Our LLM chatbot told a customer they were eligible for a refund that our policy didn't actually…

offer. The customer screenshotted it. Our support team had to honor it.

Cost: $4,700 and a very uncomfortable meeting with the VP of Customer Success.

That was the day I became obsessive about guardrails.

Guardrails that actually work in production:

Output validation against business rules. If the LLM claims something about your policies, validate the claim against your actual policy database before showing it to the user.

Topic boundaries. Define explicitly what the LLM should and shouldn't discuss. A customer service bot has no business giving medical advice, even if asked nicely.

Factual grounding enforcement. Every factual claim must be traceable to a retrieved source. No source? Don't make the claim.

Toxicity and safety filters on both input and output. Standard libraries (Guardrails AI, NeMo Guardrails) handle the basics.

Human escalation triggers. When the LLM encounters a situation it shouldn't handle autonomously (legal questions, complaints, anything involving money), route to a human.

And the simplest guardrail of all: include "I'm not sure" in the model's response options. An LLM that says "I don't know, let me connect you with a human" is infinitely more trustworthy than one that confidently makes things up.

Guardrails aren't limitations. They're what make LLM applications trustworthy enough to deploy.