and made of different stuff than it used to be.

The part that worries me most is the one furthest out. Our juniors used to grow up on the easy work. The small tickets, the boilerplate, the boring CRUD that taught them the shape of a real system by making them build it badly first. That is exactly the work the AI now does in a second. Good for this quarter's margin, quietly bad for everything after it, because I do not know where the next senior comes from if the bottom rungs of the ladder have been automated away. I do not have an answer to that one. I am not sure anyone does. I am watching us trade a problem we can measure now for one we will only feel in a few years.

I want to be honest about my own place in this, because it shapes what I can and cannot say. I am not the founder. I do not get to redraw how we sell or decide that we stop charging by the day. I am the person in the middle who feels the squeeze from both sides and turns it into schedules, scope, and hard conversations. From here the view is not a strategy, it is a weather report. The pressure is real, it is not going away, and most of the people around me are reacting to it one estimate at a time rather than deciding anything on purpose.

So what do I actually do with it. Small things. I have stopped writing estimates as if the AI discount is free money the client is owed, and started naming what did not get cheaper, the review, the responsibility, the knowing, so at least the number tells the truth. I push, where I can, to describe the work by what it delivers instead of how many bodies for how many days. I say out loud that a faster first draft is not a finished product, because someone has to. None of it is a plan. It is just refusing to pretend the old math still works.

Then five o'clock comes, I go home, and I open the little project that is mine, and the exact same tool that spent all day making my job harder to price makes being wrong cheap and building fun again. I have not reconciled those two. I am not sure they reconcile. The thing that is repricing the work I do for a living is the reason I get to make anything at all on my own. I hold both, and I go back in on Monday.