How do you specify and procure AI solutions?
Specify an AI solution by the impact you want, not the technology. Describe the process, what should get better and how it is measured. Set requirements on data, regulation, human review, operations and cost per case, and on owning your data, your prompts and the right to switch supplier. Start with a scoped pilot.
Off the shelf or built for you?
Mid-sized B2B companies rarely need another licence. They need someone who makes the solution work in their reality and sees it through. That is why the requirements should be about your process and your impact, not the supplier's feature list.
How is AI procurement different from regular IT procurement?
| Regular IT | AI solution | |
|---|---|---|
| Output | The same answer every time | Answers vary and need quality assurance |
| Requirements | Features | Impact, quality and error handling |
| Data | The system stores data | Data drives the result, and where it is processed matters |
| Regulation | GDPR | GDPR and the EU AI Act |
| Operations | Licence and support | Also model cost per case, monitoring and model changes |
What should the requirements include?
- Process and impact. Which task, what should get better, how it is measured.
- Data. Which sources, where they are stored and processed, which permissions.
- Quality and error handling. How good is good enough, where a person reviews, what happens when an answer is wrong.
- Regulation. Classification under the AI Act, transparency towards customers, GDPR and a data processing agreement.
- Operations. Monitoring, logging, cost per month and per case, who updates it.
- Ownership and exit. You own your data, your prompts and your configuration, and can switch supplier or model.
- Adoption. Training and change management are included, not just delivery.
- Pilot first. Scoped, with measurable goals and a clear decision afterwards. See From AI pilot to production.
How does procurement work?
For private companies: start with a problem and a pilot, compare two or three suppliers on a real task and sign the agreement for expansion only when the pilot has proven its impact. Public organisations in Sweden must also follow the Public Procurement Act.
What are the most common mistakes?
- Specifying the technology instead of the impact.
- Forgetting the operating cost per case.
- Locking into a supplier without the right to move data and prompts.
- Procuring a big whole before anything is proven.
- Leaving change management outside the agreement.
We write the requirements and lead the project when others deliver, so you have a buyer who knows the technology. See Services.
FAQ
Common questions about AI requirements
Should we build ourselves or buy?
It depends on how close the solution is to your core process. Test it in a pilot before you decide.
How do you compare suppliers?
On a real task from your own process, with the same input for everyone.
What should an AI contract include beyond a regular IT contract?
Requirements on quality and error handling, regulation, cost per case and ownership of data and prompts.
Are you about to buy an AI solution?
Bring the need to a conversation. We will help you phrase the requirements so you compare suppliers on the right things.
Book 30 minutes