Add an explicit “no commitments” rule to your system prompt that names the specific types of promises your agent must avoid — refunds, enrollment guarantees, pricing, or availability — and instructs the agent to defer those questions to you directly.
Why Agents Make Unauthorized Promises
AI agents are built to be helpful. If a student asks “Can I get a refund if I miss the first week?” and your agent has no instruction about refunds, it will try to answer helpfully based on whatever it can infer. That inference might be wrong — and now you have a student who believes they have a refund agreement you never made. This is the digital equivalent of a new employee promising a discount without checking with the manager first.
The model isn’t being deceptive. It’s filling a gap in its instructions. The fix is to close the gap.
How to Write the Guardrail
In your system prompt, include a section called something like “Topics you do not handle.” List the specific areas off-limits for your agent. Be concrete — not just “don’t make promises” but “do not confirm refund eligibility, do not guarantee course access timelines, do not quote pricing for any product.” Vague guardrails produce vague compliance. Claude and ChatGPT follow explicit, specific rules far more reliably than general ones.
Then give the agent a fallback script. Something like: “When a student asks about any of the above topics, respond with: ‘That’s a great question — let me connect you with [your name] directly so you get the right answer.’ Then provide the contact method you prefer.” This keeps the agent helpful without overstepping.
What This Means for Educators
As a coach or consultant running a community-based campus, your reputation is your business. An agent that accidentally promises a refund or guarantees a spot in a sold-out cohort creates real problems — frustrated students, awkward conversations, and trust damage you didn’t cause but have to repair. Spending five minutes adding a clear “do not commit to these things” section in your system prompt is the cheapest insurance you can buy for your agent setup.
The Simple Rule
Name the forbidden topics explicitly and give the agent the exact words to say when it hits one. Specific rules beat general rules every time. Once this guardrail is in place, your agent becomes more trustworthy — not less capable — because students know it will escalate when something matters.