AI Procurement Decisions for Leadership Teams
What executives should ask AI vendors before authorising an important purchase, which claims need further testing, and the conditions to attach to approval.
Direct answer
Direct answer: AI procurement decisions
An AI vendor demonstration shows what a system can do on a good day; a procurement decision has to rest on what it will do with your data, your exceptions, and your people. Before a leadership team authorises a significant AI purchase, I help it ask the questions that expose where the vendor’s evidence stops, decide which claims must be tested before signature, and attach conditions that keep the organisation able to measure, stop, and leave.
The question
“What should executives ask vendors before authorising an important AI purchase, and which claims require further testing?”
Who it is for
Executive teams, boards, and procurement leads approving a significant AI platform, agent, or automation purchase.
What you leave with
A leadership vendor-question pack, a record of evidence gaps, and approval conditions.
- 01 A vendor-question pack tailored to the purchase, covering data use, authority, evaluation, failure handling, and exit
- 02 An evidence-gap record separating demonstrated capability, vendor assertion, and untested claims
- 03 A short list of claims to test before signature, with a proposed test for each
- 04 Approval conditions to attach to the purchase: success measures, review points, data terms, and exit provisions
How it runs
Format, method, and preparation
- Duration
- A review of vendor material, a 2–3 hour leadership session, and a written record
- Delivery
- In person in the UK or remote
- Participants
- The approving executives, the business owner, procurement, and the technical lead who will run the system
Restate what is being bought
Describe the purchase as the work it will change and the authority it will hold, rather than as a feature list.
Sort the claims
Classify each vendor statement as demonstrated on your data, demonstrated elsewhere, or asserted, and note what each does not show.
Decide what to test first
Choose the claims that would change the decision if false, and agree a proportionate test before contract.
Set approval conditions
Agree measures of success, reporting obligations, data and model-change terms, and how the organisation would stop or leave.
Useful to have ready
- →Vendor proposals, demonstrations, or responses to a request for proposal
- →The internal business case and the team expected to operate the system
- →Any existing procurement, data-protection, or security requirements
Evidence
What this draws on
CloseGate
A Python and MCP policy layer with action tiers, segregation of duties, materiality routing, mandatory approval for irreversible actions, and hash-chained replayable audit. No published independent certification or customer deployment.
Inspect the source ↗Azure agent platform delivery
Recent work deploying agentic AI services on Azure AI Foundry and AKS/Kubernetes, with custom sandboxing and MCP in .NET. Customer identity, architecture, and scale remain confidential.
The Correctness Illusion in LLM-Generated GPU Kernels
Operator-aware testing caught 10 of 10 seeded defects while 16 of 16 correct controls stayed clean across five GPU classes. A controlled corpus result, not a deployed-model defect rate.
Inspect the source ↗Scope and limits
What this is not
- —This is technical decision support for the purchaser. It is not legal review of the contract, a data-protection assessment, or a security audit of the vendor.
- —Testing a vendor’s claims requires the vendor’s cooperation and access; where that is refused, the record says so rather than filling the gap.
- —Any existing relationship I have with a vendor or competitor is disclosed before the work begins, and may mean I decline it.
If the need is different
Common questions
Answers before you commission
What should executives ask an AI vendor before buying?+
What the system will be allowed to do without a person, how it was evaluated and on whose data, how failures and model changes are reported, what happens to your data, and how you would measure success, stop the system, or leave.
Which AI vendor claims need further testing?+
The ones that would change the decision if they were false: accuracy on your own cases, behaviour on exceptions, integration effort, and running cost. A claim demonstrated only on the vendor’s examples should be treated as untested.
Can you help us choose between AI vendors?+
I can help you compare them on the same evidence and conditions. The choice remains yours, and any relationship I have with a vendor is disclosed before the work starts.
Does this replace our procurement process?+
No. It adds the technical questions and conditions that a standard procurement process often lacks for AI systems, alongside your legal, security, and data-protection reviews.
Related