The promise your chatbot makes is yours
The question a board asks about AI in customer contact is short: are we liable for what the thing says? You are. Air Canada's chatbot told a customer about a bereavement discount that the airline did not offer. The customer acted on it. The tribunal ruled that the company was bound by what its own system had promised.
That ruling is the cleanest answer available to a director who wants to know where the risk sits. This piece is the first of two about trust and governance in customer contact. Here I lay down the case. The second piece covers what you build on top of it.
The question behind the question
What a board usually means is simpler: can we point at the vendor? The answer is no, and the reason is structural. A customer talking to your agent is talking to you. The interface carries your name, your tariffs, your terms. The customer has no way to tell which part of the sentence came from a model and which part came from a policy document, and no reason to care.
Everything your agent says in a conversation reaches the customer as a statement of the company. That is the whole exposure in one line.
An agent only gets authority from an organisation that can be held to what it says.
What the ruling actually settled
The Air Canada case did not create a new rule. It applied an old one to a new speaker. A company is answerable for the commitments made in its own name. Putting a model in the chair does not split that in two.
Three things follow directly:
- A promise made by your agent is a promise made by you, whatever the model was trained on.
- The customer owes you no verification. They read the answer and acted on it.
- The defence that the system misspoke is the admission that you let it speak unsupervised.
Guardrails are the mechanism of delegation
Most boards treat guardrails as the thing that slows the rollout down. They are the thing that makes the rollout possible. You can only hand an agent the authority to confirm a refund, change a booking or quote a price once you can say afterwards what it was allowed to confirm, what it actually did, and who answers for the difference.
Compare it to the authority you already hand out. A service employee can waive a fee up to an amount. The amount is written down, the waiver is logged, and a manager owns the ceiling. Nobody calls that a brake on the employee. It is the reason the employee is allowed to decide at all.
An agent earns the same kind of room on the same terms.
Trust became the commercial asset
The thing a customer is buying from you in an AI conversation is reliability of the answer. Speed is cheap now. Availability is cheap now. What is scarce is the confidence that what you were told at eleven at night is what the company will honour in the morning.
That confidence is a balance-sheet item in everything but the name. It shows up as fewer contacts, fewer escalations, fewer concessions made at the counter to repair something a system said earlier.
The company that can stand behind its agent gets to give that agent more to do.
What you do on Monday
- Write down every commitment your agents are currently able to make in a customer conversation: refunds, discounts, delivery dates, policy exceptions.
- Put one named owner on each of those commitments, at director level, the way you already do for a budget line.
- Set a ceiling per interaction that the agent may commit on its own, and route everything above it to a person.
- Keep a full log of what was promised, so a dispute starts with evidence rather than reconstruction.
- Ask your legal counsel one question this week: if our agent said it, would we honour it?
Part two takes this further into the operating side: decision rights, escalation paths and the owner behind each one. If you want to map that before then, the AI Decision Rights material lays out who may approve what, and the CX+AI Risk & Guardrails material turns the ceilings above into something you can hand to a team. Bekijk de materialen