Evaluating an AI vendor: the audit checklist

Every AI vendor says it is secure and responsible. The useful question is whether it can prove it to someone who is not on its sales team. This checklist is the one we would want a buyer to use on us.

Certifications and their scope

• Provide the full SOC 2 Type II report, not a summary letter. Which systems and which period does it cover?
• Provide the ISO 27001 certificate and its Statement of Applicability. Is the hosting environment in scope?
• When was the last independent penetration test, who performed it, and were the findings remediated?

Data handling

• List every subprocessor that can access our data, with its location.
• Is our data, or anything derived from it, used to train models for other customers?
• What is retained after the contract ends, and how is deletion verified?
• Where do prompts and outputs go during inference, and are they logged?

The model

• Which model, which version, and who provides it?
• How are we notified before the model is changed, and can we hold a version?
• What evaluation was done for our use case, and can we rerun it ourselves?
• How does the system behave when it does not know the answer?

Accountability

• For each output, can we see the inputs, the retrieved sources and the model version that produced it?
• How long are those records kept, and can we export them?
• Who is the named individual responsible for security, and for privacy?
• What is the incident notification commitment, in hours?

Exit

• In what format are our data, configurations and any fine-tuned weights returned?
• Can the system run on our own infrastructure, or another provider’s, without the vendor?
• What does transition assistance cost, and is it in the contract?

Reading the answers

A strong vendor answers most of these with a document reference. A weak one answers with a meeting. Where the answer is "we are working towards it", ask for the date and the auditor, and score it as not yet held.