The number of people calling themselves AI consultants has exploded over the last two years. In Skåne you now find everything from a one-person company with an OpenAI key to large consultancies that have added “AI” to their service catalogue. The difference in what you actually get is enormous.
These are the questions we think you should ask before signing. We are happy to answer them about ourselves too — you should get the same answer either way.
1. Have you built and operated what you are selling?
AI infrastructure is not a strategy document. It is GPU scheduling, KV caches, vector databases, and alerts that go off at three in the morning. Ask about concrete systems the consultant has built and still operates — not just presented.
A good sign: they can talk about what went wrong. A bad sign: every reference project is described as a success with no friction.
2. Where does our data end up?
This is the question that disqualifies the most vendors. If the answer is “we use an API”, ask which one, where it runs, which sub-processors are involved, and what happens to the prompts sent to it.
For many organisations — public sector, life science, finance, law firms — a data processing agreement is not sufficient. The data may not leave the organisation at all. That requires dedicated GPU capacity and self-hosted inference, which is a completely different architecture from calling an external API.
Ask: “Which of our data will leave our environment, and can you build a solution where the answer is none?”
3. What happens when the model gets it wrong?
Every model is wrong sometimes. The relevant question is what happens next: is it detected, is it logged, and can the answer be reconstructed afterwards?
Ask to see how evaluation works. Is there a test set? Do you measure quality continuously or only at handover? Do you know when a model that worked in January stops working in June? An AI service without evaluation is a service nobody owns.
4. Who actually does the work?
Many engagements are won by a senior architect and delivered by someone new to the profession. Ask who writes the code, who sits in the meetings, and how much of that person’s time is allocated to the project.
It is also worth asking how many engagements the consultant is running at once. A senior consultant with eleven parallel projects is, in practice, a project manager.
5. What does it cost after the project ends?
The large costs in AI systems are rarely development hours. They are GPU hours, token costs, and the time it takes to keep the system alive. A solution that is cheap to build can be expensive to operate for three years.
Ask for a monthly cost model at expected volume — not just a project budget. Also ask what happens at double the traffic: which costs scale linearly and which do not.
6. Can we take over the system ourselves?
A vendor that makes itself indispensable is a hidden cost. Ask whether the infrastructure is expressed as code, whether it can run in your own environment, and whether you get documentation that describes the system in plain language.
A good consultant makes themselves redundant. That is a bad short-term business model and a good long-term one — because that is the consultant you call next time.
7. What is a sensible first step?
If the answer is a six-month programme with an unknown endpoint, be wary. A sensible first step is scoped, measurable and cancellable: an architecture review, a technical feasibility study, or a proof of concept with success criteria agreed before work starts.
Ask: “If we only want to spend twenty hours before deciding — what would you do with them?”
Summary
All seven questions come down to one thing: does the vendor want you to understand the system, or only to buy it? An AI consultant who can answer straight on data residency, evaluation, cost after the project, and handover is almost always the one you want to work with.
We are happy to answer all seven — get in touch via the contact page. We work from Staffanstorp with clients in Malmö and Lund, and you can read more about our AI consultancy services.