What should a business ask when an AI vendor says its product is safe?
Turn the claim into questions about your own workflow. Ask what information goes in and who can see it, what the AI agent can do, and what record exists when someone reviews a mistake. What the vendor provides and what your organization decides are separate things.
In the summit interview, Moira Forbes asked what safety looks like, and Teresa Carlson described her view of Anthropic's approach. Both make useful prompts for a buyer's questions. Neither replaces checking a product's terms and controls.
A vendor may describe a control, but your team still has to understand whether it's available in the product and plan you'd use, how it's configured and who owns it. Ask for current documentation. Bring in the people responsible for security and privacy, plus procurement, before any sensitive material goes in.
Don't assume a general statement about safety answers a specific question about your workflow.
For your next evaluation, pick one realistic task and walk through it with the vendor. Write the unanswered questions down next to the proposed benefit, and keep them in front of whoever makes the decision.