AI guardrails: stop the bot promising what you don't have

An AI assistant invents details not because it’s «broken» but because it was handed a question without data and never told that «I don’t know» is an acceptable answer. Guardrails are the explicit list of prohibitions and stop-words that turns a guess into a handover.

One invented promise costs more than a dozen missed messages: the customer arrives with a screenshot, and there is nothing left to argue about.

Which promises should AI never make?

Four categories where an apology doesn’t fix the damage:

1. 1

   Timelines

   «It'll be there Friday» without checking stock and courier. The assistant quotes only dates that exist in data.

2. 2

   Prices and discounts

   Any figure not in the catalogue — including «we'll do it cheaper» in response to haggling.

3. 3

   Specifications

   Materials, compatibility, warranty, origin. What isn't in the description doesn't exist for the assistant.

4. 4

   Availability

   «Yes, we have it» without real inventory access. The most common and most expensive invention.

How do you phrase a guardrail so it actually holds?

A prohibition works when it can be checked unambiguously. «Be careful with prices» is not an instruction. «Quote only prices from the catalogue price field; if the field is missing, say you’ll check and hand the conversation to a manager» is.

The test: could an outsider read the rule and say yes or no about a specific reply? If not, rewrite it.

What are stop-words and why do you need them?

Words after which the assistant doesn’t answer at all — it calls a human. Not «answers more carefully». Goes quiet.

A baseline set for a shop: complaint, faulty, defective, refund, scam, lawyer, court, police, plus anything involving health, children and safety. Every niche adds its own: cosmetics — allergy; electronics — smoking, sparked; food — poisoning.

The logic is simple. In these conversations the cost of a mistake isn’t a lost sale but a public dispute or a legal problem. Automation here doesn’t save money, it manufactures risk.

Can AI message first?

Only inside the limits the platform itself sets, and this isn’t a matter of taste. On Messenger a business replies within 24 hours of the customer’s message; beyond that window there is a separate extension of up to 7 days, and it is permitted only for messages written by a human agent — automated messages are explicitly not allowed there (Meta, Messenger Platform policy, 2026).

So «nudging» a customer with a bot on day four isn’t aggressive marketing, it’s a platform-policy breach that puts your page at risk. Follow-up outside the window is written by a person.

Minimum guardrail checklist before launch

1. Data source: the assistant answers only from the catalogue; outside it — «I’ll check».
2. Forbidden promises: timelines, prices, discounts, specs, availability are never named without data.
3. Stop-words: the list that sends a conversation to a human with no attempt to answer.
4. Attempt limit: three repeated questions in one thread — handover.
5. Tone: no apologising for being automated, no manufactured empathy, no «I’ll personally make sure».
6. Outside Messenger’s 24-hour window, no automated messages are sent.