Starting with AI Without Knowing What You Measure
By Anis Hammouche·May 31, 2026
Starting an AI tool without having written down what you want to measure takes away the best way to know whether it helps. A 2026 KPMG survey gives the scale: 95% of the companies surveyed have an AI strategy, but only 8% show a measurable return. I look at what that gap suggests at your size, then give you three lines to write before you start anything.
What the survey says, and what it doesn't
KPMG surveyed 2,110 executives in 20 countries, in organisations with at least $100 million in annual revenue. According to its press release of 28 May 2026 (in French), 95% have an AI strategy and 64% say they already get real benefits from it, but only 8% show a measurable return on investment.
These are not freelancers or four-person agencies, and the survey doesn't say you're in the 8%. It doesn't say why the gap exists either. At your size, here is a plausible explanation, which I give you as a hypothesis.
The possible mechanism, at your size
Take Théo, a freelance communications consultant (an invented case). In January he subscribes to a writing assistant, in March to a transcription tool, in May to a third one to sort his emails. Each one "saves him time". In September he pays for three subscriptions and couldn't say which one he would keep if one had to go. If he had written down, before the third, that he spent six hours a week on his emails (an invented figure), he would see in one line whether the tool really gave him two of them back.
Now imagine a small agency, also fictional: "we automate the meeting notes", the tool runs, everyone is happy, and nobody wrote down how long they took before. There is nothing to compare against, so there's no answer. The calculation to run once you have your starting number is in AI ROI: How to Measure What a Tool Really Earns You.
Declaring a benefit is not measuring it
In the same survey, 64% of companies report real benefits and 8% show a measurable return. These are not the same thing: you can find a tool useful without being able to put a figure on it. The same press release notes that in France, 50% of the companies surveyed say they struggle to measure the indirect benefits of AI, against 44% internationally. The Cigref, which brings together large French groups, writes in the presentation of its report of 15 January 2026 (in French) that justifying a generative AI project like a simple server migration "is a dead end".
IBM describes a similar gap in its guide on AI ROI: according to the discussions of its Think Circle in the fourth quarter of 2025, 79% of executives see productivity gains, but only about 29% say they can measure ROI with confidence (IBM's text says "about 29%").
At your scale, the difference reads like this:
| What you feel | What you can prove |
|---|---|
| "I save time" | The number of hours spent on one specific task, before and after |
| "My replies to clients are better" | The average reply time, or the number of follow-ups needed |
| "I'm less tired in the evening" | Nothing you can count: it's comfort, to be counted as such |
| "This tool is great" | What it costs you per month against what it earns you |
The left column is felt. The right column is recorded. Moving from one to the other doesn't depend on the tool, but on having written down the starting point before you begin.
Three lines to write before you start
You don't need a dashboard. Before you pay for or plug in anything, write these three lines:
Task aimed at: [the specific task, for example "replying to quote requests"]
Starting number: [the time it takes me today, or what it costs me, recorded over two weeks]
Review date: [a date two to four weeks away, when I will record the same number]
Here is the same template filled in, with invented values:
Task aimed at: replying to quote requests
Starting number: 15 minutes per request, 20 requests a month, so 5 hours
Review date: 15 December, when I record the time spent with the tool
If you can't fill in the second line, you're not ready to start: first record how long the task takes you. If you can't fill in the first, you're subscribing to a tool, not a solution. To pick the task, five detailed automation cases will help. And it matches what I said in Stop Stacking AI Tools, Start Architecting: one more subscription doesn't replace a chosen task.
One practical consequence: on the review date, you have the right to cut. If the number hasn't moved, the tool doesn't have to stay. A subscription often renews on its own until you cancel it, so put the review date in your calendar.
What it changes for your budget
In a large company, the finance department asks for the return. You are your own finance department: money you put into an unevaluated subscription doesn't fund anything else. Take three subscriptions at €20 a month (an invented example): that is €720 over the year. If only one of those tools gave you time you then billed, you know which one to keep. For the other two, you have no proof. It's up to you to say what each line earns, including when the answer is "nothing I can count, it's comfort".
Key points
- The KPMG survey covers large companies. It shows a gap between declaring a benefit and measuring it.
- Three lines before you start: the task, the starting number, the review date.
- On the review date, if the number hasn't moved, you cut.
- The detail of the calculation, with an hourly rate or a margin, is in the article on ROI.
Frequently asked questions
What does "measuring the ROI of AI" mean for a freelancer? Comparing a number recorded before the tool with a number recorded after, on the same task: hours, a delay, a cost per project. Without the first reading, you only have an impression.
Why do so few companies show a measurable return? The KPMG survey doesn't settle it. The starting number has to be taken before launch, and people subscribe faster than they record it.
Does measuring slow down the launch? Writing three lines takes a few minutes. Recording a task's time over two weeks takes a bit longer, but you don't have to wait before trying the tool: you record while you use it as a trial.
The next step is From Solo to Small Team: What AI Replaces, and What It Will Not, which looks at what your work becomes once the tasks you measured are delegated. My next tests go out through my newsletter, you can sign up right below.
Sources
- KPMG, IA : 95 % des entreprises ont une stratégie, mais seulement 8 % en mesurent réellement le ROI, press release of 28 May 2026 (survey run from 19 February to 17 March 2026; in French)
- IBM, How to maximize AI ROI in 2026 (Think Circle discussions: 79% and about 29%)
- Cigref, Évaluer le retour sur investissement des solutions d'IA générative et agentique, 15 January 2026 (in French)
