AI ROI: quantify an agent's return before you sign
AI ROI: how to quantify what an agent gives back in an SME, from time actually freed to supervision cost, and the cases where the maths says no.
AI ROI: what the calculation actually measures
AI ROI compares what a system gains you to what it costs. The formula fits on one line: net annual gain divided by total investment, expressed as a percentage. In an SME that percentage mostly serves to reassure leadership, and it is often wrong because the gain was estimated loosely.
The number that actually decides is the payback period, meaning how many months before the agent has repaid its setup. A hesitant leadership team understands "repaid in seven months" better than "170% ROI", because the first phrasing says when cash flow turns positive again.
| Metric | What it tells you | When to use it |
|---|---|---|
| ROI percentage | Return over a period | Comparing two projects |
| Payback period | The month the investment is repaid | Getting a budget approved |
| Cost per task handled | What one unit of work costs | Deciding whether to widen scope |
We compute all three on the projects we scope, and we lead with the payback period. It is the one leadership can challenge without being an expert on the subject.
Sizing the gain: the time actually freed
An agent's gain is measured in hours handed back to people who then do something else. That second part is where it breaks down. Twenty minutes recovered three times a day across scattered tasks do not add up to one useful hour, they dissolve into the day without anyone being able to say where.
For a saved hour to count, it has to be consolidated and reassigned to something you can name. An agent that handles incoming requests overnight frees a real block in the morning. An agent that saves two minutes per email frees nothing you will be able to measure six months later.
- Consolidated hours, not time nibbled away in fragments
- Tasks that leave the schedule, not tasks that merely get faster
- A larger volume handled without hiring, when the demand already exists
- Errors avoided whose unit cost you actually know
The strongest case is still the one where the agent absorbs a peak. If you were turning work away for lack of hands, the gain reads straight off revenue, with no need for the shaky conversion of hours into euros.
Supervision cost belongs in the calculation
Cost reads on two lines: setup, paid once, and run, which comes back every month. Our guide on agent budgets details the ranges. What matters here is the line almost nobody puts in the spreadsheet: human review time.
If one person validates thirty outputs a day at a minute each, you have just added half an hour daily to someone's schedule. Over a year that is more than a hundred hours. That half hour belongs in the denominator, exactly like tokens and hosting.
The cost falls over time, provided you measure it. When the acceptance rate settles very high for several weeks, you can raise the threshold and only route the high-stakes cases to a human. Without that measurement, supervision stays at day-one levels and the ROI degrades quietly.
The traps that skew the calculation
Four mistakes keep showing up in the calculations people bring to a scoping session, and they all lean the same way: they flatter the result.
| Trap | What it produces | The fix |
|---|---|---|
| Counting raw time saved | A ROI twice too good | Count only reallocated hours |
| Leaving out human supervision | An underestimated run cost | Measure actual review time |
| Valuing at the billed rate | An inflated gain | Use the fully loaded cost of the role |
| Calculating over three years | An unverifiable promise | Calculate over twelve months, revise after |
The last trap is the most common, and the most convenient. A three-year horizon gets almost any project approved, because nobody comes back to check. Twelve months forces you to pick a case where the gain arrives quickly, which is exactly the right constraint for a first agent.
When the calculation says no
A good calculation also exists to stop a project. Four situations keep coming up, and in all four we would rather say so before sending a quote.
- The process changes every quarter, so the agent will be obsolete before it pays for itself
- The volume is too low, and one hour a month will never repay a setup
- The data you need does not exist in usable form, and producing it costs more than the gain
- The task requires judgement nobody on the team can express as rules
In those cases a scoping workshop or plain automation serves you better than an agent. It shows up in twenty minutes of conversation, and it avoids a project that drags for six months before someone dares to stop it.
The method in five steps
The method fits in five steps, and the first one takes longer than the other four combined.
- Measure the current situation for two weeks: volume handled, time per unit, error rate
- Pick a single task, the one with the steadiest volume
- Size the gain over twelve months, counting only reallocated hours
- Size the full cost, setup, run and supervision included
- State the payback period, and set a checkpoint at three months
Without step one, the next four rest on estimates nobody can confirm or contradict. That is the precise moment most ROI calculations lose their value, and it is also the step everyone wants to skip.
If you would like to quantify the return of a first agent on a real process, we can set the assumptions together in 20-40 minutes.
FAQ
- What ROI should an SME expect from an AI agent?
- No honest range exists out of context, because the gain depends on the volume handled and your team's hourly cost. The right move is to compute the payback period on your own case, using only the hours you can genuinely reallocate.
- Should you calculate ROI before or after the POC?
- Both. The one before rests on written assumptions and decides whether the subject deserves a POC. The one after uses the observed numbers and decides whether to go to production.
- How do you value an hour saved?
- Use the fully loaded cost of the role, salary plus indirect charges, not the hourly rate you bill. The billed rate overstates the gain unless you actually resell the freed hour to a client.
- Does a quality gain count in the ROI?
- It counts when you can price an error: a wrong invoice, an order to redo, a lost client. If that unit cost is unknown, keep quality as a qualitative argument and leave it out of the calculation.
- What period should you calculate over?
- Twelve months. Beyond that, uncertainty on model pricing and on your own process makes the number unverifiable, and nobody comes back to check a three-year promise.
- Do token costs make ROI unpredictable?
- They vary, but they can be bounded: you can cap the budget per request and measure real consumption over two weeks. Human supervision time is what makes a ROI unpredictable, far more than tokens.
- What if the calculation comes out negative?
- Narrow the scope before giving up. An agent on a single high-volume task is often profitable when the same agent spread across five tasks is not.
Sources and references
- AI Index Report
Adoption and economic impact figures, with their methodology. Prefer it to vendor studies when you need to quote an order of magnitude.
- Hourly labour costs
The real hourly labour cost per country. That is the figure for the numerator, not the rate you bill a client.
- DORA software delivery performance metrics
Four metrics that resist gaming. The model to copy when you want to measure an effect rather than an activity.
Scope your first AI agent
20 minutes to review your tools, data and the first useful case. No jargon, no commitment.