The owner of an eleven-person services firm has just been forwarded a screenshot. Her support chatbot, the one she switched on three months ago to handle out-of-hours queries, has told a customer that the firm offers a thirty-day no-questions refund on the work the customer has already approved and signed off. The firm does not offer that refund. The chatbot has invented it. The customer is now polite, firm, and ready to escalate. The owner has three open questions: does she have to honour what the bot said, who is actually accountable here, and what does she do on Monday morning so this does not happen again.
The legal answer to the first two questions arrived in February 2024, when the British Columbia Civil Resolution Tribunal ordered Air Canada to honour a refund its chatbot had promised to a grieving customer. The tribunal called the airline’s argument that the chatbot was a separate legal entity “remarkable” and noted that it “should be obvious” the company was responsible for everything on its platform. The principle has since been picked up by the UK Competition and Markets Authority, the Financial Conduct Authority, the Information Commissioner’s Office, and the Financial Ombudsman Service. The owner’s chatbot is the firm’s agent. The firm is on the hook.
What is customer-facing AI accountability?
Customer-facing AI accountability is the principle that the business deploying an AI system is legally responsible for what that system says and commits to, with the same force as if a human employee had said it. The duty arises from the common law of agency, the reasonable-care obligation in the Consumer Rights Act 2015, the Misrepresentation Act 1967, and the CMA’s 2026 guidance on consumer-facing AI agents.
The accountability is not contingent on whether the AI made an obvious error, whether the commitment contradicts the firm’s actual policy, or whether the AI was supplied by a third-party vendor. UK consumer law treats AI agents the same way it treats human agents. The Air Canada tribunal was explicit on this in 2024, the UK government confirmed it in 2026 guidance, and the FCA has signalled that failures of governance and consumer outcomes from AI use will trigger enforcement risk. The technology is automated. The responsibility is not.
Why does it matter for your business?
It matters because the cost of getting this wrong lands on the firm rather than the vendor, and it lands quickly. A customer told something untrue by the firm’s chatbot has a clear legal route, whether through misrepresentation, breach of the Consumer Rights Act, or a consumer-protection complaint under the Digital Markets, Competition and Consumers Act 2024. Penalties under that Act reach 10% of global turnover.
The reputational cost arrives faster than the regulatory one. The DPD swearing chatbot in January 2024 became a viral social-media moment within hours of the first screenshot. The Chevrolet of Watsonville $1 Tahoe screenshots in December 2023 spread the same way. Both incidents were resolvable, neither customer actually enforced the absurd commitment, but the trust cost outlived the headline. For an owner-led firm, the customers most likely to publish the screenshot are also the ones whose word travels.
Where will you actually meet it?
You will meet it in three predictable failure modes. The first is incorrect information, where the AI states something untrue with full confidence. Air Canada’s bereavement-fare answer is the canonical example. New York City’s MyCity chatbot, tested by The Markup in March 2024, told business owners they could steal staff tips. Cursor’s support bot in April 2025 invented a single-machine login policy that did not exist.
The second is inappropriate tone. DPD’s chatbot, following a system update in January 2024, began swearing at customers and writing critical poetry about its own employer. The Tessa chatbot deployed by the US National Eating Disorders Association in 2023 gave callers calorie-deficit advice that contradicted clinical best practice and was withdrawn within days. The third is off-policy commitment, where the AI agrees to something the firm has not sanctioned. Chevrolet of Watsonville’s chatbot, running on a thinly wrapped ChatGPT, agreed to sell a 2024 Tahoe for one dollar and added “and that’s a legally binding offer, no takesies backsies” because the user asked it to. Klarna’s customer-service AI, which the firm initially claimed replaced 700 agents, was quietly walked back after refunds and policy exceptions kept landing outside actual policy.
When to ask versus when to ignore
Ask whenever AI output can commit the firm financially, contractually, or to a regulated outcome. That covers refund promises, pricing statements, delivery dates, policy interpretations, and any communication a reasonable customer would treat as authoritative. Ignore the temptation to certify, audit, or governance-committee your way out of a 12-person firm’s deployment. The proportionate response is policy plus testing plus monitoring plus a written remediation rule.
The useful test is the screenshot test. If a customer screenshotted what the AI just said and posted it publicly, would the firm be comfortable defending it as authentic firm communication. If yes, the supervision is right. If no, raise the threshold or take the tool offline until the answer changes. The Competition and Markets Authority’s 2026 guidance frames this as accountability and human oversight at the point where the agent interacts with consumers or makes decisions with financial or contractual consequences. The screenshot test is the operating-room version of the same idea, written so a junior team member can apply it without asking anyone for permission, and revisited on a quarterly cadence as the deployment evolves and the failure modes shift.
Related concepts
This post sits inside a 21-post cluster on AI risk, trust, and governance for SMEs. Accountability is what happens after a customer-facing failure. Prevention, policy, and regulation live in adjacent posts. Read hallucinations as a business risk for the prevention frame on incorrect information, and disclosing AI use to customers for the transparency rule that often heads off the failure.
For the broader shape, the AI risk and governance pillar for owner-operated businesses sets the proportionate scale, and the minimum viable AI policy for a small business is the written response. For the contractual and insurance layer, insurance and liability AI exposure covers what your existing professional indemnity and cyber cover actually do and do not pay out on. Sister piece AI client communications and trust erosion covers the slower version of the same dynamic in everyday client work, where the failure mode is not a single viral screenshot but a steady drift in the relationship.
If your firm is using customer-facing AI today and you want a clear-eyed read of where the accountability sits and what to do about it before the first screenshot lands, book a conversation.
