Why your SME already uses AI… and still sees no productivity gain
Surveys say AI has entered SMEs. Founders say they see nothing on the P&L. Both are right. Here is why, and how to leave permanent pilot mode.
· 7 minute read
Short answer: an SME can « use AI » without changing a single process, without reliable data, and without a success metric. Then the tool adds work. It replaces nothing. Public figures agree: adoption is fast, productivity gains remain limited.
According to the France Num 2026 Barometer (presented in September 2026), about 39% of micro and small businesses report using AI tools for work, up from 5% in 2023. Banque de France (September 2026 bulletin) finds that in early 2026 more than two thirds of firms with at least 20 employees report using generative AI, mostly in a limited or experimental way, with still weak productivity gains. Two measures, two populations, one field conclusion: access to a tool is not the same as producing more.
1. Definition: what counts as an AI productivity gain?
An AI-related productivity gain is not « the team finds it handy ». It is a measurable cut in time or cost for the same outcome, or more output with the same resources, on a named process.
Without a named process and a before/after measure, you do not have an AI project. You have a pile of personal trials.
2. Cause 1: usage is real, but it is individual (shadow AI)
Someone uses ChatGPT or Copilot on a personal account for an email or a summary. Nobody decided to remove a step from the process. Time saved stays local, invisible, and non-transferable.
Result: the company « uses AI » in surveys, and the weekly schedule does not change. Until usage is tied to a shared task (quotes, follow-ups, reporting, support), there is nothing to consolidate on the P&L.
3. Cause 2: the process did not move
Adding an assistant next to Excel + email + oral validation, without removing a step, raises cognitive load. You get more drafts. You do not decide faster.
The useful question is not « which model? ». It is « which manual step do we stop if AI owns that part? ». Without an answer, the tool stays an accessory.
4. Cause 3: input data is messy or missing
A model speeds up what you feed it. If it gets three definitions of revenue, five versions of a process, or a Drive with no source of truth, it speeds up confusion.
Same finding as in a data audit: until the foundation holds, a conversational layer creates false confidence, not productivity.
5. Cause 4: no KPI was set before the trial
« We will save time » is not a metric. Useful KPIs: minutes to draft a standard quote, share of follow-ups sent within 48 hours, level-1 support response time, hours per week spent consolidating a report.
Without a baseline, you can never say whether the tool helped. After three months the debate turns subjective again.
6. Cause 5: a generic tool was bought for a business bottleneck
Banque de France notes diffusion mainly through general-purpose apps, often in support functions. Fine to start. Not enough if the real bottleneck is operational: stock, scheduling, invoicing, document compliance.
A generic tool without connectors or business rules remains a writing accelerator. It is not a production system.
7. Five-point grid: from trial to measurable gain
1. Name one task (not « AI in general »).
2. Measure current time or volume for one week.
3. Decide which manual step disappears if AI owns its share.
4. Check that input data has one definition and one owner.
5. Set a 30-day success threshold (e.g. -30% time on the task). If not met, stop or change scope.
Five points. Not a transformation programme. Minimum discipline to stop confusing adoption with results.
What this changes in practice
The barometers are right: AI has entered SMEs. Founders are right: many see nothing on the P&L. The gap is rarely the model. It is process, data, and measurement.
As long as usage stays individual, with no step removed and no metric, you accumulate trials. You do not accumulate productivity.
Already have tools, but no visible gain?
Thirty minutes is enough to pick one task, set a baseline, and see whether the blocker is process, data, or tool.