Make the proposal easier to judge
An agentic AI supplier should help you deliver useful work, connect it to your systems and keep responsibility clear. A persuasive demonstration starts the conversation. Before committing to a pilot, ask for evidence that relates to the workflow your team actually needs.
Use these questions with shortlisted suppliers, including AI Venture X. Record what is demonstrated, what is promised and what remains unknown. Agree the evidence you need before giving the pilot access to business systems.
1. Which job will improve, and why does it need an agent?
Ask for a defined workflow, accountable owner and expected outcome. Compare the proposed approach with a simpler automation or process improvement. The UK government's AI Playbook recommends choosing the approach around the problem and its constraints.
2. What information and integrations are required?
Request a map of data sources, tool access and integration dependencies. Identify who provides access, what information can leave the organisation and which system remains authoritative. Resolve data quality, ownership and support responsibilities with your commercial and technical teams. AI Playbook: buying AI.
3. What may the agent do without human approval?
Ask for an action-and-permission map, including the downstream systems. A named owner should understand its effective access. Ask which consequential actions need approval and how that rule is enforced. Microsoft's least-privilege guidance explains why tool-level boundaries and traceable authority matter.
4. What does the demonstration actually prove?
Request representative examples, difficult inputs and the failures found. Label synthetic demonstrations, limited trials and production evidence separately. Agree which assumptions still need testing in your organisation; a reference implementation cannot establish your likely return on investment.
5. Will the whole process improve?
Compare accepted output, elapsed delivery time, staff effort and review workload against the existing process. Include rework. Our AI pilot evaluation guide explains how to make that comparison before deciding whether to scale.
Illustrative example: a supplier generates a draft reporting pack in five minutes. Your team still needs to reconcile its sources and obtain approval. Measure that complete journey before treating the five-minute figure as a delivery benefit.
6. Can access be stopped and actions reconstructed?
Ask for a demonstration of revocation across connected systems and records showing what happened. Agree how uncertain or partially completed actions will be investigated before retrying them. Microsoft's guidance covers revocation and logging; implementing and maintaining these controls also adds operational effort. Least privilege for AI agents.
7. Who pays for and operates the service after the pilot?
Request assumptions for integration, usage, support and maintenance. Name the operating owner and agree responsibilities when something changes. The AI Playbook's procurement guidance includes ongoing costs and service dependencies. Buying AI.
8. What can you retain or move if the arrangement ends?
Clarify data access, intellectual property, export formats and handover obligations. Ask about switching costs and dependence on the supplier. Resolve these through the agreement before rollout. AI Playbook: buying AI.
Make the answers usable
Keep one evidence record for each proposal
Name the workflow, required output, decision owner, supplier, proposal version and comparison process. For each question, record:
Evidence and date
What was supplied, and when? Can the team inspect it?
What it establishes
Label the answer demonstrated, stated or unknown. Record its limits.
Responsibility and next action
Who resolves the gap, and what evidence will they provide?
When it must be resolved
Is this a condition before starting, or a question the pilot will test?
An unknown is a useful finding
“Access revocation has not been demonstrated” gives the team a specific action. It does not establish that the control exists or that the supplier is unsuitable.
Turn the answers into a pilot decision
For each question, record the evidence received, its limits, the person responsible and the next action. Separate conditions that must be satisfied before starting from questions the pilot is designed to answer. Choose a bounded pilot, resolve the gaps first or pursue an alternative.
The checklist is AI Venture X's proposed practical synthesis, not a certification or universal supplier score. Assurance effort should be proportionate to the workflow and its consequences. Your organisation retains the buying decision.
If you want to explore a specific workflow with AI Venture X, bring an example input, the required output and the person who owns the decision. See our agentic AI consulting service.
Sources and scope
Primary guidance behind the questions
- UK Government, AI Playbook. Published 10 February 2025. Problem selection, buying AI and lifecycle guidance inform questions 1, 2, 7 and 8. Its context is public-sector delivery.
- Microsoft, Least privilege for AI agents. Effective permissions, ownership, revocation and logging inform questions 3 and 6. Referencing its guidance does not claim an AI Venture X Microsoft-specific implementation or certification.
Sources checked 1 October 2026. The checklist, evidence record and reporting example are practical editorial guidance. They report no client result, measured saving or validated supplier score. The reporting example is illustrative.