Tech

AI in business after the hype: the three things that actually work

After the demo phase, three uses of AI genuinely hold up in a company: reading messy documents, classifying and routing incoming requests, and preparing drafts a person reviews. All three share one thing — the AI proposes, the person decides — and they must be measured like any other investment.

14 May 2026 · 9 min read · updated on 27 Aug 2026

Every company we know has followed the same path over the last two years: initial enthusiasm, a few trials that impressed everyone in a meeting, then a return to normality in which nobody uses anything any more.

It is not a failure of the technology: it is the difference between a demo and a workflow. Demos always work, because whoever runs them picks the example. Daily work does not: it arrives awkward, in an unexpected format, at six on a Friday. Here is what survives once the euphoria passes.

1. Reading documents no system could read

The most solid use and the least talked about. Supplier invoices each in their own format, crooked scans of delivery notes, orders arriving as text inside an email, fifty-page PDF specifications where you need one table.

Before, you needed hand-written rules for every format — and they broke when the supplier changed their template. Today content can be extracted even from a document never seen before, at a quality that survives real work. It is boring to show in a meeting and it is the thing that frees the most hours.

The rule we apply: extraction proposes, validation stays automatic on the numbers. If the invoice total doesn't match the sum of the lines, it doesn't pass. A person doesn't need to check everything: the system needs to know when it isn't sure.

2. Classifying and routing what comes in

Hundreds of emails, forms and messages someone reads to work out what they are about and who to pass them to. It is reading work, not deciding work, and it is exactly where a language model performs well: grasping topic, urgency, tone.

Here too the boundary is clear: route yes, reply no — or rather, reply only after a person has looked. A wrong classification costs one extra hop; a wrong automatic reply costs a customer.

3. Preparing drafts someone reviews

Summaries of a long case, first drafts of a proposal, answers to recurring questions, translations. The value is not that the AI writes better than you: it is that you no longer start from a blank page. The person corrects in five minutes what they would have written in thirty.

Two conditions make it work: the model must have your real documents in front of it (a generic proposal helps nobody), and it must be crystal clear to the reviewer that they are reviewing. That is the accountability you cannot delegate: the signature stays with whoever sends.

AI proposes and a person decides isn't a technical limitation: it is the model that makes automation defensible.The rule we keep on every project

What did not hold up

The chatbot answering customers with no safety net. Fine for frequent questions; on requests that matter it produces plausible wrong answers, which customers take as commitments.
Automatic operational decisions — what to order, what discount to give — with no human in between. The model doesn't know what it doesn't know, and won't tell you.
The generic assistant for the whole company, with no specific task. After two weeks three people use it.
Uploading everything to an external service without asking which data is leaving. It is the shortcut that later becomes a notification to the authority.

The minimum rules to set before starting

Which data may leave and which may not. Decided once, written down, known by everyone. Customer records, health data, documents under confidentiality: explicit answers are needed, not habits.
Where AI may act and where it may only propose. The line is the irreversible: if the action sends something out or deletes, it needs a traced human yes.
Who answers for what goes out. Accountability stays with whoever signs. No supplier takes it from you, and anyone claiming otherwise hasn't read the contract they are handing you.
How it is measured. Hours saved, errors avoided, response times: if after three months you can't say whether it helped, it isn't a project, it is a subscription.

Where to start, in practice

Pick one process, the one where someone reads and re-keys every day. Measure what it costs today. Build the assistance on that, with human approval where it matters. Measure again after a month.

If the number improves, you have a real case to extend and a company that has learned to use it. If it doesn't, you spent little to find out it wasn't needed there — which is still a result, and far more useful than a platform bought so as not to fall behind.

The right question is not "how do we use AI?" but "where does someone read and re-key?". The first leads to buying a tool; the second to solving a problem.

Does our data end up training the model?

It depends on the contract, and it must be read. Business services generally exclude using data for training, but the clause has to be verified and put in writing — along with where processing happens and how long it stays in the supplier's logs.

Our own model or an external service?

For almost every small company an external service with a serious contract is the rational choice: quality is higher and cost a fraction. Your own model makes sense when data cannot leave for regulatory reasons, or when volumes flip the maths.

What does trying cost?

Far less than people think: a single case is built in a few weeks and consumption is measured in tens of euros a month while it stays on one process. The real cost is your own people's time to define what is right and what isn't.

Will AI replace staff?

In the cases we have seen, no: it removes the reading and re-keying. But it is a legitimate question and it should be addressed before, not after: telling people what they will do with the hours that come back is the condition for the tool being used rather than worked around.