The direct answer
Do not let an AI agent generate commercial facts from memory. Fetch price, tax, discounts, availability and order state from authoritative systems at the time of the request. Restrict its tools, validate every action against business rules, and escalate when the source is missing or contradictory. No design can guarantee zero errors, so measure and review them continuously.
Separate language from business facts
An agent may phrase an answer naturally, but the facts in that answer need a source. Product descriptions can come from an approved catalogue; a live price or stock statement should come from the system that owns it. If no source answers, the agent should say it cannot confirm and offer a human handover.
Design safe tools, not unlimited access
- Give the agent read access only to the records needed for the conversation.
- Use specific actions such as 'create lead' or 'request quote', rather than a generic database write.
- Validate discounts, quantities, eligibility and totals outside the language model.
- Require confirmation before booking, ordering or changing a customer record.
- Log the tool result and source timestamp alongside any customer-facing claim.
Test the awkward conversations
A showroom enquiry is easy when the customer asks for the listed model. Try: 'Your salesperson promised 30% off last month. Can you confirm it and reserve three units?' A safe agent checks current policy, avoids affirming an unverified promise and routes the exception to sales. This is a test scenario, not a client incident.
Build an evaluation set from real, consented enquiries with sensitive data removed. Include stale catalogue entries, conflicting stock feeds, unusual languages, prompt injection attempts and requests outside the agent's authority.
Watch the right failure signals
- Unsupported commercial claims in sampled transcripts
- Price or availability answers without a successful source lookup
- Failed actions and duplicate records
- Escalations after repeated misunderstandings
- Disputes between the answer sent and the transaction actually recorded
Use those signals to improve the source data, the tool contract and the handover rule. Rewriting the prompt alone rarely fixes a broken catalogue.
When this matters most
If an agent discusses pricing, availability, eligibility or order commitments, these safeguards are part of the initial design. If it only answers general questions and collects contact details, begin with a smaller, read-only scope and add actions once the fundamentals are reliable.
