12. What should we do when an assistant gets our pricing, integrations or security claims wrong?

Treat factual accuracy as its own workstream, with accountable review and correction.

Put to: Lily Ray, with the buyer’s product-marketing and security owners.

Why it matters to buyers: NIST’s Generative AI Profile identifies confabulation as a risk requiring context-sensitive management. Its framework is a risk-management resource, not a guarantee that any particular correction method will change an assistant’s response. (nvlpubs.nist.gov)

Ask: “How would you investigate an incorrect product claim, prioritise its risk and verify improvement without promising control over the model?”

A good answer should cover:

• The exact statement, answer date, prompt, surface and saved evidence.
• Dated reference material and the internal owner authorised to confirm it.
• Supported, contradicted, outdated and unresolved classifications.
• Whether accessible cited material contains the same error; an inaccessible source remains unverified.
• Correction routes for owned pages, independent publishers and platform feedback.
• Repeated checks for recurrence, with escalation for high-consequence claims.

Press for: an accuracy register rather than a sentiment score. Ask how the expert handles a statement that flatters the company but is false.

For regulated or security-sensitive claims, require the appropriate specialist’s approval before publishing corrective material.

Expert quote — pending: ____________________