AI ROI: How to Measure What a Tool Really Earns You
By Anis Hammouche·May 30, 2026
To know whether an AI tool pays off, compare two numbers: the time you actually saved on one specific task, valued at your hourly rate or at a project's margin, and what the tool costs you in subscription, review time and setup time. Time saved that doesn't go back into billable work is not a gain.
Why the return on AI is so badly measured
In its guide How to maximize AI ROI in 2026, IBM cites a study of CEOs: about 25% of AI initiatives deliver the expected return, and only 16% have been scaled across the whole company. Those are large organisations, not freelancers. The mechanism applies to you too, though: you start on "it will save me time", with no figure to compare against, and three months later you can't say whether the tool was any use.
The Cigref, in the presentation of its report of 15 January 2026, Évaluer le retour sur investissement des solutions d'IA générative et agentique (in French), writes that justifying a generative AI project like a simple server migration "is a dead end". In other words, the usual calculation (one cost, one gain, done) isn't enough. You have to separate what can be counted from what can only be felt.
Separate what you can count from what you only feel
IBM distinguishes two kinds of return: "hard ROI", tangible effects directly tied to profitability, and "soft ROI", real benefits that don't show in revenue. For a small business, that gives:
| Type | Examples for a freelancer or small agency | How to check it |
|---|---|---|
| Direct | Hours saved on a specific task, billed or not; fewer errors to redo; cost per project | Count the hours before and after |
| Indirect | Calmer evenings, faster replies to clients, better morale | It matters, but it doesn't justify a subscription on its own |
A subscription is justified first by the direct return. The indirect one is a bonus: a tool that only holds up thanks to it is a tool whose profitability you'll never prove.
The calculation for a freelancer: your hourly rate
Clara is a freelance translator. She uses an AI assistant to prepare her quotes and her replies to clients. The figures below are invented for the example, and her rate is gross, before social charges and tax: the net gain is smaller. Here is the calculation, to redo with your own:
Gain for the month = hours saved in the month × [your hourly rate]
Cost for the month = [subscription] + review hours × [your hourly rate]
Return for the month = gain for the month - cost for the month
Setup (one time) = hours spent configuring the tool × [your hourly rate]
Say a rate of €60 an hour, a €20 monthly subscription and two hours saved per week, so eight hours a month. Count one hour of review a month and six hours of setup.
- Gain for the month: 8 × 60 = €480.
- Cost for the month: 20 + 1 × 60 = €80.
- Return for the month: 480 - 80 = €400.
- Setup: 6 × 60 = €360, so €40 of return in the first month and €400 in the following ones.
The trap is in the first line. Those eight hours are worth €480 only if Clara bills them to other clients. If she reinvests only four, the gain falls to €240, the monthly return to €160, and the setup is paid back during the third month. If she reinvests none, the gain is zero: she gained comfort, not money. That can be an acceptable result, as long as you know it.
To pick the task to measure, Five Concrete AI Automations for Your Business goes through five.
The calculation for a small agency: a project's margin
In Lina's five-person web agency, the hourly rate isn't what counts but the margin of a fixed-price project. Another invented example: a project sold for €6,000, estimated at 80 hours at an internal cost of €45 an hour.
| Before the tool | With the tool | |
|---|---|---|
| Hours spent (tool review included) | 80 | 70 |
| Internal cost | €3,600 | €3,150 |
| Margin | €2,400 (40%) | €2,850 (47.5%) |
The gain is €450 per project. With a tool at €60 a month and one project a month, that leaves €390. Also count six hours of setup at €45, so €270 once: the first month therefore brings in €120 (390 - 270). Same trap as for Clara: if the client asks for revisions that swallow the ten hours saved, the margin doesn't move. So the figure to track isn't the starting estimate but the hours actually spent on each project, taken from your time tracking.
The costs people forget to count
A calculation is only honest if the cost side is honest too. The subscription is the visible cost. The others add up:
| Cost | Example |
|---|---|
| Visible | Tool subscription, API calls |
| Setup | Hours spent configuring the tool, writing your prompts, testing |
| Review | The time to check each output before sending it to a client |
| Data | Checking where a tool hosts your clients' documents before you send them (the CNIL sets out its expectations, in French) |
For the detail of one of these costs, read The Real Cost of AI: The Three Hidden Costs Your Budget Forgets.
Measure over several weeks, not one day
IBM writes that "any immediate gains might be deceiving". For a narrow task, take a reading before you start, another after two weeks, and one more after a month. Invented example: Clara logs six hours a week on her quotes before the tool, four after two weeks, and still four after a month. The figure has settled, so she can decide. If nothing moves after that, the problem is the choice of task or the setup, not your lack of patience. Each time, note the same two numbers: the hours spent on the task and what the tool costs you.
Key points
- Pick a single task and record its time before you plug in the tool.
- Count the gain only on the hours you actually reinvest.
- Add setup and review to the subscription.
- Take the same readings after two weeks, then after a month.
Frequently asked questions
How long before I can tell whether an AI tool pays off? For a narrow task, two weeks give a first signal and a month confirms it or not. That's my personal rule of thumb, not a measured duration: the rarer the task, the longer you wait.
Should I count my own time in the cost? Yes. Your setup and review are time you don't bill: they weigh as much as the subscription.
How do I value saved time if I have no extra client to bill? Don't value it in euros. Write it down as a comfort gain, and decide whether that justifies the tool's price. That's a choice, not an ROI.
The next step is Starting with AI Without Knowing What You Measure, which is about starting without knowing what you want to measure. I share calculations like these in my newsletter, you can sign up right below.
Sources
- IBM, How to maximize AI ROI in 2026 (IBM study of CEOs: 25% and 16%; distinction between hard ROI and soft ROI)
- Cigref, Évaluer le retour sur investissement des solutions d'IA générative et agentique, 15 January 2026 (in French)
