Measuring the return on AI without fooling yourself
- 1Hours saved are not a financial gain until those hours are redeployed or translate into avoided costs.
- 2Without a measured baseline before deployment, no ROI can be demonstrated afterwards.
- 3Full cost includes human review, maintenance, data work and compliance, not just licences.
- 4A decision threshold and a review date set in advance prevent disappointing projects from running on indefinitely.
In many executive committees, the return on AI fits in one sentence: "we save ten hours a week." The sentence is often true and almost always insufficient. It does not say what was measured, against what, at what full cost, or what the company did with the hours it freed. This guide sets out seven steps to produce a figure that a CEO or CFO can defend without caveats.
Why people fool themselves without bad faith
Three biases recur in SMEs and mid-sized companies. First, the project sponsor is also the person measuring it. Second, benefits are counted gross while costs are counted partially, often leaving out review time and maintenance. Third, the pilot is run by the most motivated people, whose results do not carry over to the whole team.
None of this requires dishonesty. It only requires not setting rules before starting. That is what the steps below put in place. They reflect our approach to finance and performance, where measurement is part of the deployment and not a report written afterwards.
Step 1: fix the starting point
Before any deployment, measure the process as it is: volume handled, time spent, end-to-end lead time, error rate, rework. Two weeks of real observation are worth more than an estimate made in a meeting. Note seasonality as well: a quarter-end close cannot be compared with a quiet month.
Step 2: choose a unit of value tied to the P&L
An hour saved is an activity indicator. A unit of value is an effect you can connect to a financial line: an avoided hire, eliminated overtime, a shorter collection period, one dispute fewer, added sales capacity. Choose one primary indicator, two at most, and name the owner of the budget concerned. If nobody can say which line will move, the ROI remains a hypothesis.
Step 3: count the full cost
The real cost of an AI initiative goes well beyond the licence. At a minimum, we suggest listing:
- licences and usage (subscriptions, consumption, hosting);
- initial integration, amortised over three years, a prudent period for a tool category that changes quickly;
- data preparation and quality;
- human supervision and review time, often forgotten;
- maintenance, updates and training;
- compliance: personal data handling under the Swiss revised FADP (nLPD), in force since 1 September 2023, and, depending on the use, preparation for the EU AI Act, most of whose obligations apply from August 2026.
On that last point, our data, AI and compliance expertise explains how to build it into the budget from the start rather than discovering it at the end of the project.
Step 4: separate hours freed from real cash
This is the step where numbers shrink the most. A freed hour becomes a financial gain in only three cases: it avoids a spend (a hire, temporary staff, overtime), it is redeployed to an activity that produces measurable value, or it allows additional volume to be handled at no extra cost. In every other case the benefit is real, but it is capacity, not cash. Both can be managed, but they should not be added together.
Step 5: deal with attribution
To know whether AI explains the improvement, compare like with like. Depending on team size there are two options: a control group (a team or entity not using the tool), or a before-and-after comparison adjusted for seasonality and other changes (a new ERP, a departure, a reorganisation). Document these factors. A gain credited to AI when a hire happened in parallel will not survive the first review.
Step 6: include quality and risk
A faster but less reliable process can cost more than it earns. Track error rates, rework volumes and incidents. For a forecast or a report, measure the gap between forecast and actual before and after. This is what a continuous cash flow forecast makes possible, since reliability can be checked week after week.
Step 7: set the decision threshold and the review date
Before launch, write down three things: the expected net result, the threshold below which you stop or correct, and the review date, typically at 90 days and again at 12 months. This rule avoids the classic trap of prolonging disappointing projects because no one defined what failure looked like. It follows the same logic as our article on funding the AI initiatives that really matter.
A worked example, under explicit assumptions
Take a company of 120 employees whose finance team uses an AI tool to prepare the monthly report. Assumptions, purely illustrative: a loaded hourly cost of CHF 100, 46 working weeks, and a reported saving of 10 hours per week.
- Gross gain: 10 h × 46 weeks × CHF 100 = CHF 46,000 per year.
- Licences: CHF 12,000.
- Integration of CHF 25,000 amortised over three years: about CHF 8,300 per year.
- Supervision and review, 3 h per week: 138 h, or CHF 13,800.
- Maintenance, data and training: CHF 5,000.
The full annual cost comes to about CHF 39,100, and the net benefit falls to about CHF 6,900, less than 15% of the announced gain. More importantly, that amount is cash only if the 460 hours are redeployed or avoided. If the team simply absorbs them without changing how it works, the cash effect is zero and the project costs CHF 39,100 for capacity that is not used.
The example does not condemn the tool. It shows that value depends on the management decision that follows: reducing reliance on temporary staff, absorbing growth without hiring, or devoting the freed time to analysis. That is where the augmented CFO comes in, whose value lies as much in how saved time is used as in the tool itself.
Making measurement last
Once the rules are set, measurement must run without manual effort. An automated executive dashboard tracks the primary indicator, the full cost and the gap to target continuously, using the same definitions from one month to the next. Without it, ROI turns back into an annual presentation that nobody can cross-check.
Where to start
Pick one use case already in place. Reconstruct its starting point, even roughly. Recalculate its full cost. Ask the budget owner which line has really moved. Done over a few days, this exercise is usually enough to reclassify your initiatives: those to continue, those to correct, those to stop.
If you would like to run this exercise with an outside view, our diagnostic helps frame the measurement, the full cost and the decisions to take on your AI initiatives in a short time.
A personalised first reading in three minutes.