How Do AI Agents Make Money? For Vendors and For You
The question has two answers, and mixing them up is how people buy badly. One is how the companies selling agents make money. The other is how an agent makes money for you. They are not the same, and the first one is designed to be more legible than the second.
Start with the vendor side, because it tells you what the vendor is optimising for.
The four ways agent vendors charge
Model | How you pay | What it makes them optimise for | Watch for |
|---|---|---|---|
Per seat | Monthly, per user | Adding users | A role owned by 40 seats is not one role |
Per role | Monthly, per role running | The role actually working | The good model, and the rare one |
Usage | Per token, call, or action | Volume of actions | Cost grows with the problem it is solving |
Outcome | A share of what is recovered | The result | Rare, and it needs a clean baseline to work |
Programme | Fixed fee or time and materials | Scope and hours | Reported first year ranges of $500,000 to $3,000,000 |
Per seat pricing is the one to be careful with, and it is the most common. It pays the vendor more when more people touch the tool. If your goal is fewer hours spent on a workflow, you and your vendor are pulling in opposite directions.
Per role pricing aligns better. The vendor only gets paid while the role is working, so the incentive is a role that holds up rather than one that demos well. That is the model we use, and it is worth asking every vendor which one they use and why.
How an agent makes money for you
Here is where the framing matters more than the maths.
Most vendors sell the savings case, which means doing the same work for less money. That is a real thing and it is a weak case, because it asks your team to shrink and it caps your upside at whatever you were already spending.
Recovery is the better frame. The work that an agent reads line by line is usually work nobody reads line by line today, because there is no time. Retail deductions taken by customers, short pays, invoice terms that were not honoured, chargebacks that were never contested. That money left the building already. Reading every line finds it.
The difference in scale is not close. A savings case is bounded by your own payroll. A recovery case is bounded by what is leaving, which is usually larger, and it does not require anyone to lose their job for it to be true.
How to tell whether it is making money
Four things to measure, and three of them are not the agent's doing.
Baseline first. What did this workflow recover before the agent existed, over one full month? No baseline means no proof, and most vendors skip it because it is inconvenient.
Coverage. Share of arriving work actually read. 100 percent or you are still sampling.
Recovered value. Money found that would otherwise have gone. This is the headline number.
Hours returned. Time your team no longer spends on the repetitive part. This is the second number, and it is the one that gets a champion internally.
If a vendor cannot tell you what the baseline was, the claim at the end is a story rather than a result.
The uncomfortable part
An agent that is measured on recovered money can fail visibly. An agent measured on actions taken cannot, because action counts go up whether or not anything was recovered. That is why the reports look good and the bank account does not move.
Insist on a number that could come back bad.
How much do AI agents cost? covers the other side of the ledger.
Where this goes next
What does an AI agent do exactly?
Should you hire another person or deploy an AI agent?
AI agents for accounts payable teams
What is workforce design for the AI era?
Frequently asked questions
How do AI agent companies make money? Four main ways: per seat, per role running, usage based, or a share of the outcome. Larger firms also sell agent work as consulting programmes.
Is per seat or per role better for the buyer? Per role, if the goal is fewer hours spent on a workflow. Per seat pricing pays the vendor more as more people use the tool, which works against that goal.
How do AI agents make money for a business? By recovering money that was already leaving and by returning hours. Recovery is usually the larger and more defensible number, because it does not depend on anyone's salary.
What is outcome based pricing? The vendor takes a share of what the agent recovers. It aligns incentives well and it requires a measured baseline, which is why it is uncommon.
Can an AI agent save money? It can, and savings cases are typically smaller than recovery cases and harder to prove, because they depend on a headcount decision that somebody has to own.
How do I know if the agent is worth it? Measure one full month before deployment, then compare coverage, recovered value, and hours returned. If there is no baseline, the comparison cannot be made.
