The formula for ROI, return on investment, does not change when the project concerns artificial intelligence: gains, less costs, then related to those costs. It is its two terms that are established differently. The cost continues after go-live: an AI system has to be operated, monitored and re-evaluated. The gain, when it takes the form of freed-up time, remains theoretical as long as nobody has decided how that time will be used.
The ROI of artificial intelligence is therefore measured on three sets: the full cost, from build to operation; the gain observed on an indicator that the business was tracking before the project; the timescales, from the ramp-up in use to the recovery of the outlay. Each figure rests on conventions and leaves something out of scope. The assumptions are set before launch, in the business case framed by our AI strategy support; this article deals with measuring them.
The conventions behind an artificial intelligence ROI figure
In its usual form, ROI relates the net gain (gains less costs) to the costs, over a period. A result of zero then means that the project has just paid back its cost. This form is a convention and others are in circulation: relating gross gains to costs places this break-even point at one, for the same project. The figure is therefore accompanied by its formula.
The period, the scope of costs and the baseline also shift the result. A build cost incurred once weighs more over one year than over three, a calculation stopped at the build ignores operation, and a gain is read against a starting state.
Two neighbouring measures answer other questions. The payback period indicates when cumulative gains catch up with cumulative costs. Net present value adds up the expected flows after discounting them at a rate set by the company. An AI project benefits from being presented with the instrument that the finance department applies to other investments: the cases can then be compared with one another.
The full cost continues after go-live
The cost items, one-off and recurring, are described in our article on the budget for an AI project. The ROI calculation takes them all into account, including those that no invoice materialises, such as the time of business experts called upon for the evaluation set.
Operating expenses are counted for each financial year in the period. When the model is billed by the volume of text processed, this line grows with adoption: our entry on LLMOps describes tracking the cost per request. Continuous evaluation and human review of outputs consume working time, to be deducted from the time saved. For a high-risk system within the meaning of Regulation (EU) 2024/1689, there is in addition the cost of the obligations presented in our article on the AI Act’s obligations for businesses.
The pipeline that carries these operating items is described on our MLOps and LLMOps page.
This total remains tied to its assumptions. It holds for one volume and one price, both of which may change during the period of the calculation. It also depends on the share of shared foundations (platform, governance) charged to the project, according to an allocation key that is written down.
A gain is measured against a baseline
A gain is observed as the change in an indicator that the business was tracking before the project: a processing time, a rework rate, a volume processed with constant headcount, a subcontracting spend. The starting value is recorded before go-live, over a comparable period. Once that moment has passed, it can only be recovered if the applications have kept its history; failing that, it is estimated.
The before-and-after comparison also records what has changed elsewhere: the volume of activity, the season, a reorganisation. When the process lends itself to it, two groups observed over the same period, one equipped and the other not, better isolate the effect of the system.
Freed-up time requires one more convention. Hours multiplied by an hourly cost give a theoretical gain, which only becomes effective through a decision: absorbing an increase in activity, reducing subcontracting, assigning people to another task. We suggest writing a redeployment rate into the calculation: the share of freed-up hours that the business owner undertakes, in writing, to reassign.
This measure does not guarantee how long the gain will last: use may fall back once the novelty has worn off, and the quality of a model may degrade. As for the gain gathered by questionnaire from users, it measures a perception, which cannot be added to observed amounts.
Between go-live and the gain, timescales to be counted
The gain from an AI service follows its use. As long as not all the target users are using it, it remains below the value entered in the calculation, and a calculation that applies it in full from the first month brings forward the payback accordingly. The share of target users who use the system each week, recorded month after month, gives the curve to adopt.
When flows are regular, the payback period is obtained, in its simple form, by dividing the initial outlay by the annual margin, that is the annual gain less the annual operating cost; a variant discounts the flows before accumulating them. It says nothing about what comes next: two projects paid back at the same moment may then bring in very different amounts.
The ROI in a business case is a forecast. It is recalculated with the observed values, at dates set in advance: the decision points placed by the enterprise AI roadmap.
The same project changes ROI depending on what is counted
Take, by way of hypothesis, a logistics company whose customer service responds in writing to delivery complaints. It puts into service an assistant that prepares the responses, which are reviewed by an adviser. The amounts that follow are invented for convenience of calculation: they describe no real project and are not an order of magnitude.
Build and integration: 240,000 euros, once. Onboarding: 30,000 euros, once. Operation, including continuous evaluation: 80,000 euros per year. Time freed up: 4,000 hours per year, valued at 50 euros an hour, i.e. 200,000 euros per year. The calculation covers three years.
| Reading | Costs counted | Gains counted | ROI over three years |
|---|---|---|---|
| Build only, all hours valued | 240,000 euros | 600,000 euros | 1.5 |
| Full cost, all hours valued | 510,000 euros | 600,000 euros | About 0.18 |
| Full cost, 70% of hours redeployed | 510,000 euros | 420,000 euros | Negative: 90,000 euros short |
The three calculations are correct and do not answer the same question. Over five years, the third reading becomes positive again, at a little over 0.04. The payback period goes from two years and three months in the second to four and a half years in the third.
An ROI figure is therefore presented with its notes:
- the formula, the period adopted and the date of the calculation;
- the cost items counted, and those that are not;
- the business indicator, its baseline value and how it was recorded;
- the redeployment rate for freed-up time and the name of the person who undertakes to meet it;
- the date on which the calculation will be redone with the observed values.