Build an RFP and vendor scorecard for an AI delivery partner

Score each answer from 0 to 3, treat the requirements your organisation cannot waive as pass or fail, and agree the weightings before any bid arrives. Give most weight to whatever you can verify without the supplier's help, such as contract wording as drafted, certificates and reference calls.

Two things are unusually revealing in agentic work: the stopping condition each bidder will commit to, meaning the finding that would make them tell you not to proceed, and how they would sequence a second workflow once the first resets the estimate.

The ones your organisation cannot waive. The pack suggests screening before access, data residency, approval of AI tools, breach notification, ownership of the work, the terms for ending the engagement, the liability cap and insurance, plus the sector question for financial services, the public sector and industrial work.

Everything else is scored from 0 to 3. The split is a suggestion, so move a requirement either way to match your own supplier standard.

A 0 is no answer, or an answer that avoids the question. A 1 is a partial answer with no evidence. A 2 is a clear answer that meets the need, with evidence still to check. A 3 is a clear answer that meets the need and is backed by evidence you have seen.

The pack also says what a 3 looks like for each requirement, so a panel can see what it is scoring against.

Start with who is screened and to what standard, where your data is processed, which AI tools and model providers touch your content and who approves them, how quickly a personal data breach is notified and in which document that is written, and which certifications are held rather than in progress.

Then add the sub-processors with their locations and transfer mechanisms, how each agent's access is limited and revoked, and how AI-generated code is checked before it ships.

Ask whether the supplier adds margin to model, platform or licence costs, and whose accounts those costs sit on. Then ask for the run costs to expect, line by line: model inference, the platform, storage and search, evaluation and monitoring, and human review time.

A supplier's fee is only part of what the system costs. It also costs money to run, and your own team's review time is part of that cost.

Yes. The questions are written to be useful whoever you hire, and none depends on a Tenhaw answer. Tenhaw's published answers sit in a separate final section of the Word document, so you can delete it before sending the questions to bidders.

If you keep it, put the same questions to every firm on your list, Tenhaw included.

Held means Tenhaw's practice or position today, quoted from the linked page. Contract term on request means a clause in the Terms of Service or a term of the Data Processing Agreement, whose signed versions are sent on request and take precedence over the published Terms where they differ. In progress means assurance work under way and not yet certified, and only the certifications row carries it.

Where the site publishes no answer, the pack says to ask on a call and adds nothing.