# Who is liable when an AI agent makes a promise?

**Your company is.** An agent acting in your name is not a third party, and
"the model said it" has not worked anywhere it has been tried — so the
question worth your time is not *who is liable* but *can you tell what was
promised, by which agent, and on what authority*.

That second question is the one most companies cannot answer today, and it is
the one that decides whether a disagreement with a customer costs an afternoon
or a settlement.

## What a term in your contract can and cannot do

**It can allocate risk between you and your customer.** It cannot make the
promise disappear. Three things in particular tend to disappoint the people
relying on them:

- **A disclaimer the customer never read.** If the commitment arrived in a
  conversation and the disclaimer lives in terms behind a link, you are
  arguing about which one the customer was entitled to rely on. That argument
  is not one you win cheaply.
- **"The agent was not authorised."** This is only an argument if the
  authority was written down somewhere before the promise, and if what the
  agent actually had access to matched it. Otherwise it is an assertion.
- **Consumer protection does not care about your architecture.** Where it
  applies, a statement made to a consumer in your name is a statement you
  made.

**None of this is legal advice and we are not qualified to give any.** What it
is: the three places we have watched people assume they were covered, and the
thing software can actually change.

## What software can change

Not liability. **Evidence, and the authority behind it.**

| | |
|---|---|
| **what was promised** | the sentence a commitment was made in, kept so the claim is checkable rather than reconstructed from a thread |
| **by whom** | which agent, and which person's authority it was acting under |
| **on what basis** | which claims it relied on, and the state each was in **at the time it was used** — not the state it is in now |
| **under what authority** | the mandate as it stood that day, stored rather than derived, because the mandate has since changed |
| **whether it is closed** | an open promise does not expire, it rots, and nothing about it fails at a moment somebody notices |

That record is append-only. Nobody can edit it, including us — which is the
whole reason it is worth anything in a disagreement.

## The one thing to do this week

Write down the claims your agents are allowed to repeat, and put a name and a
date against each one. You do not need us for that, and
</agent-promise-register.md> describes the shape. What you will find is the
useful part: on a first run through about thirty claims, the person filling it
in reported nothing a customer would trouble them about — and a second reader
found four in ten minutes, including a borrowed certification and an undated
roadmap commitment.

## What is not on this page yet

**[unread] The decision people usually cite here.** A tribunal in Canada held
an airline to a promise its support chatbot made, and it is the case every
discussion of this reaches for. **It is not quoted or cited here, because
nothing in this session has been able to read it** — the primary decision and
the reputable reports of it are both unreachable from where this page was
written, and a citation from memory is exactly the kind of confident wrong
sentence this product exists to argue against.

`liability docs` is the reader pointed at the primary source. When a run
returns a passage, the quotation and the citation go here with the date it was
read. Until then this page says what we can stand behind, which is everything
above it.