The cost of AI in a mid-sized company: what you really pay, and what you never measure
The cost of AI is not on the subscription invoice. Checking the output, uses that creep, time saved that evaporates: the three figures to track.
A scene I keep seeing in leadership team meetings, almost word for word. I ask what artificial intelligence costs the company. The finance director has the answer, and it is precise: so many subscriptions, so many euros per month per person. The figure is small. Everyone is reassured.
Then I ask the second question: “And how much time do you spend checking what it produces?”
Nobody has a figure any more. Yet that is where the real cost of AI in a mid-sized company begins. Not on the vendor’s invoice, but in everything the invoice does not say.
Four companies in ten use AI, two in ten pay for it
The figures landed on 28 September. According to the France Num barometer 2026, published by the Direction générale des Entreprises and covering more than 9,000 companies, 40 % of French small and mid-sized companies now use artificial intelligence, fourteen points more than a year ago. Among mid-sized companies in the strict sense, the figure reaches 53 %.
But only 19 % pay for a solution. And when you look at what they use it for, generating text, voice or images comes first at 34 %, far ahead of task automation at 11 %.
Translated: AI entered these companies through the free tier, to write faster. It gives the impression of costing nothing. That impression is precisely what costs.
One detail in the same barometer stopped me. Among the companies using AI, 68 % judge its impact positive. Among those that pay, the figure rises to 85 %. You can read that as paying making you more demanding, and therefore more effective. You can also read something else into it, which I come back to below: we judge more favourably what we have chosen to buy.
The price falls, the bill rises
Internationally, BCG published its Applied AI Index 2026 on 30 September, based on 1,330 executives. Two figures deserve to be pinned up in your office. AI spending has doubled in less than a year, reaching 3.3 % of revenue. And more than 80 % of that spending now sits outside the IT budget.
Read the second one again. Four euros in every five spent on AI escape the budget of the person you assume is tracking them. In mid-sized companies I usually find them in expense claims, on the sales team’s cards, in subscriptions taken out on a Friday afternoon.
You will tell me prices are collapsing. That is true, and that is the trap. Bain documented it in a note published in June: between December 2024 and December 2025, the price of tokens, the billing unit for these models, halved. Over the same period, consumption of them rose 4.5-fold. The authors’ conclusion: the effective cost per task often stays flat. Every price cut is immediately swallowed by new uses, and by the reflex of moving to the newest model, which is also the most expensive.
Uber went through this at scale. To encourage the use of AI coding tools, the company had set up internal leaderboards ranking its engineers by consumption. Its chief technology officer acknowledged, in an interview reported by Fortune, that the year’s AI budget had been exhausted within months. The way out is instructive: Uber did not cut off access. It revised its default models, improved request caching, and made visible to each engineer what their own usage cost. It treated the spending as an engineering problem rather than a budget problem.
Hold on to the mechanism. Consumption had been rewarded. Consumption is what they got.
The cost nobody invoices: checking the work
The heaviest item appears on no invoice. It is the human time spent verifying, correcting, redoing.
Researchers at Stanford and BetterUp Labs gave it a name in 2025, workslop: documents produced by AI that look polished but have nothing inside. In their survey of 1,150 American employees, published by the Harvard Business Review, 40 % said they had received some in the previous month, and spent on average one hour and fifty-six minutes dealing with each case.
The cost does not fall on the person producing the document. It falls on the person receiving it. Picture it: your sales manager saves twenty minutes on a proposal, your operations director loses two hours understanding it. The balance is negative, and no dashboard shows it.
And the time genuinely saved, where does it go? Anders Humlum and Emilie Vestergaard, of the University of Chicago and the University of Copenhagen, followed around 25,000 Danish employees across eleven occupations exposed to AI, matching their answers with administrative employment data. Their 2025 study measures an average gain of 2.8 % of working time. Employees say they reassign 80 % of it to other tasks. AI even created new work for 17 % of users. The final result: no measurable effect on hours worked or on pay.
The time saved does not disappear. It dissolves. Nobody decided what it was for, so it serves everything, which is to say nothing measurable.
Why your brain overestimates the return
Here I change hats. Because the problem is not only an accounting one: our brain is poorly equipped to assess this kind of investment.
In 2011, Tali Sharot, Christoph Korn and Raymond Dolan published a study in Nature Neuroscience that has stayed with me. They showed that we update our beliefs asymmetrically. Unexpected good news is taken in without difficulty. Bad news is taken in far less well, and activity in the region of the prefrontal cortex that ought to process it is weaker in the most optimistic people.
Apply that to a leader who has brought AI into the company. Every successful demonstration, every email written in ten seconds, is recorded as confirmation. The two hours their operations director spends checking the output slide past. Not out of dishonesty: by construction.
The second trap is the feeling of speed. I raised it when writing about decision fatigue: when a machine answers instantly, we feel we have moved forward, whether or not we have. The feeling of progress is enough to satisfy, whether or not the result follows. Uber’s leaderboards did nothing but feed that feeling.
On one side, an AI that saves time. On the other, a brain that needs an outside figure to know whether that is true.
The manual: three figures to ask for on Monday
The good news: you do not need a full-time financial controller. In the companies I work with, three figures are enough to get out of the fog.
One: the cost per completed task. Not the price of the subscription, not the price of the token. Pick a single process, chasing an unpaid invoice, answering a tender, writing a product sheet. Work out what it costs end to end, including the human verification time. Bain puts it bluntly: most companies have no idea what they spend per task, and that figure cannot be reconstructed after the fact. Start now, on one process.
Two: where the time saved goes. Before deploying a use case, write down what the freed-up time is for. More client meetings, fewer overtime hours, a project put off for a year. If nobody can answer, the Danish study has already told you what will happen.
Three: the share of uses that do not need the best model. Using the most powerful model to write an email is flying business class to buy bread. Bain observes cost differences of three to five times between companies that match the model to the task and the rest.
I will add a question to put to any supplier before signing: “What happens to my bill if our usage goes up tenfold?” If the answer is vague, you know what you are buying.
What the cost of AI says about your organisation
I explained in an earlier piece why adopting AI fails so often in mid-sized companies: the brake is organisational, not technical. Cost confirms the diagnosis from another angle. A company that does not know what its AI costs is a company where nobody has decided what it is for.
The cost of AI is therefore not, in the first place, a finance director’s subject. It is the leader’s. The price of models will keep falling, usage will keep swelling, and between the two, only one clear decision will make the gain appear in your accounts: the one that says what the freed-up time is for, and then checks.
Do the exercise this month, on a single process. If the figure surprises you, it was time you calculated it.
If you would like us to look together at what AI actually costs and returns in your company, I am glad to take thirty minutes.
Frequently asked questions
What does AI really cost a mid-sized company?
The price of the subscriptions is only the visible part. The real cost has three items that are rarely tracked: the time spent checking and correcting what AI produces, consumption that rises as teams multiply their uses, and duplicate tools bought outside the IT budget. A BCG study published on 30 September 2026, covering 1,330 executives, shows that more than 80 % of AI spending now sits outside the IT budget. For a mid-sized company, the only reliable measure is the cost per completed task, calculated on one specific process.
How do you measure the return on investment of AI in a company?
By starting from a task, not from a tool. Pick one repetitive process, chasing an unpaid invoice, answering a tender, reconciling accounts, measure its full cost before AI, then after, including the human time spent verifying. Add a second measure: what the teams do with the time freed up. Time saved that is assigned to no identified task never shows up in the accounts.
Should you always use the most capable AI model?
No. The most advanced models cost appreciably more and are only useful for tasks that require reasoning. Writing an email, summarising a document or extracting a date does not need them. Bain observes cost differences of three to five times between companies that match the model to the task and those that use the same one everywhere.
Why do the time savings from AI never show up in the results?
Because the time saved is almost always reabsorbed. A study of around 25,000 employees in Denmark (Humlum and Vestergaard, University of Chicago, 2025) measures an average gain of 2.8 % of working time, 80 % of which is reassigned to other tasks, with no effect on hours or on pay. Until the company's leader decides what that time is for, it dissolves into everyday activity.