Consulting
From the first use case to running operation.
The part few people offer, because it is work: pick a case that proves something, examine the tool, read the contract, involve the works council, put it into operation – and come back a few weeks later to see whether it is still being used.
Step one
Choosing the right case
The first use case should not be the most important one but the most instructive. Four criteria lead there, and they are unspectacular: the process occurs often. The result can be checked in minutes. There is a recognisable „correct“. And an error is annoying, not dangerous.
What regularly works
- Summarising recurring documents – minutes, incoming mail, reports
- Drafting standard letters, so the blank page disappears
- Evaluating free-text fields nobody has ever read
- Grounded search across your own records, with the source under every answer
What fails as a starting point
- Anything touching personnel decisions – also because of the noise it creates
- Projects without a benchmark, where nobody can say what "right" would be
- The grand design whose benefit only shows after months
- Projects that first require a data foundation to be built
The last point deserves a note, because it comes up often. If your records are not searchable, you do not have an AI project — you have a digitisation project. That is not bad news, only a different task, with its own budget and a specialist provider to carry it out.
Step two
Five questions before you sign
I hold nothing with blanket decisions for or against a vendor. Decisions are made per use case, along five questions – and the answers are in the contract, not on the product page.
- Which data goes in – and do you really need all of it?
- Does it leave the European legal area?
- Is the provider training on your inputs?
- Can usage be logged in a way you could explain later?
- Is there a route to the same result with less personal data?
On top of that, any case involving personal data needs a data processing agreement under Article 28 GDPR. Without it, business use is regularly unlawful — no matter how good the tool is.
Step three
Introduce, measure, decide
Before the start we build a benchmark: twenty real past cases where the correct outcome is known. It costs half a day and is the best thing you can invest in a project like this. Without it, mood decides afterwards – and mood decides wrongly in both directions.
What we measure
Duration of the process including rework, compared with before. Share of results usable without change. Frequency and weight of corrections.
What we do not measure
User counts, request counts, satisfaction scores. They rise in every pilot and say nothing about benefit.
The most honest figure
Whether the tool is still in use eight weeks after the pilot. That question is rarely asked, because the answer can be uncomfortable.
Boundaries
What I do not do
I build no models and train none. I write no bespoke applications and take over no system operation. Legal advice belongs to a lawyer: I can tell you which obligation applies to you and how others meet it, not whether your particular case would hold up in court.
I also provide no scanning or digitisation services. Where opening up analogue records becomes a precondition, I support requirements, selection and steering of a specialist provider. For completeness: I advise a provider in that market and disclose it here.
And I run no workshops whose output is a wall of sticky notes. There are enough of those.
Frequent questions
About the consulting.
Where should a company start with AI?
How long does an introduction take?
Do you build the solution as well?
Do you recommend a particular vendor?
First conversation
A conversation that is worth something on its own.
Thirty minutes in which we go through your situation and I tell you what I would tackle first in your position — even if no engagement comes of it.