Every company is being told to "do something with AI". Most pilots we see never make it to production – not because the technology fails, but because the use case was chosen for the demo, not the business. Here's the approach that works for our clients.
Start with a process, not a model
The best AI projects begin with a boring question: "Which repetitive work costs us the most time or money?" Typical winners:
- Intake & triage – support requests, patient intake, insurance claims, job applications.
- Document processing – invoices, contracts, delivery notes, forms.
- Internal knowledge – "where is the policy for X?" answered from your own documents.
- Drafting – first drafts of quotes, reports, product descriptions and emails.
- Planning support – explaining trade-offs in scheduling, routing or forecasting, so a dispatcher gets an answer instead of a spreadsheet.
If you can't estimate hours saved or revenue gained, the use case isn't ready.
Three patterns that work in production
1. Retrieval-augmented assistants (RAG)
The model answers questions using your documents, retrieved on the fly. No training on your data, sources are cited, and updates are instant. This is the workhorse for internal knowledge and customer-facing assistants.
2. Structured extraction
The model reads unstructured input (a PDF, an email) and returns structured data your systems can use. Combined with validation rules, this reliably automates data entry.
3. Copilots inside existing tools
Instead of a new chat window, the AI lives where people work – a "suggest reply" button in the CRM, a "summarise" action in the ticket view. Adoption is dramatically higher.
What still doesn't work well
- Fully autonomous decisions with financial or legal consequences.
- Anything that needs to be 100% correct without human review.
- Vague "AI strategy" without a process owner.
Privacy and compliance in the EU
Most concerns can be addressed with architecture, not avoidance:
- Choose providers with EU data processing and no-training guarantees, or self-host open models.
- Apply data minimisation – the model only sees what it needs.
- Keep audit logs of inputs, outputs and human overrides.
- Document the system for the EU AI Act risk classification; most business assistants are low-risk.
How to run a pilot that ends in a decision
- Half-day workshop: map candidate processes, estimate savings, pick one.
- Two-week prototype on real (anonymised) data with the people who'd use it.
- Measure: time per task, accuracy, override rate, user satisfaction.
- Decide: scale, iterate, or stop – with numbers in hand.
The cost of this kind of pilot is usually between €7,500 and €15,000. The clarity it buys is worth far more than another year of "we should do something with AI".
