# Can an AI agent offer a discount?

**It can propose one safely long before it can agree one.** Those are two
different permissions, and most companies control the second — who signs —
while leaving the first wide open, which is backwards: by the time a signature
is needed, the shape of the deal has already been set in a sentence somebody
sent on Tuesday.

A buyer who has been told *we can probably do fifteen percent* has been given
a number. Whether anyone had authority to agree it changes what happens next,
but it does not change what the buyer now expects.

## Proposing and agreeing, split on purpose

| | proposing | agreeing |
|---|---|---|
| **the sentence** | *there is usually room at this volume* | *fifteen percent, confirmed* |
| **what it needs** | a range somebody owns, and the words that keep it a range | the authority to bind the company |
| **what it costs when wrong** | an expectation to walk back | a price you have to honour |
| **who holds it** | can be the agent, within a ceiling | a person, every time |

**An agent can hold the first and never the second.** It describes what is
usually possible, says what it depends on, and hands over — and the handover
is recorded as something open, so the thing it half-promised is on a list
somebody is looking at rather than in a thread nobody reopens.

## The part nobody budgets for

**A discount does not come back.** Most of what an agent spends replenishes:
a prospect's patience recovers, an enrichment credit resets at midnight,
approval capacity clears as the queue clears. A discount precedent does the
opposite — it holds, or it gets worse, and it keeps doing that after you have
stopped giving them. The next buyer in that segment has heard what the last
one got.

That is why the honest control here is a reported trend rather than another
ceiling. The rule about what may be offered governs the deal; watching where
the average has drifted tells you whether the rule is still being followed —
and putting a cap in both places would be the same control disagreeing with
itself.

## What to actually set up

1. **Write the range down with an owner.** Not *be sensible* — the actual
   figures, with the name of whoever is accountable for them being current.
2. **Let the agent propose inside it**, in words that keep it a proposal:
   what it depends on, and who confirms.
3. **Keep agreeing with a person**, and make the handover visible rather than
   a dead end.
4. **Read the drift**, by segment. A discount policy that nobody has looked
   at in two quarters is a policy that has already changed.

## What this does not do

It does not stop a human from agreeing something unwise, and it does not
price your deals. It records what was offered, by whom, and on whose
authority, and refuses to let an agent state a figure that no current source
supports.

</what-should-an-ai-agent-be-allowed-to-promise.md> is the full list of eight
permissions this is one of. </agent-promise-register.md> is where the range
lives.