← All articles
Trust & accuracy12 August 20266 min read

Stopping an AI assistant from inventing your policies

The single biggest risk in customer-facing AI is not rudeness or bad grammar — it is confident inaccuracy. A made-up cancellation window or an imagined discount lands you in a dispute with a customer who has a screenshot.

Ground every answer in a knowledge base you control

Your assistant should answer from structured business knowledge: services and prices, opening hours per branch, delivery terms, promotions with end dates, and written policies. If it is not in there, it is not an answer.

Importing your existing website is a fast starting point, but review what was imported. Old pages carry old prices.

Make 'I don't know' a designed response

A well-written fallback keeps the conversation moving: acknowledge the question, say the detail isn't confirmed, offer to get a person to confirm, and capture the contact detail needed to do that. Customers accept uncertainty. They do not accept being misled.

Keep policy text in one place and dated

Cancellation, refund and deposit terms should live in the workspace as the single source of truth, and be reviewed when they change. Anything quoted to a customer should be traceable back to that record.

  • One owner responsible for policy content
  • A review date on pricing and promotions
  • Escalation to a human for any dispute about terms

Takeaway

Accuracy is a configuration problem, not a personality problem. Fill the knowledge base, date the content, and make honest fallbacks the default.

Keep reading