How Sanma keeps your AI agent on topic and your costs predictable
Why a business AI agent needs guardrails, and the layers Sanma uses to stop misuse, runaway conversations, and surprise bills.
By the Sanma team ·
Put an AI chat on a public website and, sooner or later, someone will ask it to write their homework or a computer program. Every one of those requests costs you money and does nothing for your business. Sanma is built so your agent stays focused on your customers. Here is how.
Layer 1: misuse is declined before any AI cost
Requests that are clearly unrelated (writing code, essays, homework, translations, or attempts to change the agent's instructions) are recognised and politely declined before the AI model is ever called. That reply costs nothing. If someone keeps trying, the chat is ended.
Layer 2: the agent knows its job
Every agent is told to help only with your business: services, prices, hours, bookings, orders, and policies. Subtle off-topic questions get a one-sentence decline and an offer to help with something relevant.
Layer 3: replies are kept short
Each reply has a length limit, messages from visitors are capped in size, and the number of new chats one visitor can open in a day is limited. Long, expensive answers simply can't happen by accident.
Layer 4: long conversations go to a person
A genuine customer who needs a lot of help is better served by your team. Very long conversations are handed over with the full context, rather than being cut off or running forever.
Your own trade is never blocked
If your business is, say, a translation agency or a coding school, you can allow that topic for your own agent, so customers asking about your actual services are always answered.
Accuracy guardrails too
The agent answers only from knowledge you have approved, prices orders only from your catalog, reads totals back before confirming, and says "I'm not sure" rather than guessing. Want to see it in action? Ask our assistant to write you a poem.
