Who is liable if an AI agent makes a mistake?

Always a person or a company — a machine has no legal personality and cannot be a defendant. The European Parliament put this on record in its resolution of 20 October 2020, noting that it is not necessary to give legal personality to AI systems, because harm is nearly always the result of someone building, deploying or interfering with the system. Which of those three will be liable in your case depends on the roles set out in the contract and on national law; there is no single European rule for this.

Do the AI Act obligations for high-risk systems already apply?

No. Regulation (EU) 2026/1744, in force since 27 July 2026, postponed their application: to 2 December 2027 for stand-alone systems listed in Annex III, and to 2 August 2028 for systems embedded in regulated products covered by Annex I. The stated reason is that harmonized standards are not ready and the national competent authorities have not all been set up. The transparency obligations under Article 50, meanwhile, still apply from 2 August 2026.

What happened to the AI Liability Directive?

It was withdrawn. The European Commission decided to withdraw proposal COM(2022) 496 final at its meeting on 16 July 2025, and the notice was published in the Official Journal on 6 October 2025. The Commission's work programme gives the reason as no foreseeable agreement, with a promise to assess whether another proposal should be tabled. The directive was meant to introduce presumptions of causality and a right to request disclosure of evidence — without it, the burden of proof on the injured party remains a matter of national law.

Does the new Product Liability Directive cover a business's losses from an agent's mistake?

For the most part, no. Directive (EU) 2024/2853 does name software as a product in Article 4(1) and has to be transposed by 9 December 2026. But under Article 5(1) only a natural person has the right to compensation, and the list of types of damage in Article 6(1) is closed: death or personal injury, damage to property, and destruction of data — with property used exclusively for professional purposes and professional data excluded. A typical business loss from an agent's mistake — a lost customer, a deal that fell through, corrupted CRM records — falls outside this directive.

Can we accidentally become the provider of an AI system without developing anything?

Yes, and it is one of the most commonly underestimated risks. Under Article 25 of the AI Act you are considered a provider, with all the obligations under Article 16, if you put your name or trademark on a system already placed on the market, make a substantial modification to it, or modify the system's intended purpose so that it becomes a high-risk system. The last of these covers an everyday scenario: taking a general-purpose model and wrapping it in your own prompts for candidate screening or creditworthiness assessment. The initial provider then drops out of that role.

What AI law is in force in Ukraine?

As of August 2026, no dedicated one. There is no law on artificial intelligence in force and no bill registered in the Verkhovna Rada, Ukraine's parliament. The Ministry of Digital Transformation published a regulatory roadmap on 7 October 2023 and a White Paper on 26 June 2024, but neither document is a legal act. Cabinet of Ministers Order No. 457-r of 9 May 2025 requires the ministry to submit a bill to the Cabinet only in the fourth quarter of 2026. Until then, disputes are resolved under the general rules of civil and contract law.

What should we do now, before the rules take effect?

Three things, and all of them have to be in place before an incident: a log of the agent's actions tied to the prompt and model version, a written boundary of authority (what the agent does without human confirmation), and a clear allocation of roles in the contract with your contractor. The absence of a government requirement does not reduce the loss — it only means the dispute will be decided on the contract and the evidence. A party that cannot reconstruct exactly what the agent did is in the weaker position in such a dispute, regardless of who is actually at fault.

The agent made a mistake: who is legally liable

The European Union's obligations for high-risk AI systems were due to start applying on 2 August 2026. A week before that date, Regulation (EU) 2026/1744 entered into force and moved them to 2 December 2027.

This is no mere calendar detail. That is the very chapter of the AI Act that sets out what anyone who has deployed a system in their own business must do: oversee it, keep logs, respond to failures. In other words, precisely the part that names a specific person next to a specific agent has been pushed back by sixteen months, while agents are already at work in companies.

That has not made the question "who is liable if the agent makes a mistake" go away. It simply no longer has an answer in the form of an article of a regulation, and it has gone back to where it came from: your contract and your logs.

A machine has no legal personality — that was settled six years ago

The temptation to think that liability for an autonomous decision dissolves somewhere arises every time a system is called "autonomous." Legally, it does not dissolve anywhere: a machine cannot be held liable, so liability stays with people and companies. A claim is brought against a person — natural or legal. An agent is neither.

The European Parliament said so plainly back in 2020, in its resolution on a civil liability regime for artificial intelligence:

Notes that all physical or virtual activities, devices or processes that are driven by AI-systems may technically be the direct or indirect cause of harm or damage, yet are nearly always the result of someone building, deploying or interfering with the systems; notes in this respect that it is not necessary to give legal personality to AI-systems. — European Parliament resolution of 20 October 2020, 2020/2014(INL), paragraph 7

The key part is not the denial of legal personality but what comes before it: harm from an AI system is "nearly always the result of someone building, deploying or interfering with the systems." Three verbs — three different people. The question is not whether someone is liable, but which of the three.

The philosophical debate about machine rights, meanwhile, is very much alive. In Import AI 470 of 24 August 2026, Jack Clark summarizes a lengthy post by Taylor Belrose on why giving AI systems rights is a bad idea; the argument there rests on the impossibility of consciousness, not on the law. Clark himself adds that he feels "deeply confused" about issues of AI consciousness and AI rights. It is worth keeping these two conversations apart: whether a machine deserves rights is an open question, whereas whether it bears liability is a closed one — and not closed in its favor.

"That was the agent's decision" is not a defense in a dispute, because no such defendant exists.

The rule that was supposed to name who is at fault has been pushed back by sixteen months

Regulation (EU) 2026/1744 of 8 July 2026 — the Digital Omnibus on AI — was published in the Official Journal on 24 July and has been in force since 27 July 2026. It amended Article 113 of the AI Act and set new dates of application:

What	Before	Now
Stand-alone high-risk systems (Article 6(2), Annex III)	2 August 2026	2 December 2027
High-risk systems embedded in regulated products (Article 6(1), Annex I)	2 August 2027	2 August 2028
Transparency obligations (Article 50)	2 August 2026	unchanged

The legislator stated the reason candidly in the recitals: the harmonized standards and common specifications are not ready, and the national competent authorities have not been set up. The regulator moved the date because it was the one that ran out of time.

The obligations of whoever deploys the system in their own business — the deployer, in the regulation's terms — moved along with that chapter. In a typical project that party is the client: not the one that developed the agent, but the one that put it to work on its own customers and its own data. Article 26, which sets out those obligations, was not changed by the Omnibus: it will simply start to apply later.

Until the end of 2027 there will be no official checklist you can go through and say "we did everything that was required." The absence of a requirement does not mean the absence of a loss.

The directive that was supposed to make it easier to sue was withdrawn

The other half of the structure was removed even earlier, and far fewer people know about it.

In September 2022 the European Commission proposed a separate directive on liability for AI — the AI Liability Directive, COM(2022) 496 final. It was meant to solve the injured party's hardest problem: proving that it was the AI system that caused the harm when you have seen neither its code, nor its data, nor its logs. The directive introduced presumptions of causality and a right to request disclosure of evidence.

The Commission withdrew that proposal. The decision was taken at its meeting on 16 July 2025, and the notice was published in the Official Journal on 6 October 2025. The Commission work programme for 2025 words the reason as follows, verbatim:

No foreseeable agreement — the Commission will assess whether another proposal should be tabled or another type of approach should be chosen. — COM(2025) 45 final, Annex IV, item 32

It is easy to get the date wrong here, and people often do: February 2025 was only the announcement of the intention in the work programme, whereas legally the proposal was withdrawn on 6 October 2025.

There is no common European rule on how to prove fault in a case involving AI, and there will not be one any time soon. Everything goes back to the national law of the Member State, and that differs from one to the next. That is exactly why any article that promises you "the EU rules on AI liability" in a single list is selling a product that does not exist.

What does arrive on time — and why it is most likely not about you

Against the backdrop of two postponed regimes, one is moving on schedule, and it is worth getting exactly right, because it is the one surrounded by the most confusion.

Directive (EU) 2024/2853 on liability for defective products has to be transposed by the Member States by 9 December 2026 and applies to products placed on the market after that date. Its headline change is the definition of a product:

'product' means all movables, even if integrated into, or inter-connected with, another movable or an immovable; it includes electricity, digital manufacturing files, raw materials and software. — Directive (EU) 2024/2853, Article 4(1)

Software is now expressly named as a product. Hence headlines along the lines of "the developer will be strictly liable for an AI's mistake." Formally, that is true. But further on, the text contains two limitations that usually go unquoted.

The first is who can sue. Article 5(1) gives the right to compensation to "any natural person who suffers damage caused by a defective product." The claimant is a natural person. Your company does not appear in that role.

The second is what for. Article 6(1) contains a closed list of types of damage: death or personal injury, damage to or destruction of property, and destruction or corruption of data. And the last two items come with exceptions: property used exclusively for professional purposes, and data used for professional purposes, are excluded from the directive.

Put this together. The agent miscalculated a discount and you lost a customer worth forty thousand euros. The agent sent a counterparty someone else's terms and the deal fell through. The agent corrupted records in the CRM. None of these cases is covered by the new directive: the claimant here is a legal person, the loss is purely economic, and the data is professional. This is the territory of contract and national law, not product liability.

A detail that also trips people up: the exception concerns property used exclusively for professional purposes. A laptop used both for work and at home does not fall under this exception: "exclusively" is not in the text for decoration.

The trap that does not wait for 2027: you can become a provider

The most expensive mistake in allocating roles is to assume that the role is determined by who issued the invoice. The regulation determines it differently.

Article 25 of the AI Act says that a distributor, importer, deployer or any other third party is considered to be a provider of a high-risk system, and takes on all the obligations of a provider under Article 16, in any of three circumstances:

they put their name or trademark on a system already placed on the market;
they make a substantial modification to a high-risk system, and it remains a high-risk system;
they modify the intended purpose of a system — including a general-purpose model — so that it becomes a high-risk system.

The third item describes an ordinary Monday at many companies. You take a general-purpose model, wrap it in your own prompts for candidate screening or creditworthiness assessment, and by the letter of the regulation you are no longer the vendor's customer but the provider of a high-risk system, with the full set of obligations. The vendor, meanwhile, drops out of that role: Article 25(2) lifts it from whoever initially placed the system on the market — not unconditionally, though, but on condition that they hand over a defined package of documentation and access to the new provider.

The Omnibus did not repeal this article; it strengthened it. It rewrote paragraph 2, spelling out exactly what the initial provider has to hand over to the new one: technical documentation sufficient to assess compliance with Article 16, information about known limitations and failure modes, and targeted technical access for testing and validation. The obligation to conclude a written agreement with component suppliers has been extended to those that supply AI models, tools, services and processes. And fines for infringing these particular paragraphs were added separately, through a new point (da) of Article 99(4).

This article has the same date of application: December 2027. But the contracts you sign this week will easily still be in force by then. The question "what will we be in the regulation's terms when it starts to apply" is cheaper to ask now than to redo the paperwork for the whole supply chain after the fact.

What already applies while the rest waits

There is one block of obligations the Omnibus left alone: the transparency obligations under Article 50 apply from 2 August 2026, which means they are already in effect. These are disclosure requirements: that a person is told they are interacting with an AI system, and that generated content is marked as generated.

And next to it sits the only near-term deadline in this whole area. The Omnibus added a new paragraph 4 to Article 111 of the AI Act, which covers systems already on the market: providers of systems that generate synthetic audio, image, video or text content and were placed on the market before 2 August 2026 have to bring them into line with the marking requirements of Article 50(2) by 2 December 2026.

The resulting order is paradoxical, but that is how it stands: today you are required to tell customers that they are talking to a machine, and nothing is required of you as to who will answer if the machine makes a mistake.

Ukraine has no AI law, and that is an answer too

For a Ukrainian company, the story is shorter. There is no law on artificial intelligence in force, and no bill registered in the Verkhovna Rada (Ukraine's parliament) either.

What does exist is a sequence of plans. The Ministry of Digital Transformation presented a roadmap for AI regulation on 7 October 2023 and a White Paper on 26 June 2024; neither document is a legal act. Cabinet of Ministers Order No. 457-r of 9 May 2025 approved an action plan, and its first item sets the ministry a deadline for drafting a bill on the legal regulation of artificial intelligence development and submitting it to the Cabinet: the fourth quarter of 2026. So by then the draft only has to reach the government, and it will reach parliament later still. In June 2026 a working group on the future law started work under the relevant Rada committee.

So today a dispute over an agent's mistake in Ukraine is resolved under the general rules of civil and contract law. Exactly which rules a court will apply in your situation is a question for your lawyer, with your facts in hand, not for a blog article. We deliberately give no article numbers here: the choice of rule depends on who you are in the particular chain and what harm was caused, and a mistake in that choice costs more than the consultation itself.

Who is actually liable: whoever cannot show what happened

There is no common European rule, there is no dedicated law in Ukraine, and product liability does not cover business losses. Two things remain: what the contract says and what can be proven.

The second is more interesting than the first. In almost any dispute over an autonomous action, the losing side will be the one that cannot reconstruct exactly what happened: what data the agent based its decision on, which prompt and model version it was running, whether a human was there and what they saw. Without a log you can neither prove your case nor rebut the other side's version: you simply have nothing to do it with.

There is, by the way, no requirement to keep such logs yet either: Article 12 of the AI Act on record-keeping and Article 19 on automatically generated logs sit in the same postponed chapter. You keep a log not for the regulator but for the week when you have to explain to a customer or a court why the agent did what it did.

Three things decide a dispute, and none of them can be created after the fact.

The first and most important is the action log. Not "the system is logged," but specifically: what the agent did, when, on what input data, with which prompt and model version, and which human saw that step. Logs that are purged after thirty days are worth nothing in a dispute over a claim that is six months old. The other two are simpler: a written boundary of authority (what the agent does without human confirmation and what it does not) and an allocation of roles in the contract: who is responsible for the quality of the model, who for configuring it to your process, who for oversight in day-to-day work.

Liability falls on whoever did not take care of the boundaries. The dispute is won by whoever still has the records.

Questions to ask before launch

Six questions worth settling before the agent starts working with real customers. They do not replace a lawyer; they save the lawyer hours and you money.

What are we in AI Act terms — a deployer or a provider? If you put your name on someone else's system or change its purpose, the answer may turn out not to be the one you expected.
Does our use case fall under Annex III? Recruitment, creditworthiness assessment, access to services — these are typical candidates. This determines whether December 2027 concerns you at all.
What is the agent allowed to do without a human? The boundary has to be written down, not implied. The difference between "offers a discount" and "grants a discount" is the difference in the cost of a mistake.
Where are our logs and how long are they kept? If logs are purged after thirty days and the claim arrives six months later, you have nothing.
What does the contract with the contractor say about liability for the agent's actions? Not about code quality — about actions in production.
Who is our counterparty in law if the harm is done to a customer? The customer will sue you, not your vendor or the model provider. Recourse against them is a separate story and a separate clause in the contract.

Where this article stops working

This is not legal advice, and the absence of advice on any specific situation here is deliberate: the answer depends on the jurisdiction, the roles of the parties and the nature of the harm, and we know none of those three parameters.

Everything said here about the AI Act and product liability concerns EU law. Ukraine is not a Member State, so these rules cannot be mechanically carried over to a Ukrainian contract, and we deliberately did not try to do that for you.

And dates move. The best proof is this very piece: the obligations that for half a year were written about as starting on 2 August 2026 moved to December 2027 a week before that date. Before relying on any date from here, check it against the version in force on EUR-Lex.

If you already have an agent running that starts the conversation with the customer or changes something in your data, the shortest check takes one evening: take the last week of its work and try to reconstruct three arbitrary decisions. If you cannot, your first risk is not even a legal one but an engineering one.

Grow2.ai builds front-office agents around the client's process and, along with them, the boundaries of authority and the logs from which an agent's decision can be reconstructed. If you would like to go through your own scenario, come and talk to us.

Read next

Sources