If an AI agent makes a mistake, who pays?
Who pays when an AI agent buys the wrong thing or a chatbot gives wrong advice: real cases, the makers' terms, EU and Italian rules, and what to do.
AI-generated image: Internet of Things News / OpenAI ChatGPT
After reading this article you will be able to:
- Explain who usually pays when an AI agent or a company's chatbot gets something wrong
- Tell what the new EU rules on faulty products cover and what they leave out
- Set up an agent so its mistakes stay small, and know whom to contact if one happens
Drafted with AI from primary sources · checked by a separate AI agent · spot-checked by our editors · How we check
On this page
- What did a tribunal decide when a company’s chatbot got it wrong?
- Has an agent ever spent someone’s money by mistake?
- What do the makers’ own terms say?
- What do the EU and Italian rules add?
- How do you keep an agent’s mistakes small?
- What can you do if an agent got it wrong?
- Where does the answer stop?
- Where to go next
- Common questions
Today it is usually you, the person who set the agent to work. An AI agent is an AI that clicks, types and fills in forms on websites for you, and the companies behind ChatGPT, Claude, Copilot and Gemini write that you are responsible for what it does; Anthropic and Google name purchases explicitly. There are exceptions: a company answers for what the chatbot on its own website tells you, and new EU rules, for products put on sale from 9 December 2026, let you claim from the maker when faulty software, AI included, injures you or damages your things or data. When a mistake costs you money, call the shop and your bank first.
This site is written with AI tools, including Claude by Anthropic. This page uses only what makers, public authorities and published studies say, with links. The pages cited were opened on 15 November 2026.
What did a tribunal decide when a company’s chatbot got it wrong?
In November 2022, after a death in the family, Jake Moffatt asked the chatbot on Air Canada’s website about the airline’s lower fares for such trips. The chatbot said a ticket already bought could be sent in for the lower fare within 90 days, even after travelling. That was wrong: the airline’s own page said the discount did not apply once the trip was over. The passenger took the airline to British Columbia’s Civil Resolution Tribunal, an online small-claims tribunal in Canada.
Air Canada argued that it could not be held liable for what the chatbot said. The tribunal member called this “a remarkable submission” and wrote: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.” In Moffatt v. Air Canada, 2024 BCCRT 149, on 14 February 2024, the airline was ordered to pay CA$812.02: CA$650.88 in damages, plus interest and the tribunal fees.
The passenger had kept a screenshot of the chatbot’s answer. Note that the case is about a chatbot that gave wrong information, not an agent that acted.
Has an agent ever spent someone’s money by mistake?
Yes. In February 2025 Geoffrey A. Fowler, technology columnist of The Washington Post, wrote about trying Operator, an early OpenAI agent (republished by the New Zealand Herald). He asked it to find the cheapest dozen eggs for delivery. He wrote that “without my approval” it used his credit card to buy them: US$31.43 once fees and a tip were added. OpenAI told the newspaper it was examining “why Operator occasionally doesn’t send confirmations” and adding “stricter confirmation requirements”. The order went through on his card, so the cost fell on him; the article does not say whether the money came back.
What do the makers’ own terms say?
Four makers of agents that act on websites, in alphabetical order:
| Maker and agent | Who is responsible, in the maker’s words | Before a purchase |
|---|---|---|
| Anthropic, Claude in Chrome | “You remain responsible for all browser actions taken by Claude”, listing “Purchases or financial transactions” among them (help page) | “Claude asks for permission before accessing financial sites” |
| Google, Gemini in Chrome | “You’re responsible for Gemini’s actions during a task, including mistakes and unexpected results like purchases” (help page) | May ask you to take over for steps such as “Finalizing financial transactions” (same page) |
| Microsoft, Copilot | “When you request that Copilot take Actions on your behalf, you are solely responsible for those Actions and any results or consequences” (terms for individuals, effective 18 August 2026) | “Copilot will ask for your attention and supervision for certain actions, such as buying an item” (help page) |
| OpenAI, ChatGPT | You “are responsible for all activities that occur under your account” (terms of use for Europe, updated 16 January 2026) | Asks in the chat before actions such as “confirming a booking or making a payment” (help page) |
Microsoft adds that what you buy through Copilot is “sold and shipped by a third party”, not by Microsoft. Terms are the maker’s side of the contract and do not replace the law of your country, as OpenAI’s European terms say: “You still have the full protections of the laws applicable to you.”
What do the EU and Italian rules add?
Faulty products, AI included. Since 1985, EU rules have made makers pay when a faulty product hurts someone. The new Product Liability Directive (EU) 2024/2853 replaces them and counts software as a product, naming “AI systems” among the examples, whether the software sits on your device or runs over the internet. EU countries must make it national law by 9 December 2026, and it applies to products placed on the market, that is put on sale or into use, after 8 December 2026 (a date corrected on 7 May 2026); older products stay under the old rules.
It covers death or injury, including medically recognised harm to mental health; damage to your things, except the faulty product itself and things used only for work; and data destroyed or corrupted, such as family photos, if it is not used for work. It does not cover “pure economic loss”, money lost without anyone hurt or anything damaged, so a wrong order is outside it. You claim from the maker or, if the maker is outside the EU, from the importer or the maker’s representative in the EU. When a case is too technical to prove, a court can presume that the product was faulty, or that the fault caused the harm, if you show it is likely. You have three years from when you learned, or should have learned, of the damage, the fault and who is liable.
A proposal that was dropped. In 2022 the European Commission proposed a separate AI Liability Directive for claims about harm caused by AI. It was withdrawn; the notice appeared in the EU’s Official Journal on 6 October 2025.
Italy. Under Italy’s AI law, Law 132/2025, the government adopted Legislative Decree 160/2026, in force since 30 September 2026. Its civil part, headed “procedural tools” for compensation, makes a claim for harm caused while using an AI system easier to prove. If you bring facts that make your claim believable, a judge can order the other side to show evidence of how the AI system worked. If the harm came from breaking a rule of the EU AI Act, the link between the two is presumed unless the other side proves otherwise, and the other side is not cleared just because its AI system met the AI Act’s rules. A consumer can also sue in the court where they live. Separately, on 28 October 2026 the Council of Ministers, Italy’s cabinet, gave final approval to the decree that brings the new EU rules on faulty products into Italian law, rewriting articles 114 to 127 of the Consumer Code, according to the parliamentary dossier on the draft.
How do you keep an agent’s mistakes small?
- Give it small jobs you can check. “Add these five items to my cart and stop before checkout” is safer than “do my shopping”.
- Keep the confirmations on. What is an AI agent? gives the menu names, agent by agent.
- Keep your card out of its reach. Do not save a card in the shop accounts the agent uses, and never type card numbers or codes into the chat. If your bank’s app lets you set a low spending limit on a card, use it.
- Keep a record. Save screenshots of what you asked and what the agent did, and the order emails. In the Air Canada case, the screenshot was the passenger’s proof.
What can you do if an agent got it wrong?
- Stop the agent and save the chat, screenshots and order emails.
- Contact the shop at once and ask to cancel. In the EU you can cancel most online purchases within 14 days of delivery without giving a reason, says Your Europe, the EU’s official guide. Exceptions include travel, hotel, car rental and event bookings for set dates, and goods made to order. You usually pay the return postage.
- Call your bank or card issuer if the payment itself is the problem. Italy’s central bank, the Banca d’Italia, says to tell your bank at once when a card is used without your permission. In the US, the Federal Trade Commission explains how to dispute a credit card charge in writing: your letter must reach the card issuer within 60 days after the first bill with the error was sent to you. Say that an agent made the purchase: the official pages do not say whether a bank must treat it as a payment you did not authorise.
- Put it in writing. In Italy, the Banca d’Italia says, your bank must answer a written complaint (reclamo) about a payment service within 15 working days; if you are not satisfied, you can turn to the Arbitro Bancario Finanziario, the banking ombudsman.
- If it was a scam, for example a fake shop the agent was sent to, follow I’ve been scammed: what should I do first?.
Where does the answer stop?
The Air Canada decision is one small-claims decision in one Canadian province, about a chatbot that gave information. The new EU rules on faulty products leave out wrong purchases and apply only to products placed on the market after 8 December 2026. Whether a maker could still answer for a faulty agent under your country’s general law differs by country: for a large sum, ask a consumer association or a lawyer.
Where to go next
- What is an AI agent?: what agents can do, and the settings that keep you in control.
- What can an AI agent do for my small business?: agents at work, and what never to let them do.
- AI for everyone: the complete guide to today’s AI.
- I’ve been scammed: what should I do first?: the first steps if money has gone.
Common questions
Can I get my money back if an AI agent buys the wrong thing?
Often yes, but from the shop, not from the maker of the AI. In the EU you can cancel most online purchases within 14 days of delivery without giving a reason, with exceptions such as travel, hotel and event bookings for set dates and personalised goods. Contact the shop at once, and if the payment itself is the problem, call your bank or card issuer.
Is a company responsible for what its chatbot says?
In the one decision on this page, yes. In Moffatt v. Air Canada (2024), a Canadian small-claims tribunal held the airline responsible for wrong information its website chatbot gave about fares for people travelling after a death in the family, and ordered it to pay CA$812.02. It is one decision in one Canadian province, not a rule for every country.
Is the AI maker responsible if its agent makes a mistake?
The makers' own terms put the responsibility on you. Google writes "You're responsible for Gemini's actions during a task, including mistakes and unexpected results like purchases"; Microsoft's Copilot terms and Anthropic's Claude in Chrome help page say much the same, and OpenAI's terms make you responsible for all activity on your account. These terms do not replace the law of the country you live in.
What happened to the EU AI Liability Directive?
It was a 2022 proposal for EU rules on claims for harm caused by AI. The European Commission withdrew it, and the withdrawal was published in the EU's Official Journal on 6 October 2025. The new Product Liability Directive, which counts software and AI systems as products, does go ahead, for products placed on the market after 8 December 2026.
Does the new EU law cover an AI that deletes my files or photos?
For products placed on the market after 8 December 2026, yes: the new Product Liability Directive covers data that a faulty product destroys or corrupts, as long as the data is not used for work. It also covers injury and damage to your things. It does not cover money lost without anything being damaged, such as a wrong purchase.
Topics: Who answers?, AI agents, consumer rights, liability, Italy