The case that set the standard
The chatbot promised, the company paid
In 2022, Canadian Jake Moffatt lost his grandmother and asked Air Canada’s chatbot about bereavement fares. The bot replied that he could buy a full-price ticket and apply for the discount retroactively. That was wrong – the airline’s own policy did not allow retroactive refunds, and Air Canada rejected his claim.
Moffatt took the airline to a tribunal, which ruled in his favour in 2024. Air Canada’s defence was remarkable: the company argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal flatly rejected this – the chatbot is part of the company’s website, and the company is responsible for all information on it, whether it comes from a static page or a bot.
The amount was modest, some 650 Canadian dollars. The principle was not: a company owns what its AI says and does. That principle has since become the foundation of European regulation.
The regulation
The AI Act: where the requirements stand today
The EU’s AI Act has entered into force in stages. The bans on unacceptable AI have applied since February 2025, and on 2 August 2026 the transparency requirements in Article 50 became enforceable: chatbots and other AI systems that interact with people must clearly disclose that the user is talking to an AI, and AI-generated content must be machine-detectable. Breaches can cost up to EUR 15 million or 3 per cent of global turnover.
The requirements for high-risk systems, including AI used for recruitment, credit scoring and access to essential services, have been postponed to December 2027 by the so-called Digital Omnibus. The postponement is no free pass, however: the requirements for risk management, logging, human oversight and documentation are settled, and systems deployed now must be able to meet them.
At the same time, the European Commission withdrew its proposal for a separate AI Liability Directive in 2025. The consequence is that liability for AI failures is decided under national law – and there, the Moffatt logic applies: whoever puts the AI into operation is liable for what it says and does towards customers and the outside world.
The risk
The questions you must be able to answer
- Which AI systems do you have? Most organisations have more than they think – purchased, embedded in SaaS products and built in-house. Without a complete inventory, nobody can assess the risk.
- Who owns them? Every AI needs a named owner responsible for what it says and does – as Air Canada should have had for its chatbot.
- What can they access? A chatbot that can read price lists and customer data can also promise discounts and leak data. Its access determines the damage potential.
- Can you document it? Logging and traceability are both a requirement in the regulation and your only evidence if a case arises.
- Do customers know they are talking to an AI? The Article 50 transparency requirement applies now and is enforced by national supervisory authorities.
AI compliance is access management
The discussion about AI liability almost always ends in the same place: what could the system do, and who had allowed it? An AI agent or chatbot is in practice a non-human identity with access, entitlements and the ability to act – and it must be governed as one.
That means least privilege: the customer service bot should be able to read the approved answers, not the entire intranet. It means ownership and lifecycle: the AI is created through a controlled process, recertified at fixed intervals and decommissioned when no longer used. And it means traceability: who gave the AI which access when, and what has it used it for?
With those disciplines in place, Article 50 and the documentation requirements are not a fire drill but an extract from systems you already control. Without them, every new AI system is a Moffatt case waiting to happen.
Where should you start?
With the inventory and the access. Map your AI systems, give each one an owner, and review what they can actually reach. It is the same exercise as for service accounts and API keys – AI agents are simply the newest non-human identities in the line.
From there, compliance can be built on top: transparency towards users, logging of system behaviour and documentation you can produce when the supervisory authority or an unhappy customer asks. The sooner access management is in place, the cheaper the rest becomes.
Do you know what your AI can promise on your behalf?
An IAM Healthcheck maps your identities and access – people, service accounts and AI agents – and gives you a prioritised plan before the regulator or your customers start asking questions.
