The AI diagnostic in micro-businesses: what it involves and what it finds

A dark screen showing an AI assistant's welcome prompt
In short: an AI diagnostic in a company of fewer than ten people takes two or three days and answers one question: where would automation actually save hours here? Roughly half the time, the honest answer is that the data has to be cleaned up before anything can be automated at all.

An AI diagnostic is a review of how a business actually runs, carried out to find where artificial intelligence would return real hours. In a company of fewer than ten people the exercise is short, and the conclusions are blunter than in a larger organisation, because there is nowhere for a bad decision to hide. You either bought the right tool or you are paying a subscription for something nobody opens.

Why very small companies bother

The usual assumption is that AI is a large-company subject. The economics point the other way. A firm with fifty people has someone whose job is administration. A firm with four does not, so the administration lands on the owner in the evening. An hour saved there is an hour of the most expensive person in the business.

There is also a regulatory push. In France, electronic invoicing becomes mandatory to receive from 1 September 2026 and mandatory to issue for small companies from 1 September 2027. Businesses that were going to modernise their bookkeeping eventually now have a date. A diagnostic run at the same time avoids doing the work twice, once for the tax obligation and once for the productivity gain.

How the diagnostic runs

A code editor showing an AI actions menu with options such as Explain Code and Find Problems

It starts by watching the work rather than discussing it. What software is actually in use, as opposed to what was purchased. How data gets keyed in, and how many times the same figure is retyped. Where the day gets stuck. This part is uncomfortable for the same reason it is useful: people describe the process they designed, and the process that runs is usually a different one.

The second stage puts numbers on the potential gain, in hours per week and in errors avoided. Estimates get stress-tested against how the work varies, because a task that takes ten minutes on a normal Tuesday and ninety minutes at month end is not the same task.

The third stage is a roadmap ordered by immediate effect on the bottom line against difficulty of implementation. Small companies should do the cheap boring thing first. It funds the interesting thing, and it builds the confidence that makes the second project easier to sell internally.

Data protection and where files live

A monitor displaying an AI assistant's welcome screen

Any technology project handling personal data sits under GDPR, and a diagnostic covers it as a matter of course rather than as a separate exercise. Three questions carry most of the weight. Where is the data hosted? Who has access, including former staff and outside providers? What does the algorithm do with it once it has it? France's data protection authority publishes practical guidance on all three.

This is not only a legal exercise. In small businesses the answer to "who still has access" is frequently a name nobody had thought about in two years, and finding that out is worth the afternoon on its own.

The objections that come up

The first is fear of the tools themselves, usually expressed as a belief that AI needs a technical department to run it. That was true a few years ago. It is not now, and the more common failure today is buying something too sophisticated rather than too simple.

The second objection is more serious and comes up in most diagnostics. If what goes into your systems is wrong, automation reproduces the errors faster and at greater volume. Duplicate customer records, product references that never matched between two tools, an address field used for three different things. Cleaning that up is unglamorous and it is the actual first project. Any consultant who skips it is selling you a subscription.

Funding and outside help

French companies have routes to co-funding. The training component can usually be covered by your OPCO, the sector body that collects and redistributes employer training contributions, and the level of support is generally more generous for companies under eleven employees. Bpifrance co-funds a dedicated data and AI diagnostic, though its size thresholds put it out of reach for the smallest firms. None of it is automatic, and the criteria change, so the sensible first call is to your usual adviser.

The case for an outside pair of eyes is simply that the owner cannot see their own process any more. Someone who has watched thirty other companies do the same thing spots in an hour what took you a year to stop noticing. That is the whole value, and it is worth being clear-eyed that it is a limited one: the diagnostic produces a decision document, not a working system.

Want to know where automation would actually save you hours?
The initial conversation is free and without obligation.

Discover AI diagnostics & audit

FAQ

What is an AI diagnostic?

A review of how your business actually runs, aimed at finding where artificial intelligence would save real hours and cut data entry errors, and where it would not.

Is it only worthwhile for large companies?

No. Very small companies often gain more, because they have no back office to absorb administrative work and every hour saved goes straight back to the owner.

Can the diagnostic be funded?

In France, sector training funds known as OPCOs cover the training component, and Bpifrance co-funds a data and AI diagnostic for companies above a certain size. Eligibility depends on your sector and headcount, so check with your usual adviser.

What are the GDPR obligations on an AI project?

You need to know where the data is hosted, who can reach it, and on what legal basis it is processed. Where personal data is involved at scale, a data protection impact assessment is expected.

How long does a diagnostic take?

For a company of fewer than ten people, two to three days spread over a fortnight is typical. Less than that and the findings come from what you were told rather than what happens.