What an AI agent is really worth: the ROI worksheet
“How much does it cost?” is the second question. The first — the one that actually decides — is “what does it return?” This is the method we use to put a number on a use case before launching it: six variables, one calculator, and the four mistakes that make the result lie.
The calculator below can answer “never.” That’s deliberate: a use case that doesn’t pay for itself is useful to know, and far better discovered on this page than six months into a project.
The calculator
Move the six sliders to match your situation. The starting values describe a common case: a team of six, each losing three hours a week to a repetitive task.
Those who actually do the task today — not the department headcount.
Time spent on the target task. Measured over a week, not estimated from memory.
Gross salary plus employer charges, divided by hours actually worked. Roughly €45/h for a €45k salary.
The fraction of the task the agent handles on its own. Above 80%, be suspicious: there are always edge cases.
Design, integration with your tools, testing and team training. One-off.
Model calls, hosting, monitoring. It keeps running as long as the agent does.
Under twelve months, the project makes its own case to a CFO.
- Hours recovered per month
- 47 h
- Gross monthly gain
- €2,104
- Net monthly gain
- €1,954
- Year-one result
- €15,453
- Then, every year
- €23,453
Recovered hours only turn into money if they are redeployed. An hour handed back to someone who spends it differently isn’t a saving — it’s comfort, and it’s better to call it that in front of a leadership team.
The six variables — and what they hide
Six is enough. Beyond that, you’re no longer calculating a return on investment: you’re building a justification — and a thirty-line model always ends up finding the answer you wanted.
1. Time actually spent — measured
Not estimated from memory: measured over one week, by the people who do the task. The gap between estimated and actual time routinely exceeds 40%, in both directions.
2. How many people do it
The people who perform the task, not the department headcount. A task done by two people out of twelve isn’t costed on twelve.
3. Fully loaded hourly cost
Gross salary plus employer charges, divided by hours actually worked. Roughly €45/h for a €45k salary — a number your payroll team can give you in five minutes.
4. The share that can really be automated
This is the variable everyone overestimates. An agent handles the normal flow; edge cases come back to a human, and there are more of them than anyone admits. Above 80%, the number is probably wrong.
5. Setup cost
Design, integration with your tools, testing — and training the teams who will use it. That last line is the one people forget, and it’s the one that decides whether the tool is still in use six months later.
6. Running cost
Model calls, hosting, monitoring. It runs as long as the agent does, and it scales with volume: an agent that pays off at 100 cases a month may not at 10,000.
The four mistakes that make the numbers lie
Counting hours that won’t be redeployed
Twenty minutes a day handed back to eight people adds up to about three-quarters of a full-time role — but only if that time actually goes somewhere else. Otherwise it’s comfort, not savings. Calling it that in front of a finance team saves you from being contradicted at the first review.
Forgetting the cost of review
An agent whose every output has to be fully re-read doesn’t save much. What saves time is an agent that’s reliable on the normal flow and flags the cases it can’t handle on its own. The review rate you’ll need is a project input — measure it during the pilot.
Costing the build but not the adoption
Development is the visible part. Training, unlearning old habits and the three weeks when the team runs both the old process and the new one are a real cost. Counting them in the setup makes the calculation honest — and surprises less likely.
Modelling a task that’s about to disappear
Automating data entry that next year’s ERP migration will eliminate is the surest way to pay twice. Before calculating anything, check the task will still exist in two years — the least technical question on the list, and the most profitable one.
The threshold below which you shouldn’t bother
Our internal rule, learned from projects that should never have started: below fifteen hours recovered per month, a custom agent isn’t worth the project. Not because the maths would be negative — it can look positive on paper — but because a tool that saves little is a tool people forget to use, and a tool nobody uses returns nothing at all.
Below that threshold, two options do better: a simple automation with no agent, or training — teaching the team to do the task twice as fast with the tools they already have costs far less and often delivers the same gain. The 12 highest-ROI automations →