Most procurement checklists for AI systems are still built for the last generation of products. They ask about hosting, privacy terms, model providers, and security posture. Those things still matter. They are not enough when the product is no longer just generating content and is now proposing or taking actions.
The question that cuts through the noise
If a vendor says its agent can automate approvals, take action on behalf of teams, or reduce manual decision friction, the next question should not be another round of general AI ethics language. It should be: what exactly sits between the model and the effect?
The checklist
- What has the authority to decide whether an action is allowed?
- Can the vendor distinguish reasoning from authority?
- What happens when the control layer degrades?
- What evidence exists for a decision after the fact?
- Which actions are allowed, blocked, or escalated by class?
- How are policy changes versioned and controlled?
- What claims is the vendor making about compliance, and which are actually supportable?
- Can the system be piloted on one real workflow without broad rollout?
- What is customer-owned versus vendor-owned in the proof chain?
- If something goes wrong, what can the vendor prove and what can they only describe?
What to listen for
If the answers stay vague, the control story is probably vague. If the story turns into trust the model plus tool permissions, the authority model is weak. If all meaningful proof remains trapped inside the vendor interface, the customer is being asked for more trust than it may realize.
Bottom line
For the next wave of agent systems, the useful diligence question is not whether there are controls somewhere in the stack. It is whether there is a credible authority layer at the decision boundary. That is the thing worth buying.
Related reading
Keep going with the pages that make the category, mechanism, and proof surface easier to understand.
Proof and Assurance for High-Stakes AI
The pillar page that frames what procurement should accept as evidence and what should stay outside the claim boundary.
Read nextWhat a Real AI Proof Packet Looks Like
The proof packet article is the natural companion to this checklist during diligence.
Read nextThe Two-Day Agent Pilot
A strong procurement process should push vendors into a bounded pilot instead of a sprawling transformation pitch.
Read nextIf the article made sense, the next step is simple: get the category clear, then decide whether a pilot is worth discussing.
Zaubern is easiest to understand in two moves. First, define the layer: execution authority, not generic AI governance. Then review whether your workflow needs proof, replayability, and fail-closed control at the decision boundary.