AI integration and consulting

What this is about

A language model is a tool, not a member of staff. It is strong at ordering text, translating, summarising and making structure out of the unstructured. It is weak at knowing things that are written down nowhere — and it does not tell you so.

The work therefore lies less in connecting than in drawing lines: where the model decides, where it only suggests, and where a person signs off.

Where it usually carries

• Sorting and routing what comes in — enquiries, receipts, forms.
• Turning free text into fields another system can process.
• Preparing drafts that a person checks and answers for.
• Searching your own holdings, with the source given.

Where we advise against it

• Anywhere a wrong answer is expensive and nobody notices it.
• As the sole basis for statements with legal effect.
• Where a simple rule does the same — a rule can be read back, a model cannot.

What we put in writing

• Disclosure. Where users are talking to a system or seeing machine-generated content, it says so there — Art. 50 of the AI Act requires it, and it is the least one can do anyway.
• Where the data goes. Which provider, which country, what retention, and whether inputs are used for training. That belongs before the decision, not in the footnote.
• What happens on failure. A service that does not answer must not bring the process to a halt.
• How it is checked. What would tell you the results are getting worse — before a customer notices.

Consulting without installation

Often the most useful answer is that installing it is not worth it. We look at the process, say where a model takes something off your hands, where it only moves work around — and roughly what it means in operation. Even when no commission follows from that.

Scope, process and price are stated in the quote — once it is clear what the job is. All amounts net, plus VAT.