
AI agents are often sold with a number attached: "paid back in three months". That number rarely comes from a calculation you can reproduce. It comes from an assumption about time saved, multiplied by an hourly rate, with no rework rate and no minimum volume. The day real volume arrives, the token bill and the supervision hours change the conclusion.
Calculating an ROI is not complicated. It is a division. What is difficult is choosing the inputs and accepting the result. Here is the method we apply, with a model you can copy straight into a spreadsheet.
Key Takeaways
- The ROI of an AI agent is calculated over a fixed period and a fixed scope: cumulative gains minus cumulative costs, divided by cumulative costs. Without a period, the figure means nothing.
- The gain is measured in hours actually recovered, valued at loaded hourly cost. In France, Insee puts that cost at 44.2 euros per hour in the market services sector in 2025, for companies with 10 employees or more.
- The most underestimated cost line is not the licence: it is human supervision and the handling of edge cases.
- A profitable agent almost always has the same profile: many repetitive tasks, a high unit time, a low cost of error.
- The break-even volume matters more than the headline ROI. It tells you from how many cases per month the agent stops costing you money.
The formula, and the three choices it forces
The formula fits on one line:
ROI = (cumulative gains - cumulative costs) / cumulative costs
It says nothing until three choices have been made explicitly.
The period. Twelve months is a good compromise for an agent in production: long enough to absorb the ramp-up, short enough for volume to stay predictable. A three-month ROI flatters the result by counting only the months when everything works.
The scope. One task, one team, one process. If you mix three use cases into the same calculation, you will never know which one is profitable and which one is losing money.
The denominator. Everything you pay, including your team's time. This is where most models go wrong, almost always in the same direction.
Why the denominator is almost always wrong
An AI agent proposal describes a build. It bills for scoping, development and acceptance testing. It rarely mentions what comes next: tokens consumed every month, hosting, supervision of doubtful answers, fixes after a model version change.
As a result, the cost kept in the calculation is the first invoice, not the cost of the year. The headline ROI can be halved simply by adding the recurring lines. Those lines are detailed in the real cost of an agent in production; here we keep only what affects the calculation.
Step 1: measure the task before deploying anything
A baseline is measured, not estimated. Over two to three weeks, record:
- the real volume of tasks per month, not the volume announced in a meeting;
- the unit processing time, timed on about ten representative cases;
- the share of cases that need rework afterwards, and why;
- the cost of an error when it happens (a wrong quote, a legally risky answer, a badly booked appointment).
Without these four numbers, any ROI is an opinion. External scoping mainly helps you avoid forgetting the edge cases, the ones that account for 20% of volume and 60% of the time; that is the job of an audit and advisory engagement.
The hourly cost to use, and the one to avoid
Use loaded hourly cost, not gross salary. In 2025, in France, Insee (2025) measures an hourly labour cost of 44.2 euros in market services, for companies with 10 employees or more. Across the non-agricultural market sector, the same publication reports 44.7 euros per hour, of which 31.0 euros are gross wages and salaries. Employer contributions therefore represent about 44% of gross pay: working from gross pay underestimates the gain by a third.
Two precautions. The Insee scope covers companies with 10 employees or more; if your company is smaller, use your actual payroll cost. And if a single person handles the task, do not value their hour at a consultant's rate: the gain is a cost avoided, not a selling price.
Step 2: value the time saved without fooling yourself
Three corrections separate a credible gain from a slide-deck gain.
The recovery rate. Out of ten hours freed up, how many are actually redeployed to something other than waiting in front of an inbox? Nobody knows that for you, but the number must be set before the calculation, not adjusted afterwards. A 50% rate is a prudent, defensible assumption; 100% assumes every saved hour turns into billable output, which rarely happens.
The rework rate. Is every output reviewed? How long does that review take? An agent that handles 100% of requests in two minutes but requires three minutes of checking per answer has not removed the work: it has moved it.
The volume. Gains grow with volume, fixed costs do not. An agent sized for 600 cases a month that only receives 120 is expensive per unit. That is why the break-even volume, calculated below, matters more than the average ROI.
Step 3: list the costs, including those on no invoice
| Line | Nature | In our example |
|---|---|---|
| Implementation | one-off | 12,000 euros of services |
| Internal scoping time | one-off | 68 hours |
| Tokens | variable, per task | about 0.02 dollars per request |
| Hosting and infrastructure | fixed monthly | 40 euros per month |
| Response supervision | monthly, in hours | 4 hours per month |
| Fixes and enhancements | annual | to be taken from the proposal |
| User training | one-off | one day |
The last three lines are the ones people forget. Supervision is not a detail: it is a human cost that returns every month, and it only falls with the quality of the configuration. Fixes depend on a contract: the clauses to demand before signing are detailed in the maintenance contract.
Step 4: calculate a cost per task, not a monthly cost
A monthly cost cannot be compared from one client to another. A cost per task can, and it follows from the provider's public price list.
Example with a mid-range model on OpenAI's price list (read on 19 November 2026): 1 dollar per million input tokens, 5 dollars per million output tokens. One request handled by the agent consumes about 12,000 input tokens (the request, the customer context, the instructions) and 1,500 output tokens.
- input: 12,000 tokens at 1 dollar per million, i.e. 0.012 dollars;
- output: 1,500 tokens at 5 dollars per million, i.e. 0.0075 dollars;
- total: 0.0195 dollars per request, rounded to 2 cents.
At 600 requests a month, that is 11.70 dollars in tokens. Compare that with 4 hours of supervision at 44.2 euros: 176.80 euros. Supervision costs fifteen times more than tokens. That is the most useful finding in this calculation, and it contradicts the most widespread intuition.
One methodological point: keep tokens in dollars and human costs in euros, then convert at the rate of the day when you produce the figure. Fixing a conversion rate from memory in a twelve-month model creates an error that grows over time.
Step 5: calculate the payback period and the volume threshold
Take the case of a small company handling 600 quote requests a month, at 12 minutes each.
| Line | Calculation | Result |
|---|---|---|
| Current cost | 120 hours x 44.2 euros | 5,304 euros per month |
| Volume handled by the agent | 480 requests at 3 minutes of validation | 24 hours |
| Volume handled manually | 120 requests at 12 minutes | 24 hours |
| Cost after the agent | 48 hours x 44.2 euros | 2,121.60 euros per month |
| Gross gain | 5,304 - 2,121.60 | 3,182.40 euros per month |
| Retained gain at 50% | 3,182.40 / 2 | 1,591.20 euros per month |
| Recurring costs | 40 + 176.80 + tokens | 216.80 euros and 12 dollars |
| Net gain | 1,591.20 - 216.80 | about 1,362 euros per month |
| Implementation | 12,000 + 68 hours at 44.2 euros | about 15,000 euros |
| Payback | 15,000 / 1,362 | about 11 months |
The volume threshold follows from the same table. Each request handled by the agent saves 9 minutes, i.e. 6.63 euros gross at 44.2 euros per hour, so 3.31 euros after applying the 50% recovery rate. The model's recurring fixed costs come to 216.80 euros (hosting and supervision). Threshold: 216.80 / 3.31, i.e. about 65 requests a month. Below that, the agent costs money even when it works perfectly.
That threshold is the strongest argument you can bring to a management committee: it turns a promise into a verifiable condition. Before committing the budget, it is better to validate these assumptions on a reduced scope; that is the purpose of a scoped pilot before production.
What this model does not tell you
It says nothing about quality. An agent that handles 480 requests with 8% of answers needing substantial rework does not save what the table claims. The rework rate must be measured during the pilot, then fed back into the calculation.
It says nothing about risk. A wrong answer sent to a customer costs more than the time it saved. On tasks with high legal or financial stakes, human validation stays in the cost, and rightly so.
It says nothing about drift. Volume grows, prompts get longer, the provider changes prices, a model version is deprecated. A ROI model must be revised every six months, otherwise it becomes a memory.
Finally, it does not replace the decision. An executive who did not measure a baseline before the agent cannot demonstrate the gain afterwards: all they will have left is the vendor's word. Measurement beforehand is the only part of the work that cannot be recovered.
FAQ
What ROI can a small company expect from an AI agent?
There is no universal figure. The result depends almost entirely on three variables: task volume, unit time saved and the recovery rate you are willing to retain. With the assumptions in the table above, payback takes eleven months. With a third of the volume, it never happens. The calculation itself stays the same.
Should internal implementation time be counted?
Yes, and it is the line most often missing from presentations. Scoping, data preparation, acceptance testing, training: in our example, 68 hours, about 3,000 euros. A project that ignores this time shows a payback period one to two months too short.
How do you value saved time when nobody redeploys it?
Two honest options. Either you value it at zero in the cash calculation and present the gain as available capacity, or you decide in advance what that time will be used for (sales, quality, new cases) and you verify that it happened. The second option is the only one that produces a verifiable financial gain.
What minimum volume makes an AI agent profitable?
It is calculated, not announced. Divide your recurring fixed costs by the net gain per task. In our example, 216.80 euros divided by 3.31 euros gives about 65 requests a month. Below that threshold, automation remains a convenience, not an investment.
Who should validate the assumptions in the calculation?
Management for the hourly cost and the recovery rate, the business owner for volume and unit time. A calculation validated only by the vendor selling the agent is not a calculation, it is a sales argument.
Conclusion
The ROI of an AI agent is defended line by line: a measured baseline, a loaded hourly cost, an explicit recovery rate, complete recurring costs, then a division. The model above fits in a spreadsheet and can be revised in an hour. What it demands is that you measure before you promise, and choose the assumptions before you know the result.
If you want to build this calculation on your own volumes, with a defined scope and verifiable figures, you can use the contact page: we start from your data, not from a benchmark. Deployment, for its part, is prepared upstream, and is presented through what it changes in day-to-day work (our AI agent services).


