Part 5: Rules, Responsibility and Your First Agent

What the EU AI Act asks of organisations that use AI agents today, who is liable when an agent gets something wrong, and how to start with a first agent in your own work without regretting it. The last of five parts on AI agents at work.

Tue, 03 Nov 2026 · 14 min

Made for you, by AI, from the sources below.

Listen to it in AI Agents at Work

Transcript

The script the two voices were given, with the sound cues taken out. A voice may say a word a little differently, so the audio is the episode.

Mia: Welcome to Walking Points. A man named Jake Moffatt asked an airline's website chatbot about a bereavement fare after his grandmother died. The chatbot gave him the wrong answer, the airline refused to honour it, and then it argued to a tribunal that the chatbot was a separate entity responsible for its own actions.

Adam: And I'm guessing the tribunal didn't go for that.

Mia: It did not, and we'll see exactly why in a few minutes. That case is the doorway into this last part of AI Agents at Work. The topic is Rules, responsibility and your first agent. This part was made in October 2026. It's the part nobody finds exciting until it goes wrong. It earns a whole episode because the rules and the liability decide who answers for what an agent does. We'll take three ideas in order: the EU AI Act in plain words and what applies now, who is liable when an agent acts, and a practical way to start with a first agent.

Adam: Quick thread check first. Where did the series leave off?

Mia: Last Tuesday we covered The People Next to the Agents, about how agents change tasks and roles inside organisations. Nothing in those stories has moved in a way that changes what we said. So, the law. The EU AI Act, formally Regulation 2024/1689, is the European Union's general law on artificial intelligence. It entered into force back in August and switches on in stages, which is why people argue about which date applies to what.

Adam: So what's the basic idea? A list of things you can't do?

Mia: Partly. It regulates by risk. A few uses are banned outright. A larger group counts as high-risk, for example AI used in hiring, in evaluating workers, in credit decisions or in education, and those carry heavy duties. A third group only has to be honest with people, which is the transparency tier. Most everyday uses sit outside all three.

Adam: Okay, but who does the law actually talk to? "The company" is a pretty vague target.

Mia: That's the second thing to understand. The Act talks about roles, not technologies. A provider is whoever develops an AI system, or has one developed, and puts it on the market under their own name. A deployer is whoever uses an AI system under their own authority in a professional setting. A company that buys an agent product from a vendor is the deployer, and the vendor is the provider. Most of the heavy duties fall on providers, and the deployer's duties are lighter but real.

Adam: That sounds tidy. Maybe too tidy. What if a company builds something on top of someone else's model?

Mia: Good push, because the roles aren't fixed by who you are. A company that builds its own customer-facing chatbot on someone else's model may well be the provider of that chatbot as far as the Act is concerned. Where that line falls for a given project is a question for counsel, not for a podcast.

Adam: Fair. So what applies right now, in October 2026, to an ordinary organisation?

Mia: Two obligations. The first is AI literacy, Article 4. Since 2 February 2025, providers and deployers must take measures to support AI literacy among their staff and the other people who work with AI systems on their behalf. According to the European Commission's own questions and answers, nobody needs a certificate and there's no standard course. It can be a workshop or internal guidance, tailored to people's background and the risks of the systems in use, and organisations should keep an internal record of what they did.

Adam: So it's a box to tick, basically.

Mia: Less than that, and a revision called the Digital Omnibus made it softer. Law firm Gibson Dunn summarises it as having to support the development of literacy rather than guarantee a specific level. So it's an obligation of effort, not of result.

Adam: And the second one?

Mia: Transparency, Article 50, which applied from 2 August and was not postponed. The Goodwin law firm is blunt that the core duties are in force. Where an AI system is designed to talk directly to people, the provider must make sure people are told they're dealing with an AI, unless that's obvious. Providers of generative systems must mark outputs in a machine-readable way. Deployers of emotion recognition systems must inform the people exposed to them. And deployers who publish deepfakes, or AI-generated text meant to inform the public, must disclose that.

Adam: Wait, that last one sounds like it would catch a lot of company blog posts.

Mia: There's an exception, and it's the interesting bit. If a piece has gone through human review or editorial control, and someone holds editorial responsibility, the duty to disclose falls away. Breaches can bring fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher, as Goodwin reports. The practical reading is plain. If an agent writes to customers, suppliers or applicants, say it's an AI, early and without making a show of it. If it produces text for the public, a human who takes responsibility changes the picture.

Adam: Right. Now, a lot of people say the whole Act got delayed. Is that wrong?

Mia: It's true of one big block, and that's where people get tripped up. The Omnibus moved the high-risk rules for stand-alone systems, the ones on the Act's list of sensitive uses like employment, credit and education, from this summer to 2 December 2027. High-risk AI built into already regulated products, a medical device for example, moved to August 2028.

Adam: So if I'm deploying an agent for hiring, I've got until the end of next year. That sounds like a relief.

Mia: I'd push back on the relief part. That's when Article 26 starts to bite, and it's a decent description of responsible use anyway. A deployer of a high-risk system must use it according to the provider's instructions, give human oversight to people with the competence and authority to do the job, monitor the system, report serious incidents, and keep the automatically generated logs for at least six months. It must also tell people when a decision about them is made or helped by the system. Commentators point out the content of those duties didn't change, only the date.

Adam: So the sensible move is to design for them before the thing exists.

Mia: Exactly. The working rule: if the agent touches hiring, evaluating people, or credit and access to services, the high-risk list is where to look. If it sorts an inbox, drafts reports or reconciles invoices, the AI Act mostly asks for literacy and honesty.

Adam: Okay, that's the rulebook. Now back to Mr Moffatt. Who pays when the agent gets it wrong?

Mia: One note first: this is general information, this isn't legal advice, and a lawyer who knows your sector is the right person for a real decision. Now the case, because no exotic law was involved. Moffatt v. Air Canada was decided by the Civil Resolution Tribunal of British Columbia. The chatbot said he could buy a ticket and apply for the bereavement discount afterwards, within 90 days. The airline's actual policy didn't allow that. The tribunal found that Air Canada had negligently misrepresented its policy, and that it made no difference whether the information came from a static page or a chatbot. It ordered the airline to pay about CA$812 in total, according to the McCarthy Tétrault law firm.

Adam: That's a tiny amount for a case everyone cites.

Mia: The amount is small and the principle is the point. A business is answerable for what its automated front door says, and "the bot did it" is not a defence. It's a Canadian tribunal, so it doesn't bind anyone in the EU, but it shows how courts reason.

Adam: Hang on, if I hire a receptionist and they say something wrong, I'm liable too. Is a chatbot really different?

Mia: Not on the tribunal's reasoning. It made no difference whether the words came from a page or a chatbot, and courts reason this way because the company chose to put the tool in front of customers. So the company answers for what comes out.

Adam: Okay, fair. So in Europe, is it just that one idea?

Mia: The same instinct runs through several layers. First, contracts. Between a company and the vendor, the risk is mostly written down: what the vendor promises, what it excludes, what it caps. A company that says "we just use the tool" can discover it carries all the risk. Second, ordinary negligence toward the people on the receiving end, the Moffatt layer, which lands on whoever the customer dealt with. Third, product liability. A new EU Product Liability Directive has to be turned into national law by 9 December. As Reed Smith and the European Commission describe it, software counts as a product, including AI systems. Liability is strict, so the injured person only has to show the product was defective and caused harm, and a vendor can't disclaim its way out.

Adam: Mia, that sounds like it would scare every company using an agent.

Mia: For most office uses, no. It protects injured individuals, for things like injury and damage to private property, so it's mainly a vendor's problem. But a company that builds its own system, or substantially modifies one, should ask whether it has become the manufacturer. There was also a proposed AI Liability Directive, but the Commission withdrew it in October. And the fourth layer is data protection. Under Article 22 of the GDPR, people have a right not to be subject to a decision based solely on automated processing. In the SCHUFA case in December 2023, the Court of Justice of the European Union ruled that a credit agency's automated score counted as such a decision when the lender draws strongly on it, even though a human at the bank made the final call.

Adam: So a person clicking approve doesn't automatically fix it.

Mia: Right, and that ties back to the failures part, where approving became a reflex. So the honest answer to who is liable is: usually the organisation that deployed the agent, toward the people it affected, with a claim against the vendor if the contract allows. Inside an organisation, the answer is a named human.

Adam: Then let's get practical. If I'm starting from nothing, what's the base rate of this going badly?

Mia: Gartner predicted in June that over 40 percent of agentic AI projects would be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls. That's a forecast, not a measurement, but its three causes match what the earlier parts found. What follows is the show's own synthesis of the series, not a rule from any source.

Adam: Go on then. Where does somebody begin?

Mia: Pick the task before the tool. A good first task is bounded, repeated often, easy to check and reversible if wrong, like a first draft of a weekly report, matching invoices to orders or triaging a shared inbox. Anything that talks to customers about money, health, contracts or hiring doesn't fit, not for the first one. Then climb a ladder one rung at a time: the agent only reads and suggests, then drafts with a person sending, then acts on small reversible things with approval, and only after a stretch of checked results acts alone on a narrow slice.

Adam: That sounds slow. Won't people skip rungs?

Mia: They will, and the cost of a mistake decides how many rungs you need. Access matters too. The lethal trifecta from the failures part, private data, untrusted content and a way to send things out, is answered on day one by never giving all three to the same agent. Add a budget cap and an off switch, because a scheduled agent ran up about 1,800 dollars in two nights.

Adam: And who's responsible day to day?

Mia: One human owner. Not a committee: one person who reads the logs, answers for the results and can say stop. Then tell people. Say in the message that an AI handled it, tell colleagues what the agent does and doesn't do, and give them a way to flag a mistake. For literacy, one hour where the people who'll work with the agent see what it gets wrong, with real examples from their own task, plus a note that you did it, meets the spirit better than a certificate.

Adam: And the last piece?

Mia: Deciding in advance when to stop, with a date and a result at which the experiment is judged.

Adam: I'd guess that's where the Gartner number comes from, projects that never get judged.

Mia: That's the reading. So what changes in a working role? Know which tasks the law treats as high-risk and keep agents away from them for now. Expect to be asked to say when an AI is talking. And treat yourself as the named human, because for the courts and regulators we've covered, a person is the one who answers.

Adam: Is there one thing a listener could do this week?

Mia: One step. Find one place where an AI at your work writes to customers, suppliers or applicants, and check whether the message says it's an AI. The reason is that saying so is already expected since 2 August, and when that message gets something wrong, the organisation answers for it, not the bot.

Adam: That's a clear one. And it closes the loop on the series.

Mia: It does. Part 1 explained what an agent is, a system that works in a loop and acts. Part 2 looked at where agents earn their keep, Part 3 at how they fail, Part 4 at the people working next to them. This part is the frame around all of it: a law that mostly asks for literacy and honesty, a liability picture where the deploying organisation is usually the one answering, and a small, careful way to start. Which brings us back to Jake Moffatt, whose airline learned that the bot's answer was its own.

Mia: Want a show like this about your own work? walkingpoints.fm

Adam: Thanks for walking with us.

Mia: Take care out there, and I hope the walk is a good one.

Heard something wrong? Tell us