Why AI Agents Get More Expensive Over Time

One of my agents hit a wall at 2am.

It needed permission for something small. Nobody was awake to give it. So it stopped. No error. No alert. No failed run sitting in a log waiting to be found.

I found out the next morning, and not because anything had broken. I found out because nothing had happened at all. A shift that should have produced six pieces of work produced none, and the only evidence was an absence.

That is the part people miss about running AI inside a business. The expensive failures are not the loud ones. A crash gets fixed by lunch. This sat there quietly costing me a night, and it would have kept sitting there if I had not gone looking.

Not a bug. A silence.

It worked on day one.

Every one of these stories starts the same way. The demo worked.

The demo always works. Someone shows you the tool doing the thing, correctly, on a clean example, and you think: right, that solves it. You buy it. You roll it out. For about three weeks it does roughly what it did in the demo.

Then six months later it is open in nobody's browser, and nobody can quite tell you what went wrong.

Here is the uncomfortable answer. Nothing went wrong. Nothing broke. There was no incident, no outage, no bad release to point at. The tool got slightly less trustworthy every week until using it cost more than not using it, and your team quietly routed around it.

Most software you buy fails all at once. This fails by decay. Decay does not show up on a status page, and nobody budgets for it.

Somebody just named this.

For a long time I did not have a word for this, and neither did the people I talked to. They described it constantly, but they described it as a feeling. The AI is fine, it is just a lot of work.

The word arrived this year, and it did not come from me. A paper published in May 2026, Governing Technical Debt in Agentic AI Systems (Hydari, Iqbal and Ramasubbu), called it the stochastic tax: the recurring cost of keeping a system that behaves probabilistically inside the bounds you need it to stay in.

I want to be straight that I did not coin it. What strikes me is how many people arrived at the same idea at once. Inside about ninety days there were five separate frameworks describing adjacent versions of this problem, from a consultancy, a university, a cloud provider, and two independent research groups. None of them cite each other. That is not one idea spreading. That is a lot of people independently hitting the same wall.

Gartner put a number nearby. They predict more than 40% of agentic AI projects will be cancelled by the end of 2027, and they name three reasons: escalating costs, unclear business value, and inadequate risk controls.

I would not claim the tax explains all three. It explains two of them cleanly, and you can check that against your own operation rather than take my word for it. Escalating costs and unclear business value are exactly what this looks like from the outside when a team is quietly paying it every week.

Where you are actually paying it.

The tax does not appear on the invoice from the vendor. That is the whole problem. If it arrived as a line item, somebody would have cancelled it already.

It appears on your payroll instead. Four places, and you will recognize all of them.

Somebody re-checks work the tool already finished. Not because they were told to. Because it was wrong once, months ago, and the trust never came back. Four hours a week, every week, that nobody logs against the AI budget.

Somebody patches the same rule a third time. Same rule. Third patch. Each one made sense on its own.

Somebody hand-verifies before anything goes out. The tool does the work and a person checks the work, which means you are now paying for the tool and for the labor it was bought to replace.

A prompt exists that nobody owns. It runs. It matters. The person who wrote it moved teams. Nobody will touch it now, because nobody knows what breaks if they do.

Add those up in hours and you get a real number. Most teams have never added them up, because individually each one just looks like somebody being careful.

Why the rate goes up.

Here is the part that makes this compound rather than simply cost.

Every workaround you add makes the next fix more expensive. Not because the fix itself is harder, but because the next person has to understand every previous workaround before they can safely add theirs.

Patch one takes ten minutes. Patch four takes an afternoon, and most of that afternoon is reading. The work is not the patch. The work is the archaeology.

That is your interest rate, and you can find it without any tooling at all. Look for the thing you have patched twice. That thing is telling you what you are paying now, and it is the thing that will cost the most to fix six months from now.

It is also why "we will clean it up later" is the most expensive sentence in this whole area. Later is strictly more expensive than now, every time, by an amount that grows on its own while nobody is watching.

There is good news in that, though. Debt you can name is debt you can pay down. The teams that get out from under this are not the ones with better models. They are the ones who decided to look.

Four questions to ask on Monday.

You do not need a maturity model to find out what you are paying. You need about twenty minutes and a willingness to hear the answer.

1. Who owns this prompt? If the answer is nobody, you are paying.

2. What happens when it fails at 2am? If the answer is that it waits, you are paying. That was my night.

3. How would you find out? If the answer is that someone would notice, you are paying. Someone noticing is not a control. It is a hope.

4. What have you patched twice? That is your interest rate. Start there.

None of this means the tools do not work. Mine work. I run a fleet of them and I would not go back.

It means they are not free once you have bought them, and the part you pay after you buy them is the part that never makes it into the demo.

- Jaron