Guide
To calculate the ROI of an AI project, value the hours actually freed (hours × fully loaded hourly cost), subtract build and running costs, then set the net gain against the cost. The payback period, in months, is the figure a committee reads first; it holds only if every assumption is written down.
Contents
Calculating ROI and payback in six lines
Value the time that is really freed, subtract the costs, then work out how many months it takes to recover the investment. Each line hides an assumption to be written down.
- Hours freed per year = people concerned × weekly hours spent on the task × weeks worked × share that can be automated.
- Gross annual value = hours freed × fully loaded hourly cost (salary and employer charges, per hour worked).
- Realised value = gross value × realisation rate (the share of freed time that is actually redeployed or avoided).
- Net monthly gain = realised value ÷ 12 − monthly running costs.
- Payback period (months) = build cost ÷ net monthly gain.
- ROI over a period = (net monthly gain × number of months − build cost) ÷ build cost.
A freed hour is not a euro saved
Multiplying hours by an hourly cost gives a capacity value, not a cash saving. What happens to the freed time decides how solid the figure is.
| What happens to the freed time | Financial effect | Strength before a committee |
|---|---|---|
| Overtime, temporary staff or subcontracting avoided | Lower spend, visible in the accounts | Strong |
| Hiring avoided or postponed | Payroll cost avoided, to be decided explicitly | Strong once the decision is made |
| Extra volume absorbed with the same headcount | Cost avoided, checkable against volumes handled | Medium |
| Time redeployed to higher-value work | Real value if the work exists and can be measured | Medium to low |
| Time freed with no new use | None | Nil |
The costs to count: build, running, internal time
A defensible figure separates what you pay once from what you pay every month, then adds what nobody invoices.
| Item | Nature | Why it gets forgotten |
|---|---|---|
| Scoping, development, testing, integration | One-off | Often the only line in the quote |
| Preparing data and access | One-off | Its scale only shows up along the way |
| Business teams’ time: interviews, tests, validation | One-off, internal | No invoice, so no line |
| Training and change management | One-off, then ongoing | Missing from a technical quote |
| Subscriptions, model usage, maintenance | Every month | Absent from calculations that stop at launch; consumption grows with usage |
| Human oversight: reviewing and validating outputs | Every month | It reduces the net gain but remains necessary |
Illustrative example (estimated figures): an accounting firm
SHIFTLABS’s first reference engagement: an accounting firm in France, client anonymised. Every figure is an estimate: 6 agents in production, ~1,337 h/year that can be automated, ~€60k a year of estimated value, a ~€14k build, an estimated payback of ~4 months.
Read them as a committee would. First, €60k for 1,337 h implies an hourly cost of about €45; in your own paper, that rate will come from finance. Second, this is freed time priced at an hourly cost, not a cash saving. The straight-line calculation gives €14k ÷ €5k a month, or 2.8 months; the estimate says about 4 months, because use ramps up after the two-week go-live. Finally, these estimates do not break down running costs: your own paper will need to quantify them.
What happens to the payback if only part of the value is obtained?
| Share of the estimated value actually obtained | Annual value | Payback (€14k build) |
|---|---|---|
| 100% | €60k | 2.8 months |
| 75% | €45k | 3.7 months |
| 50% | €30k | 5.6 months |
| 25% | €15k | 11.2 months |
Six pitfalls that distort the ROI of an AI project
These errors inflate a figure with no warning before the first measurement.
- Unrealised gains: counting 100% of the hours as saved. A task is rarely automated in full; justify each share.
- Adoption: an agent nobody uses earns nothing. Track usage from the first month and cost a half-usage scenario.
- Hidden costs: forgetting running costs, oversight and your teams’ time. The net gain falls accordingly.
- Double counting: counting the same hour in two use cases that touch the same task.
- No baseline: without times recorded before the project, no gain can be proved afterwards. Measure the work actually done, not the work declared (the example interview, which is fictional, shows how a task is quantified).
- Single measurement: one reading, too early or at peak usage, misleads. Set the measurement dates in advance.
Presenting a defensible figure to an executive committee
A committee judges the strength of a figure more than its precision. One page is enough if it holds these six elements.
Have the relevant business lead validate each assumption: a figure with no owner does not survive the first question.
| Element of the paper | What the committee should find |
|---|---|
| The decision requested | Budget, scope, date, and the payback in months on the first line |
| Three scenarios | Cautious, central, ambitious, and the assumptions that change |
| The assumptions register | For each one: value, source, status (measured, estimated or assumed) and business owner |
| The full costs | Build, running, oversight, internal time |
| The cumulative curve | Net value month by month, with the break-even point marked |
| The stop threshold and follow-up | Usage level below which you correct or stop; indicators (time per task, volume handled, active users) and measurement dates |
Frequently asked questions
How do you calculate the ROI of an AI project?
Value the hours actually freed (hours × fully loaded hourly cost × realisation rate), subtract running costs, then set the result against the build cost. The gain simulator runs this calculation on your own assumptions.
What payback period can you expect from an AI project?
We do not quote a market average: it would depend on scope, the volume of repetitive tasks and adoption. Our reference engagement, on a well-scoped set of repetitive tasks, is estimated at about 4 months. Be wary of “average ROI” figures given without a source or a scope.
What is the difference between ROI and payback?
ROI compares net gain with cost over a period. Payback says how many months it takes to recover the investment. The second is quicker to read, the first tells you what follows. Give both, with the period and the denominator used for ROI.
How do you value the gain if nobody loses their job?
Value what the freed time concretely allows: overtime or temporary staff avoided, volume absorbed without hiring, fewer errors. Otherwise, present a capacity gain and say so, without converting it into a cash saving.
Can you estimate ROI before building anything?
Yes, as an order of magnitude: the simulator gives a first figure in minutes. A figure based on the work actually done has to be measured: SHIFTLABS’s two-week diagnostic quantifies each use case (impact × feasibility × cost), and business experts validate every conclusion.
Further reading
- Gain simulatorEstimate the gain from a first AI agent using your own assumptions.
- Express diagnosticAssess your AI maturity in 3 minutes.
- Example interviewAn AI interview replayed step by step.
- ContactBook a 30-minute call with SHIFTLABS.
And in your organisation?
In two weeks, AI agents interview your teams and domain experts validate every conclusion: you leave with costed use cases and a roadmap.