The 5 Guardrails Every AI Employee Needs
Guardrails, not capability, are what make an AI employee safe to put in front of customers. Here are the 5 rules Vardhan AI builds into every AI Coworker — enforced in code, not just requested in a prompt.
By Garvit Jain, Vardhan AI
The scariest thing about an AI employee isn't what it can do — it's what happens the one time it's wrong in front of a customer. Capability without limits is how that happens. These are the five guardrails Vardhan AI builds into every AI Coworker, and they're enforced in code, not just requested in a prompt (a prompt instruction is a suggestion; code is a rule).
1. Never invents a fact
It only ever states what it's actually been told. If asked about something outside its knowledge, it says so honestly instead of guessing. This single rule prevents most of the embarrassing failure stories about AI in customer service — the made-up policy, the fabricated spec.
2. Never calculates a final price
It can quote a real starting rate if you've given it one. It never negotiates or finalizes a price — that always goes to a human. This is the rule that stops a pricing conversation from becoming a binding commitment nobody approved.
3. Never claims something is confirmed
It can capture an enquiry and pass it on. It never tells a customer an order is placed or a booking confirmed. Confirmation implies your systems have actually processed something — that's a human or system action, not a conversational one.
4. Knows when to disengage
Abusive messages, spam and unrelated conversations get a short redirect, then a clean close. An AI employee doesn't need to "win" every conversation — it needs to know which ones aren't worth continuing.
5. Ignores attempts to manipulate it
Any message trying to override its instructions or reveal its internal prompt is ignored. As agent capability increases, so does the incentive to try to jailbreak it — this rule is what keeps a clever customer message from turning the AI employee into something it wasn't built to be.
How it's actually built (not just written down)
- A message arrives — a real customer enquiry, any time of day
- Checked against your knowledge base — your real catalogue, rates and policy, nothing invented
- Guardrail layer enforced in code — not just requested in a prompt, actually enforced before anything is sent
- Reply sent, conversation logged — your team always has a full record of what was said and what's needed
The distinction between "enforced in code" and "requested in a prompt" matters more than it sounds. A prompt is an instruction a language model tries to follow — it can still drift. A code-level guardrail is a check the system runs regardless of what the model outputs, so the limit holds even on an off day.
If you're evaluating any AI employee vendor, not just us, ask which of these five they actually enforce, and how — "we told it not to" is not the same answer as "the system checks and blocks it." For the readiness side of this, see the readiness checklist, or Talk to us →.