What it actually costs

The most honest thing I can say about AI in production is that it is a very good deal whose price you pay later, somewhere else, and never on the invoice.

If you have read the last few editions, you have probably noticed I keep circling the same shape from different sides. This week I want to name it directly, because it is the most honest thing I can tell you about running AI in real systems.

AI in production is a genuinely good deal. The leverage is real and I am not going to pretend otherwise. But it is the kind of deal where the savings show up immediately and the costs show up later, somewhere else, spread thin enough that they never land on anyone's invoice. That asymmetry is the whole problem.

Think about why teams keep getting surprised. The savings from AI are immediate, visible, and concentrated. You feel them today, you can see them in the demo, and they sit right where you used the tool. The costs are the opposite. Delayed, invisible, and diffuse. They arrive next quarter, they show up in a different part of the system than the part where you saved, and they are spread across enough people and enough time that nobody can point to the moment they began.

So every individual decision to reach for AI looks like a clean win, because in the moment you make it, only the savings are visible. The bill is real. It just does not arrive addressed to you, today. Here is what is on it, in my experience, that almost nobody counts.

The first cost is the review tax. AI moved the work from writing to verifying, and verifying generated code you do not fully understand is slow, vigilant, attention-heavy work that does not look like work. You did not save the time. You converted it into a kind of effort your tools do not measure and your people find quietly draining. I spent a whole edition on this, so I will leave it there.

The second is understanding debt. Every time someone ships code they cannot fully explain, the team takes a small loan against the future. It costs nothing the day you take it. It comes due all at once, at the worst possible moment, in the incident or the audit where someone has to actually reason about a system nobody truly understands. Most teams are carrying far more of this debt than they realize, because debt is invisible until it is called.

The third is the training ground you are giving away. The work AI now does was how juniors became seniors. Hand all of it to the machine and you save money now and quietly stop producing the experienced people you will need to supervise the machine in five years. That cost is the most delayed and the most diffuse of all, which is exactly why it is the easiest to ignore and the most expensive to discover late.

And then there is the one I think is the most dangerous, because it does not touch your code at all. It touches your judgment.

A founder told me a while back that he had started running every major decision through AI before acting on it. Pricing, hiring, product direction, partnerships. All of it went through the models first. And every time, the answer came back reasonable, balanced, confident. So he moved fast. Three big calls in six months, all of which felt right in the moment, all of which turned out to be wrong.

When he looked back, he saw what had happened. He had stopped stress testing his own thinking. No advisor pushing back, no co-founder poking holes, just AI taking his assumptions and handing them back to him in cleaner, more confident language. The danger was never that AI gave him bad advice. The danger was that it gave him confident, well structured, completely unchallenging advice, and that feels like clarity when it is actually just comfort.

This is the cost nobody puts on the ledger, because it does not look like a cost at all. It looks like good decision making. AI is a brilliant thinking partner and a terrible devil's advocate. It has no skin in your game, it does not know what it does not know about your business, and it will rarely just tell you that you are wrong. Real thinking needs friction, and AI is the most efficient way ever built to remove it. Use it to validate your thinking instead of to pressure test it, and the quality of your decisions degrades so smoothly you will never feel it happening.

None of this is an argument against the tools. I want to be clear about that, because the honest accounting is not "AI is too expensive." It is "AI is a deal, and you should know its real price before you sign, every single time."

A cost you have not named is not a cost you have avoided. It is just a cost you will pay by surprise, usually larger than if you had seen it coming. The teams that come out of this era ahead will not be the ones who got the most for free, because there is no free. They will be the ones who counted both columns honestly, took the leverage with their eyes open, and paid the real price on purpose instead of getting handed an invoice they never expected.

So here is the exercise I would actually do this week. List the three biggest things AI has saved you this quarter. Then next to each one, write down honestly what it cost you somewhere else. The review hours. The understanding you skipped. The junior you did not hire. The decision you did not pressure test. If that second column comes out blank, it is not because the costs were zero. It is because you have not looked yet.

What is on your second column that nobody has written down?

Reply and tell me. I read every one.