Your chatbot's answer is your company's promise

GembaOS is a zero-trust governance runtime for enterprise AI agents: policy-decided access, single-use clearance tokens bound to exact parameters, and accountable human approvals, running on-premises or in a private VPC.

This article examines the simplest way an AI agent can inflict severe financial and legal damage on an enterprise without touching a single production database: by stating an unauthorized commitment that the law subsequently forces the company to honor.

Incident breakdown: The legal fallout of Moffatt v. Air Canada

In November 2022, a passenger visited Air Canada’s website to book flights following the passing of a family member. Seeking clarification on bereavement fare policies, the passenger consulted the airline’s official AI-powered customer service chatbot.

The chatbot provided clear, plausible, and reassuring advice: the passenger could purchase standard full-price tickets immediately and submit an application for a retroactive bereavement refund within 90 days. Relying on this explicit instruction, the passenger purchased the tickets.

When the passenger subsequently submitted the refund claim, Air Canada rejected it. The airline’s actual, static policy stated that bereavement rates could never be claimed retroactively once travel had occurred. The passenger filed suit with the British Columbia Civil Resolution Tribunal.

Air Canada’s legal defense was striking: the airline argued that the chatbot was “a separate legal entity that is responsible for its own actions” and that the customer should have cross-referenced the static policy page elsewhere on the site.

The tribunal rejected this argument in its entirety:

1. No separate legal existence: An enterprise chatbot is not an independent contractor or third party. It is a direct digital representative of the brand.
2. Duty of care: The enterprise owes consumers a duty of care regarding all representations made across its official touchpoints.
3. Negligent misrepresentation: Holding the airline liable, the tribunal ordered Air Canada to pay damages and costs.

While the financial damages in this initial case were modest, the precedent was seismic: what an AI agent tells a customer constitutes a binding representation by the enterprise.

The blast radius: From chat conversation to legal commitment

Rebooking after flight disruption, fee waivers, compensation following delivery delays, goodwill vouchers, and discount matches—these represent the highest-volume inquiries in enterprise customer support.

Because modern LLMs are trained to be helpful, agreeable, and persuasive, they naturally generate fluent, empathetic, and contextually plausible promises. When an unconstrained conversational model is deployed to front-line customer service, a dangerous illusion is created:

Three architectural rules before an agent speaks to customers

To protect against misrepresentation liability, enterprises must implement three structural controls:

1. Official commitments are actions, not prose

A refund, compensation voucher, or policy exception must never be promised within the conversational chat thread. The conversational agent has zero transactional authority. It can draft a structured proposal, but that proposal remains inactive until a designated human supervisor cryptographically signs the authorization.

2. Verified policy text over generated summaries

Where customer questions touch upon legal terms, cancellation rights, or eligibility criteria, the agent must not improvise explanatory prose. Official determinations must cite canonical policy blocks directly approved by compliance and legal teams.

3. Structural separation of duties

The system that talks to the customer must not be the system that executes financial payouts. If a customer is eligible for a refund, the front-line agent issues an intake ticket. The governance runtime routes that ticket through the approval chain, validates real-time account balances, and executes the refund via single-use clearance tokens only after human sign-off.

Summary

The lesson of Moffatt v. Air Canada is clear: disclaimers and terms-of-service footers cannot protect a company when its AI agent makes an explicit representation to a customer.

An AI agent’s explanation must never be allowed to turn into an accidental promise. By stripping conversational models of execution authority and enforcing human-signed clearance tokens for every financial commitment, GembaOS ensures that companies remain in complete control of their promises and their liabilities.

Explore live scenarios

See how GembaOS enforces governed financial actions:

• Customer refunds scenario: Reviewing refund thresholds and human cryptographic sign-off.
• Flight rebooking scenario: Handling travel exceptions and fare class governance under regulatory oversight.