Say you run a small car workshop. You set up an AI assistant to answer WhatsApp messages while you're under a bonnet. A customer asks what a pre-inspection check costs. The assistant says 40 euros. Your price is 90. The customer turns up, shows you the screenshot, and looks at you.
Who pays the 50 euros? You, almost certainly. And the reasoning behind that is worth understanding, because the same logic applies to every mistake an agent can make, not just the cheap ones.
The customer never met your software
A customer doesn't see "the assistant" and "the workshop" as two parties. They see one thing: your business, answering on your number. So if the answer is wrong, they hold you to it.
This isn't a theory. In 2024 a Canadian tribunal (British Columbia's Civil Resolution Tribunal, in Moffatt v Air Canada) dealt with a customer who was told by the airline's website chatbot that he could claim a bereavement discount after travelling. The real policy said otherwise. The airline argued, in effect, that the chatbot was its own thing and the airline wasn't answerable for what it said. The tribunal didn't accept that, and ordered the airline to pay. The sum was small, a few hundred Canadian dollars. The principle was not.
Be careful what you take from this. It's one tribunal, in Canada, and a law firm that summarised it calls its weight "only persuasive" elsewhere. It involved a scripted chatbot, not an agent that acts on its own. It doesn't tell you what a Spanish or any other EU court would do. What it does show is how people instinctively reason: a tool you put in front of customers speaks for you.
What Europe has (and hasn't) written down
You might expect a clear AI liability law to settle this. There isn't one. The European Commission proposed an AI Liability Directive and then withdrew it in February 2025, because member states couldn't agree a text. The revised EU Product Liability Directive now treats software, including AI, as a product, but that's mostly about what happens when the maker ships something defective, not about what you owe a customer when your own agent misquotes a price.
So the honest answer is: for most everyday mistakes, ordinary rules about what you promised your customers and what care you took still do the work. I'm not a lawyer, and whether a specific mistake costs you money depends on the contract, the law that applies and the details. For anything involving invoices, tax or a dispute, ask your gestor or a lawyer. The rest of this piece is about what's in your hands.
Your software supplier probably won't carry it for you
Here's the part that surprises people. You might assume that if the tool is wrong, the company that made it shares the blame. Read the terms. Standard AI service terms often push the risk of outputs onto you, cap what the vendor will ever pay, and sometimes ask you to cover the vendor if a customer claims. As the law firm Astris puts it in its guide on agent liability, those terms bind you and the vendor, not your customers. Your customer can still come to you, and you may have little to claim back.
A negotiated contract can change that. A click-through plan almost never does.
So where does a human belong?
If the answer to "who pays?" is usually "you", the practical question becomes: which mistakes can you afford?
A useful way to sort them: ask what happens if the agent is wrong and nobody notices before it leaves.
- Cheap and reversible, like a draft reply to yourself, a note in a calendar, a summary of yesterday's messages. Let it run. Check now and then.
- Seen by a customer, like a price, a deadline, a promise ("we can do it Friday"), a refund. A person looks first. This is where the 40-euro-versus-90 mistakes live.
- Moves money or can't be undone, like payments, anything filed with an authority, deleting records. A person approves each one, every time.
That's the same line we draw at Ausavia: the agent prepares, you tap approve. It isn't because agents are useless. It's because the person whose name is on the business is the one answering afterwards, so that person should see it first. If you want the longer version of that argument, Is It Legal for AI to Act Without You Checking? goes into what the rules say about oversight, and Who Tells You When Your AI Tool Goes Wrong? covers how you find out something broke.
Where this advice runs out
If your agent only drafts things for you, with nothing going out unseen, most of this worry shrinks. Responsibility questions get sharp when the agent speaks to customers alone, so "not yet" is a fair answer for that step. And a person approving is no magic shield: if you tap approve on forty messages without reading them, you've taken the risk and kept none of the protection. Equally, if a customer is harmed in a way involving personal data, health or a legal dispute, you need proper advice, not a blog post.
One thing to do this week
Take ten minutes and write a short list: "What my assistant may do alone, what I check first, what only I do." Put the price list in a place the assistant must read from, not remember. Then open your AI tool's terms and look for three words: liability, indemnify, outputs. Whatever they say, you'll know what you're really carrying.