Artificial intelligence in a small business: where it helps, and where it costs a lot for nothing

AI is not a starting point, it is one building block among others. In a small business, most of the gain still comes from simple rules. This page describes where AI really adds something, and where it adds cost and unpredictability.

The distinction that saves money

A rule always gives the same result for the same input. A language model produces a plausible answer, not a guaranteed one. Anything that can be described by a rule should be handled by a rule: it is cheaper, faster and verifiable. AI starts where rules fail: free text, unstructured documents, judgement calls.

Use cases that hold up

• Sorting and routing incoming messages. A quote request, a complaint and an invoice do not follow the same path. A model can sort free text.
• Extracting information from a document. Pulling an amount, a date and a number out of a PDF that changes shape with every supplier.
• Summarising a long exchange so a decision-maker can decide without reading forty messages.
• Producing a first draft of a reply, report or record that a human corrects. The gain is the blank page avoided, not the review removed.
• Translating and adapting content that has already been validated.

Where AI disappoints

1. Calculations and numeric decisions. A language model is not a reliable calculator. These stay with code.
2. Anything that commits the company without review. A quote, a contract, a legal answer: human validation is not optional.
3. Unstable processes. If the rules change every month, neither AI nor automation will hold. Stabilise first.
4. Replacing a knowledge base. A model does not know your prices or procedures. You must provide the data, which means it must exist and be up to date.

What it costs

For a small business, model calls amount to tens of euros a month for well-targeted use, and climb fast when whole documents are sent with every request without filtering. Two habits cut the bill: send the model only what it needs, and handle by rule everything that can be before the call.

The hidden cost lies elsewhere: poorly framed AI produces outputs that must be checked one by one. Checking time cancels the time saved. A good integration is judged by the share of outputs accepted without correction.

Confidentiality and personal data

Three questions before sending anything to a model: what data leaves, where is it processed, and is it reused for training. The business offers of the main providers contractually exclude reuse of data, which consumer versions do not always guarantee.

For sensitive data, two strategies exist: anonymise before sending, or run a model locally. The second costs hardware and skills, and is only justified when the nature of the data requires it.

Where to start without getting it wrong

Pick a point in the flow where information arrives as free text, measure how much time someone spends on it, then test on thirty real cases before industrialising. If the result is right on twenty-seven, the system is usable with a review. If it is right on fifteen, the problem is usually the data provided, not the model.

Frequently asked questions

Does a small business really need artificial intelligence?

Not first. Most of the gain comes from simple rules that move information between tools. AI becomes useful for what a rule cannot describe: free text, unstructured documents, sorting that needs judgement.

What are the most profitable AI use cases in a company?

Sorting and routing incoming messages, extracting information from documents in varying formats, summarising long exchanges and producing a first draft corrected by a human. All four deal with unstructured text.

ChatGPT or Claude for a business?

Both suit common uses. What decides lies elsewhere: the business plan chosen, data reuse commitments, integration with existing tools and the cost at real volume.

Is data sent to an AI reused?

The business offers of the main providers contractually exclude reuse of submitted data for training. Consumer versions do not always. It is the first point to check before connecting a model to customer data.

Should you run a local AI to protect your data?

Only if the nature of the data requires it. A local model avoids sending anything outside but needs hardware, skills and maintenance. Anonymising before sending is enough in many cases.