AI Procurement Framework: How to Evaluate an AI Development Partner in 2026

Frequently Asked Questions

01.How long does it take to evaluate an AI development partner?

Plan for three stages: a shortlist from documents, a paid scoping sprint with two finalists, and contract redlines. The sprint is usually the longest stage, and we suggest two weeks of partner time on the same de-identified sample. Sign the acceptance memo before any of it starts, because writing the threshold after you have seen results lets it drift toward the favourite.

02.Do I need a technical team to run an AI procurement?

You need one engineer you trust, in house or independent, to run each finalist's evaluation code on your holdout set and to read the model card. Everything else in the framework is a business decision: the metric, the threshold, the cost of an error, the weights and the contract terms. A finance lead and the business owner should sign the acceptance memo, not the engineer.

03.What should I own at the end of a custom AI build?

Ask for assignment of the trained weights, adapters and fine-tuning code, the evaluation harness, the holdout results, the model card and a reproducibility pack that lets another team retrain the model. Third-party foundation models stay licensed, not owned, so the contract should list them by name with their licence terms and say what happens if a licence changes.

04.Is buying custom AI development regulated in the EU?

The purchase is not regulated as such, but the role is. If you have an AI system developed and put it into service under your own name, you are its provider. For high-risk systems, from 2 December 2027 for Annex III uses, you and the partner will need a written agreement on information, capabilities, technical access and assistance, and breaching it will be fineable (Regulation (EU) 2024/1689, Articles 25(4) and 99(4)(da), as amended by Regulation (EU) 2026/1744).