What should an AI agent policy for a sales team include?

Last updated 17 October 2026

What each agent may promise, what it relies on and who owns that, what it may spend of a buyer's attention, when it must ask a person, and what happens when nobody answers. Then who reads the policy itself, and how often.

Most policies written for agents are a list of things not to say. That covers manners. It does not cover what the agent commits your company to, which is where the cost is.

The sections, in plain words

1. Who granted the authority, to which agent, until when. A policy with no end date is one nobody re-reads.
2. What it may promise. Eight separate decisions rather than one setting from cautious to bold. Three of them are never yours to open: suggesting something the agent never said, giving a person's word, and promising a date nothing is in place to keep.
3. What it relies on, and who owns each piece. Every price, feature and certification it may repeat, with a named owner and how often it is checked. A rule that depends on information nobody owns is decorative.
4. What it may spend. A buyer's attention, your sending reputation, your approvers' time. Show the totals first; cap only what you choose to.
5. When it asks first. In terms of the deal and the line, not the agent: a number outside the range, a line that is overdue, a variation on a term.
6. What silence means. An ask nobody answers is not a yes.
7. People and accounts no agent touches, whatever else is true.
8. What happens when the rules run out: what stops, what carries on, and who is told.
9. Who watches the policy itself, for being too loose and too tight, and on what day it is next read.

The test of a good one

Every rule says why it is there in a sentence somebody who sells would understand. A limit nobody can explain gets switched off. And every rule names the information it depends on and who keeps that current, because an agent following a rule perfectly against wrong information is still wrong.

What to leave out

Anything that is a judgement about a person. A policy that becomes a way to grade the rep behind an agent stops being followed honestly, and the signals that tell you an agent is drifting disappear with it.

What this does not do

Reynforce does not write your policy. It holds it in a form an agent's checks can read, and records which version every promise was made under.

/what-should-an-ai-agent-be-allowed-to-promise.md goes through the eight decisions. /get-an-ai-agent-approved-by-legal.md covers what a reviewer asks of the policy.