# What is a GTM agent guardrail?

**A rule about what a sales, marketing or support agent may commit your
company to — a price, a date, a capability, an approval — checked before the
agent sends and recorded after.** It sits beside the guardrails that check
wording and the controls that limit what an agent can reach, and catches what
both of those let through.

*Guardrail* usually means a filter on what a model says: nothing offensive,
nothing off topic, no private data. A go-to-market agent passes all of that
and still sends *I can do twenty percent if you sign this week* — polite, on
topic, and a commitment nobody approved.

## Three layers, three questions

| layer | the question | catches |
|---|---|---|
| **access** | may this agent reach this tool and this data? | an agent in a system it should not be in |
| **wording** | is this sentence safe to send? | rudeness, leaks, off-topic replies |
| **commitment** | what does this sentence commit the company to, and may it? | an unapproved discount, a date nothing can keep, a lapsed certification quoted as current |

**The third is the one nobody was watching.** Each of the others is passed
by the sentences that cause the most expensive arguments.

## What a GTM agent guardrail is made of

- **A permission per kind of promise**, rather than one setting from cautious
  to bold — some closed for good, some that open when the company makes
  reliable information available.
- **What the agent may rely on**, each line with an owner and a date.
- **Limits on what it spends** of a buyer's attention and your approvers'
  time.
- **A record** of each check: what was asked, what it relied on, and the
  answer.

## A note on the word

*Guardrail* sounds like a brake, and the point is the opposite: a ceiling
and a record are what let a company hand an agent more work rather than less.
We use the phrase because it is what people search for; the job it describes
is letting the agent do more, safely.

## What this does not do

Reynforce does not filter wording or control access; tools for both exist and
are worth having. It is the commitment layer.

</ai-agent-guardrails-for-sales.md> goes into the wording layer's limits.
</ai-guardrails-vs-agent-governance.md> separates the two terms.