Criterio

Jurisdiction record: refund agent for a travel company

Legal is drawing the line on the wrong decision. In Moffatt v. Air Canada (2024) the chatbot never issued a refund. It told a customer he could claim a bereavement fare after travelling, which was wrong, and the tribunal held the airline to that statement. Air Canada argued the chatbot was responsible for its own words, and the tribunal rejected that. The harm came from what the agent says about entitlement, not from the money it pays out. Requiring human approval of every refund leaves that risk untouched: an agent that can't pay can still promise. Air Canada's failure was an execution error inside a mandate that was otherwise sound, which was "answer policy questions." It is fixed by grounding the answers, not by removing payment authority.

Status of the mandate. The €200 autonomy is a proposal, not yet a delegation in force. Still missing: how refunds under €200 are distributed by cause, which policies can be decided from booking data alone, and who owns policy interpretation (legal, revenue or support).

Zones

The agent's own decisions. It can answer questions about a specific booking using the fare rules it retrieved for that booking, and cite them. It can issue refunds up to €200 when three things hold: entitlement follows from a written rule plus booking data (refundable fare class, cancellation inside the window, the company itself cancelled), the amount is calculated by the system rather than the agent, and nothing is negotiated. A wrong refund here is a bounded cost that shows up in the ledger. That is reversible enough to be covered by the next zone's review.

Edge cases. These are still executed, but they trigger an alert and a review:

• refunds near the cap;
• several refunds against the same booking or customer within a short period, which may be splitting a claim to stay under the cap;
• a customer whose history shows earlier claims.

Behind this zone sits an aggregate daily cap that acts as a circuit breaker. A daily human review of a sample, plus every flagged case, is early enough because each loss is capped and can be traced. The aggregate cap protects against the scenario where review comes too late: mass manipulation of the agent through social engineering or prompt injection.

Decisions outside its mandate. For these the agent stops and hands the case to a person:

• Any denial of a statutory right. This covers EU261 compensation, Package Travel Directive cancellations and "extraordinary circumstances" judgments. A wrong denial is the least visible error in the system: the customer leaves, no ledger entry is created and nobody reviews it. A wrong payment can be audited; a wrong "no" cannot.
• Any promise of a future or conditional refund ("you can claim afterwards").
• Medical, bereavement or other compassionate exceptions.
• Goodwill gestures, unless a defined budget and criteria exist.
• Cases where the customer says another agent told them something different.
• Chargebacks and disputes.

Within this zone the agent may explain the process. It may not anticipate the outcome.

Escalation

• Recipient: the support refund queue, which is still to be designated.
• Backup: the shift lead, also to be designated.
• Policy interpretations: whoever owns the policy, which is pending.
• While waiting: the agent tells the customer the case has been referred and gives a timeframe. It makes no commitment on the outcome and records what was said. Silence never turns into a refund or a denial.

Metric versus purpose

Support will probably measure containment and CSAT. Neither prevents an agent from improving its numbers by saying "yes" or promising refunds, which is exactly how Air Canada failed. If refund cost is measured instead, the incentive flips toward wrong denials. This is a plausible incentive, not observed behaviour. The fix is a sampled audit against policy that counts wrongful payments and wrongful denials, and that weighs as much as containment does.

Evidence that would move the line

Legal's position works well as a time-limited phase with exit criteria. Run the agent in recommend-only mode for a few weeks and compare its decisions with the humans' (approvals, denials and amounts). The disagreement rate per refund category tells you which categories move into the agent's own zone and whether €200 is the right cap.

Revisit the line if any of these change

• regulation (EU261 reform, AI liability rules);
• the observed error rate;
• the volume of fraud attempts;
• an expansion of the agent's channels.

This record proposes how the work is split operationally. It does not certify legal compliance.

Referral. If the fare rules were written for human agents who interpreted them case by case, the agent will execute the wording without that judgment. Whether those rules still hold up belongs to /sedimento.

Open question: Which refunds under €200 can today be decided from booking data and written rules alone, with no interpretation, and who gets the case when they can't?