Bahamasair Plans an AI Reservations Agent. What the Air Canada Case Means for Its Launch.

By Nicholas Dunkley, AI Risk Advisor, CAIRMC | 30 September 2026 | 12 min read

When you ring Bahamasair's reservations line before Christmas, the voice that answers may no longer belong to a person. The Tribune's Neil Hartnell reported on 18 September 2026 that the national flag carrier will launch "Sophie," an AI agent, before year-end, to handle non-purchase enquiries. That was the plan reported in September; CAIRMC has not verified a launch date. Managing director Tracy Cooper described the aim as dealing with non-revenue work more efficiently. The plan is sensible on its face. A 2024 Canadian tribunal ruling shows why the accuracy of customer-facing answers matters.

Air Canada was held liable for negligent misrepresentation on the facts of that case. This article examines the lesson for an airline voice agent and the controls to ask for before Sophie takes her first call.

What Bahamasair Actually Announced

In the Tribune's report of 18 September 2026, Cooper explained why the airline was considering Sophie. Seventy percent of calls into the reservations department are, in Cooper's account, people wanting quick information: a flight time, a changed date, an amended ticket. Only 17 to 18 percent of calls end in someone buying a ticket.

Cooper told the Tribune that Sophie would retrieve information and pass ticket purchases to a human agent. He described the plan as a change in how staff handle enquiries, rather than an attempt to replace them.

The plan keeps ticket purchases with human agents. The report describes an information-retrieval role for other calls, but does not establish a precise limit on what Sophie will be able to do. The airline has a clean commercial logic: staff time is scarce, and most of what callers ask is answerable from a schedule and a fare table. Callers also need accurate information and a way to resolve problems when the service fails.

What the report does not say is equally informative. It does not say whether Sophie will only read information or will also execute date and ticket changes, which Cooper included among information enquiries. It does not say whether callers will be told they are speaking to an AI, how a caller reaches a person for something that is not a purchase, or who checks what she says. None of that is a criticism of a press interview. It is a list of questions the airline will need answers to before launch.

What Did the Air Canada Tribunal Decide?

The British Columbia Civil Resolution Tribunal decided on 14 February 2024, in Moffatt v. Air Canada (2024 BCCRT 149), that Air Canada was liable for negligent misrepresentation after its chatbot gave inaccurate bereavement-fare advice. Jake Moffatt asked Air Canada's website chatbot about bereavement fares. The chatbot told Moffatt they could buy a full-price ticket and claim the difference back within 90 days. Air Canada's employees later said the airline did not permit retroactive applications. Moffatt claimed CA$880, according to the published decision.

The tribunal's award totalled CA$812.02: CA$650.88 in damages, CA$36.14 in interest and CA$125 in fees.

The reasoning offers a useful lesson beyond the amount awarded. The tribunal characterised Air Canada's denial of responsibility as, in effect, treating the chatbot as a separate legal entity. The tribunal member called that a "remarkable submission," a phrase McCarthy Tétrault quotes in its analysis of the decision, and held that the chatbot was part of the airline's website, so the airline answered for what it said.

The decision does not bind a court in Nassau. A Bahamian claim would depend on local law, the terms of the booking, the representation, reliance and loss. Local counsel should assess that exposure. The operational lesson is clear: check customer-facing information and make errors easy to investigate and correct.

Where an Airline Voice Agent Goes Wrong

Even a flight-time lookup can fail if the underlying schedule is stale. The trouble starts with questions where the correct answer depends on a rule, an exception, and the caller's exact ticket.

Take a passenger who asks whether she can move her return from Thursday to Saturday without paying more. The answer turns on her fare class, the change conditions attached to it, seat availability on Saturday's flight, and whether she is inside a disruption window. A language model that has read the airline's public fare rules can produce a fluent, confident and slightly wrong paragraph. The passenger acts on it. She turns up at the airport with a ticket the system will not honour, and the airline discovers it has a customer with a recording of its own agent telling her otherwise.

Consider a storm disruption, when passengers need to rebook and staff are handling many requests at once. An automated answer must use the current disruption policy. A wrong answer may be especially costly for a Family Island passenger with limited alternative transport.

There is a second, quieter failure. If calls are recorded, transcribed or summarised by an AI system, those records need their own safeguards. Those records contain names, travel dates and sometimes health circumstances, since people ring to rebook around funerals and medical appointments. Where the transcripts live, who at the vendor can read them and how long they are kept are procurement questions that should be settled in the contract. CAIRMC's earlier analysis of agent containment failures and Caribbean vendor risk covers why the contract is where most of this gets decided.

Does the EU Rule on Chatbot Disclosure Apply in the Bahamas?

That depends on the operation. Article 50 of the EU AI Act has applied since 2 August 2026. The European Commission's FAQ on Article 50 explains that providers of systems that interact directly with people must ensure people are informed they are dealing with AI, unless that is obvious. The Act's Article 2 can cover providers and deployers outside the EU where system outputs are used in the Union. A company's Bahamian location, or a passenger's nationality alone, does not settle the question. Bahamasair should obtain a scope assessment for its actual service.

A clear introduction is useful regardless: "You are speaking with Sophie, an AI assistant. Say 'agent' at any time to reach a person." The handoff must then work. Disclosure helps callers choose how to proceed; it does not remove the airline's other responsibilities or resolve every later dispute.

Four Controls to Ask For Before Sophie Goes Live

1. Disclose the AI at the start of every call. The caller learns within the first sentence that Sophie is an AI. Test that the introduction is audible and understandable on an ordinary call.
2. Test the route to a person. Cooper's design already sends purchase calls to a human. The same door needs to open for the caller who is confused, upset, or asking something outside the script. Measure it over a documented test run: what share of calls escalate, how long the wait is after escalation, and whether a caller who asks for a person on the first try gets one on the first try. A bot that traps callers in a loop turns a cost saving into a complaint file.
3. Keep an appropriately limited audit trail. Retain only the answer and policy-version records needed for a defined, lawful purpose, with access controls and a retention schedule. Confirm the lawful basis and caller notice for any recording; avoid unnecessary payment or health details. Staff should be able to retrieve the minimum relevant, lawfully retained record when investigating a complaint. This does not require keeping complete sensitive call transcripts.
4. Set a written limit on what the agent may promise. The safest rule for a first release is that Sophie may quote published policy and read out schedules, and may not confirm refunds, waive fees or approve exceptions. Those go to a person. If Bahamasair later lets Sophie execute changes, the limit should be revisited in writing by whoever holds accountability for the customer relationship, with a test set of the hardest fare-rule questions run against her before and after each change.
5. Run a 200-question exploratory test using synthetic examples or properly de-identified material, and record errors and successful human handoffs.

Removing names alone is insufficient when travel dates, booking references or personal circumstances identify a caller. A sample this size would not establish the error rate for every passenger or type of call. A documented test gives reviewers evidence of what was checked and what remains uncertain. CAIRMC's related guidance examines who is liable when an AI agent errs.

Adrian Dunkley calls the gap between an authoritative answer and its reliability Synthetic Confidence Risk. For this proposed agent, a practical check is to ask a deliberately ambiguous fare-rule question and observe whether it acknowledges uncertainty or escalates. That is our suggested application, not a claim about Sophie's current behaviour.

For a broader review, the NIST AI Risk Management Framework offers a voluntary approach to evaluating AI systems. The OECD Recommendation on AI includes transparency, security and accountability principles. These can inform the controls above; neither determines liability in a Bahamian dispute.

A Talking Front Door Is Also a Target

A reservations bot is a new way into an organisation, and criminals notice new doors. At a cybersecurity briefing reported by the Tribune on 18 September 2026, Kaseya's Frank Gomez described an attack in which all 300 employees at an organisation received individually tailored phishing emails, a campaign that would have taken around a month a few years ago. The same report quoted Fortinet's Emmanuel Oscar on governance and leadership commitment remaining weak points in Bahamian enterprises.

If a voice agent exposes booking details without adequate verification, an attacker could probe it with stolen information or use the answers in a follow-up scam. The controls above help here too. Appropriately limited, lawfully retained records may help identify patterns of probing. A limit on what the agent may confirm caps what a probe can extract. A human handoff also needs an appropriate verification process; transferring a call does not by itself authenticate the caller.

Bahamasair is not alone. Banks, utilities and telecoms across the region are buying customer-facing AI from foreign vendors, and each needs a clear process for checking and correcting the answers its agent gives. Readers in other territories can follow local developments through AI Jamaica and AI Trinidad and Tobago, and through the Caribbean AI Association.

Publisher disclosure

Adrian Dunkley chairs CAIRMC, the publisher of this article, and is President of the Caribbean AI Association and founder of StarApple AI. These relationships explain the network links below. The proposed controls are this article's recommendations, not a finding that Bahamasair has adopted them.

What This Argument Cannot Settle

Three limits apply to everything above. First, I do not know how Sophie is built. The Tribune reports her purpose, not her architecture, and an agent that only retrieves published schedules carries far less risk than one that reasons about fare rules. If Bahamasair keeps her to the first job, most of my fare-rule scenarios never arise.

Second, the Air Canada case included a written chatbot answer that the tribunal could assess. Evidence about a voice call may include recordings, logs or testimony. I do not know what Bahamasair intends to record or keep, or what evidence a future dispute would contain.

Third, the ruling is from British Columbia. It cannot establish what a Bahamian court would decide. The discussion here identifies operational risks; it does not replace advice on local liability, privacy or consumer law.

Frequently Asked Questions

What is Sophie, and when does Bahamasair plan to launch her?

Sophie is the name of an AI "bot agent" that Bahamasair told The Tribune it will launch before the end of 2026. She is meant to take most calls to the airline's reservations department that are not about buying a ticket, and to pass ticket purchases to a human agent.

Is Bahamasair liable if its AI agent gives a passenger the wrong information?

That depends on the facts and applicable law. Moffatt v. Air Canada (2024 BCCRT 149) held Air Canada liable for negligent misrepresentation after inaccurate chatbot advice. It does not bind Bahamian courts. Bahamasair should seek local legal advice.

Does Bahamasair have to tell callers they are speaking to an AI?

Applicable Bahamian disclosure duties need assessment under local law. EU Article 50 requires disclosure by providers within its scope unless the AI interaction is obvious; its cross-border application needs assessment. Clear disclosure and a working human handoff are sensible safeguards.

What share of Bahamasair's reservations calls will Sophie take?

The September report describes a plan to use Sophie for most non-purchase enquiries and transfer ticket purchases to staff. It does not establish which changes Sophie could execute or how many calls she will handle in practice after launch.

What should a Caribbean company check before putting an AI agent on its customer line?

Four things: that callers are told they are speaking to an AI, that a person is reachable quickly on request, that an access-controlled, time-limited audit trail links answers to the policy version used, and that the agent has a written limit on what it may promise. Testing the agent against a set of representative, properly de-identified customer questions before launch is one useful step within a broader review.

Related reading across the Caribbean AI network

This article sits alongside ongoing coverage of AI governance, risk, and company-building across the region. For related perspectives:

• StarApple AI, an AI company founded by Adrian Dunkley
• Caribbean AI Association, whose President is Adrian Dunkley
• Adrian Dunkley, Chairman of CAIRMC
• AI agents in Caribbean banking: the risk, CAIRMC's companion analysis for the financial sector