Guardrails3 min read
Why our sales agent cannot invent a price
Telling a model “never make up prices” is a wish. Here is the rule we wrote in code instead.

Mitra is the AI salesperson inside Resandhi. It answers shop customers on WhatsApp, in Hindi, English or Hinglish, and sells from the shop’s own catalogue. The one thing it must never do is quote a price or a stock level that is not real.
Prompts are not controls
You can tell a model not to invent numbers. Most of the time it listens. Most of the time is not good enough when a customer pays the price it quoted.
Any price or stock number in a reply must have come back from a tool in the same turn. If it did not, the reply is blocked.
What happens to a reply
- Your systemsCustomer asks about a product
- AI agentMitra calls the catalogue tool
- AI agentDraft reply checked for numbers
- Your teamUnbacked number? One retry, then a person
- Your systemsReply sent
What stays with the owner
- Refunds and discounts wait for the owner
- An address or amount changes only when the customer confirms
- The team can take over any chat, at any time
The same pattern goes into every agent we build for clients: decide which mistakes are unacceptable, then make them impossible in code rather than unlikely in a prompt.
