AI Agent Costs: Why You Pay for the Attempt and Not the Result

Think about how strange this is. A plumber who cracks your sink does not invoice you for the second visit. A lawyer who files the wrong motion does not bill you for the refiling. Software is the only trade that convinced its customers to pay for the swing rather than the hit. We agreed to it without arguing. For thirty years, software did the same thing every time you ran it. A spreadsheet does not have a success rate.

An agent does.

AI agent pricing bills the attempt, not the result

Look at what is actually published. Intercom’s Fin charges $0.99 per resolved outcome. If it does not resolve the issue, you are not billed. Zendesk prices in the same neighborhood, roughly a dollar to two dollars per automated resolution. Salesforce’s Agentforce charges $2.00 per conversation, and that charge applies whether or not the issue gets resolved.

Same category. Same buyer. Opposite arrangement.

Run the numbers on ten thousand conversations a month. Fin reports an average resolution rate around 76%. At $0.99 per outcome, you pay for the 7,600 that landed, which comes to $7,524. At $2.00 per conversation, you pay for all ten thousand, which comes to $20,000.

That is 2.66 times the cost for the same volume of work.

Sit with the second half. Of the ten thousand conversations, 2,400 went unresolved. Under per-conversation pricing you paid $2.00 for each of them anyway. That is $4,800 a month, roughly $57,600 a year, for work the machine did not do. You bought an attempt. The attempt failed. The invoice did not notice.

None of this makes Salesforce a villain. Agentforce is a serious product and the company has been visibly wrestling with how to price it. It launched per conversation, added a credit model at roughly ten cents per action, then added per-user licensing at $125 a seat. Three models running at once, changed twice inside eight months. In June it agreed to buy Fin for about $3.6 billion, which is one way of saying the outcome-priced model was worth billions to own.

That is not a company failing. That is a company discovering that the meter is hard to defend.

The tokens make it worse, which is the part that stings

Here is where the Retry Tax stops being an annoyance and starts being a design problem.

Anthropic published a detailed writeup of its multi-agent research system. Buried in it is the most useful number in the industry. Its agents used roughly four times the tokens of a chat. Its multi-agent setup used about fifteen times. And token usage alone explained 80% of the variance in performance on one of its evaluations.

Read that plainly. These systems work in large part because they spend. In the same post, Anthropic said something else. Work that needs shared context, or heavy coordination between agents, is a poor fit today. They disclosed the cost, they disclosed the benefit, and they named the boundary. That is more honesty than most vendors offer, and it deserves saying out loud.

But if spending is the lever, then the buyer is holding the wrong end of it.

Now add the finding almost nobody has absorbed. The research team at Chroma tested eighteen frontier models. That included the ones selling million-token context windows. They measured how accuracy holds up as input grows. All eighteen degraded. Not some. Every one. Effective context landed somewhere around half to two thirds of what was printed on the box. Strangest of all, the models did better on shuffled text than on well-organized documents.

Put the two findings side by side and the sequence appears.

You pay per token. More tokens raise the odds of a good answer, up to a point. Past that point, more tokens make the answer worse. The agent produces something wrong. You retry. The retry costs tokens too. The failed attempt now sits in the context. So the second run starts from a worse place than the first.

You paid three times. Once for the failure. Once for the degradation. Once for the fix.

The bill nobody forecast

The most quoted defense of all this is that inference is getting cheaper, so the problem solves itself. Prices did fall. One analysis of enterprise API traffic put the blended cost per million tokens down about 67% year over year. And yet 73% of enterprises still exceeded their AI budgets.

Total spend is price times volume. Prices fell. Volume did not. Agentic architectures multiply volume on purpose, because volume is the thing that makes them work.

Uber burned its entire 2026 generative AI allocation in four months. Five thousand engineers, adoption climbing from 32% to 84% in a single month, power users clearing two thousand dollars each. Then a $1,500 cap. Tesla, which once ran leaderboards ranking engineers by token consumption, now caps spend at $200 a week.

These are not careless companies. They learned that agentic spend cannot be forecast, because the thing that drives it is failure, and nobody forecasts failure.

The villain is the arrangement, not the vendor

It would be lazy to say the model providers designed this to enrich themselves. There is no evidence for it, and the accusation is beneath the argument. Anthropic published the token multiplier itself. Nobody hid anything.

The problem is structural. When a vendor is paid for consumption, a wrong answer costs the vendor nothing. When a vendor is paid for outcomes, a wrong answer costs the vendor everything.

Those two structures produce different engineering. Not different morals. Under the meter, the Retry Tax is your problem. Under outcome pricing, it becomes the vendor’s problem. Vendors solve their own problems fast.

So ask the question that actually matters

Everyone evaluating an AI agent asks which model it runs on. It is the wrong question, and it is the one every vendor is delighted to answer.

Here is the right one. When this system is wrong, who pays?

If the answer is you, you are not buying an assistant. You are underwriting one. You are financing the machine’s education, one failed run at a time, and calling it a software subscription.

If the answer is the vendor, something useful happens. The vendor starts caring about context rot. The vendor starts caring whether an agent stops and asks before it does something irreversible. The vendor starts spending fewer tokens and getting it right the first time, because that is the only way the vendor gets paid.

Same technology. Same models. Opposite incentives. That difference will decide which agentic AI projects survive 2027, and which become the 40% Gartner expects to be canceled.

We think the second arrangement is right, and we are building on that premise. We have not published a price, and we will not until we can defend every number in it. That is the same standard that kept our graymail research unpublished until the arithmetic held.

You do not have to wait for us. Pull your last AI invoice. Find the line items you paid for. Then find the ones that worked.

Ask us who pays when Pax is wrong

It is a fair question and we would rather answer it in a conversation than in a brochure. Bring your invoice, your failure rate, and your hardest procurement objection.