Which AI use cases actually pay off for a small or midsize business?The ranking, with the evidence

For a small or midsize business, the AI use case with the best-established return is the customer service assistant. In the study by Brynjolfsson, Li and Raymond, published in 2025 in the Quarterly Journal of Economics, the tool raised the number of requests handled per hour by 15% on average across 5,172 agents, and by about 30% among the least experienced. The measurement comes from a single company, a large software vendor, and covers one specific task. Even with that limit, it is one of the most solid measurements a CEO has today.

I am publishing this ranking because the one you find when you search for “best AI use cases” is almost always a league table with no sources: return‑on-investment percentages lined up, with no word on who measured them or how. No leadership team should commit budget on that basis. For years I ran a vocational training group of around a thousand employees, and today I run my own AI agents. In both roles, I learned not to commit money to a number when I did not know the method behind it, and that kept me from several costly mistakes.

Why a sourced ranking now?

French small and midsize businesses are only now starting to adopt AI, behind large companies. According to Insee (France’s national statistics office), 18% of companies with 10 or more employees used at least one AI technology in 2025, compared with 6% in 2023. The gap by size is clear: 15% between 10 and 49 employees, 31% between 50 and 249, 58% above that (Insee Première No. 2120, July 2026).

That lag has an upside: you can now choose based on published results. They are still rare. In McKinsey’s global survey, conducted in May and June 2026 among 1,719 respondents, 37% attribute at least some positive effect on operating income to AI, and only 6% attribute at least 5% of their operating income to it while judging the value created significant (McKinsey, The State of AI, August 2026). For a small or midsize business, the point is to pick a first use case with a real chance of showing up in the P&L.

How is this ranking built?

Every line was screened against three rules.

A measured result. A use case makes the list only if its gain was measured against a group that did not have the tool: a controlled trial or a field experiment, published with its method. Self-reported surveys and vendor figures serve as context, never as a score.

A line in the P&L. Each use case is tied to the line it affects: revenue, personnel costs, purchasing. Time saved counts only if it turns into extra capacity or revenue.

The entry cost for a small or midsize business. For a comparable gain, a use case that starts with a subscription ranks ahead of one that requires clean historical data and a project lasting several months.

Rank	Use case	P&L line affected	Measured gain	Evidence	Entry cost
1	Customer service assistant	Personnel costs	+15% requests handled per hour, about +30% for the least experienced	Staggered rollout observed in real-world conditions, 5,172 agents	Low
2	Drafting everyday business documents	Personnel costs	-40% time, +18% rated quality	Controlled trial, 453 professionals	Very low
3	Demand forecasting and pricing	Revenue	+9.7% revenue in the test group	Field experiment, one online retailer	High
4	Software development assistant	Personnel costs	+26% tasks completed, but -19% speed for experienced developers on their own code	Three field experiments, 4,867 developers; one controlled trial with the opposite result	Low
5	Analysis and deliverables	Personnel costs	+25% speed and about +32% quality on suitable tasks, lower accuracy on a task outside the tool’s range	Controlled trial, 758 consultants	Low

Why does customer service come first?

It combines what the others have only in part: a measurement taken in real-world conditions, over several months, and a gain that shows up directly as capacity. The most useful finding for a small or midsize business is in the detail: the tool mainly helps recently hired agents, and does very little for the most experienced (Brynjolfsson, Li and Raymond, NBER, then Quarterly Journal of Economics, 2025). In practice, the tool passes part of the veterans’ know-how on to newcomers.

Small and midsize businesses know this problem well: every departure from customer service opens a period during which the replacement answers more slowly and less accurately. For a team of four, a 15% gain is a little more than half a person of capacity. That is an indicative calculation, and it assumes the volume of requests justifies that capacity. It also takes an up-to-date knowledge base of answers, since the assistant draws on what the company has already written.

Does everyday writing really deserve second place?

It does, on cost. In the trial by Noy and Zhang published in Science in July 2023, 453 managers and professionals wrote letters, memos and analysis plans. Those with access to the tool took 40% less time, for quality rated 18% higher by independent evaluators (MIT, study summary).

The measured time saving is the highest in the ranking, and getting started costs a subscription. What moves it down one place: the measurement comes from an online trial, on tasks of about half an hour, not from a company in operation. And the time saved shows up in the P&L only if someone decides what to reassign it to.

Why does the use case that generates the most revenue rank only third?

It is the only one on the list that directly affects the top line. At Rue La La, a US online flash-sale fashion retailer, a demand forecasting tool combined with price optimization raised revenue in the test group by about 9.7%, with a 90% confidence interval of 2.3% to 17.8% (Ferreira, Lee and Simchi-Levi, Manufacturing and Service Operations Management, 2016).

But it requires what many small and midsize businesses do not have yet: a clean sales history, a catalog broad enough to give the forecast something to work with, and prices that can be changed often. The result comes from a single retailer, with a very particular sales model. It is probably the most powerful lever on the list for a wholesaler, an online retailer or a manufacturer with a catalog, and much less so for a service company that works on fixed fees.

Why do two studies on software development contradict each other?

Three field experiments at Microsoft, Accenture and a Fortune 100 company, covering 4,867 developers, measure 26% more completed tasks with a coding assistant, with the gain going mainly to the least experienced (Cui et al., Management Science, 2026). By contrast, METR’s controlled trial published in July 2025, with 16 experienced developers and 246 tasks in their own projects, measures a 19% slowdown. The same developers expected to save 24% of their time and, once the work was done, still believed they had saved 20% (METR, 2025).

The two results do not necessarily contradict each other. One plausible explanation: on code they know by heart, developers lose time reviewing the tool’s suggestions, whereas on less familiar code they would save time. For a small business with one or two developers, it is better to decide task by task where the tool helps.

What about analysis, presentations and briefing notes?

This is the experiment run in 2023 with 758 Boston Consulting Group consultants, published in 2026 in Organization Science (Dell’Acqua et al., Harvard Business School). On the tasks the tool was suited to, consultants completed more than 12% more tasks, more than 25% faster, with quality rated about 32% higher.

On a task chosen to go beyond what the tool can do, the share of correct answers fell from 84.5% without AI to 70.6% with it, and to 60% among those who had also received training in how to use it. For me, this is the main finding: the tool does not signal when it is outside its range, and it is up to the company to spot those tasks and keep human oversight on them. Hence fifth place, despite a high gain.

What is not ranked, and why?

Predictive maintenance, vision-based quality control, fraud detection, autonomous agents that chain tasks end to end. These use cases can be useful, but to date I have not found a published study with a control group whose result carries over to a company with 20 to 200 employees. The figures in circulation come from vendors or consulting firms, often with no published method. They will be added if such a measurement is published.

The McKinsey survey at least gives a direction: cost reductions are most often reported in supply chain, service operations and manufacturing, and revenue increases in marketing and sales. These are self-reported, with no controlled measurement, but they point to the functions to examine first.

Where should you start?

With rank one if you have a customer service team, rank two if you do not: these are the two cheapest and best-proven entry points. Then do what the studies in this ranking do, at your own scale. Equip half the team for six to eight weeks, keep the other half as a control group, and measure the same indicator on both sides. For example, requests handled per hour, response time, the number of documents produced or the rework rate.

That way you will know whether the tool pays off in your company, not just in a study. We build this kind of test with the CEOs we work with, starting from the decision maturity diagnostic. Definitions of the terms used here are in the glossary of the economics of decision-making.

Frequently asked questions

Peter Guillon, founder of Exelvys. This ranking is revised with each new published measurement.