Updated 1 November 2026: rewritten with the liability routes in Turkish and EU law, cited and linked.

When an AI agent acts on its own, the business that deployed it is usually the first to answer, because the software itself cannot hold rights or debts. Depending on the facts, liability then runs through contract and tort law (in Türkiye, the Code of Obligations, TBK), product liability against the developer (in the EU, Directive (EU) 2024/2853, for products placed on the market after 9 December 2026), and regulatory duties under the EU AI Act. No single law deals with agents as such, and this note found no published Turkish high court ruling on one, so the Turkish analysis below is a reading of general provisions.

Can an AI agent itself be liable?

No. Article 8 of the Turkish Civil Code gives capacity to hold rights and debts to human beings, and the Code treats legal persons separately. Nothing gives it to software. Courts elsewhere have started from the same point. According to Chinese press reports, the Hangzhou Internet Court held that AI has no civil-law personality and cannot declare intent. In Moffatt v. Air Canada, a Canadian tribunal rejected the idea that a chatbot was separate from the airline. That small-claims decision has no binding force.

Does a contract made by an agent bind the company?

It may well, and the answer turns on the authority the company created towards the other side. Under TBK Article 1, a contract forms by mutual declarations of intent, which may be express or implied. Because the agent is not a person, it fits poorly as a "representative" under Article 40. The more natural reading is that the company speaks through a tool, with the agency rules applying by analogy to questions of scope.

  • Where authority has been communicated to third parties, its extent follows that communication (TBK Article 41). Limits kept internal may not help against a counterparty who never saw them.
  • An act without authority binds the principal only if the principal approves it afterwards (ratification). The other party may ask for an answer within a reasonable time (TBK Article 46).
  • The duty of honesty in Civil Code Article 2 protects reasonable reliance. An agent using a company's domain and signature block invites it.

A slow reaction to a known error is also risky, because the facts may be argued as ratification. The site's AI agent liability guide gives scenarios and a template. Courts do not all agree on outward statements, however. The Hangzhou court reportedly held that a chatbot's own compensation "promise" did not bind the operator, while Moffatt held the airline responsible for wrong information.

What if the agent causes harm to someone outside a contract?

Turkish tort law offers three possible routes, each with a different burden of proof.

  • Fault (TBK Articles 49 and 50). The injured party proves the harm and the defendant's fault. For agents, fault would mean poor design, oversight or setup by a person or company.
  • Employer liability (TBK Article 66). An employer answers for employees unless it proves due care. Article 66(3) adds that a business answers for damage caused by its activities unless it proves its working order was suited to preventing it. An agent is not an employee, but this organisational duty fits badly governed automation well, in this note's reading.
  • Significantly dangerous businesses (TBK Article 71). This no-fault route applies to a business capable of causing frequent or serious harm even with full care. Most office agents are unlikely to meet that test.

Claims lapse two years after learning of the harm and the liable person, and in any event ten years after the act, unless criminal law sets a longer period (TBK Article 72). The site's analysis of strict liability and algorithmic error looks at this further. Between contract parties, TBK Article 112 is easier for claimants: the debtor must show it was not at fault. Under Article 115, a clause excusing gross fault in advance is void.

Can the developer be sued under the new EU Product Liability Directive?

Yes, for some harm. Directive 2024/2853 applies to products placed on the market or put into service after 9 December 2026, which is also the transposition deadline.

  • Software counts. Article 4(1) includes software as a product, and recital 13 says AI system providers should be treated as manufacturers, whether supplied on a device, through the cloud or as a service. Free and open-source software supplied outside a commercial activity is excluded (Article 2(2)).
  • Learning counts. Defectiveness takes into account a product's ability to continue to learn or acquire new features (Article 7(2)(c)).
  • Non-EU makers do not escape. Importers, authorised representatives and, failing those, fulfilment service providers can be liable (Article 8(1)(c)).
  • Proof is easier. Courts can order disclosure of evidence (Article 9). Defect or causation can be presumed, including where technical complexity makes proof excessively difficult (Article 10).
  • Exclusion clauses do not work against the injured person (Article 15).

The limit is important: only natural persons can claim, and only for death or personal injury, damage to non-professional property, and destruction or corruption of non-professional data (Article 6). A company losing money to a mistaken agent payment is outside the Directive and must use contract or national tort law, which Article 2(4) preserves. The separate AI Liability Directive proposal was withdrawn. The Commission's notice in the Official Journal of 6 October 2025 lists it among the withdrawn proposals, approved at the Commission's meeting of 16 July 2025, and points to Annex IV of COM(2025) 45 for the reasons.

In Türkiye, Law No. 7223 makes manufacturers and importers liable for damage caused by a product, bars contractual exclusion, and sets limits of three and ten years (Article 6). It defines a product as any "substance, preparation or goods" and requires a link between non-conformity and damage (Articles 3 and 6). It does not mention software, so its application to a stand-alone agent is uncertain.

What does the EU AI Act require of those deploying agents?

The Act does not name agents as a category, so duties depend on what the agent does. It is also a regulatory law: it does not itself create a damages claim.

  • Article 50(1): since 2 August 2026, providers must design systems so people know they are dealing with AI, unless that is obvious.
  • Article 26: if the agent's use falls within Annex III, for example ranking job applicants, deployers must from 2 December 2027 (Regulation (EU) 2026/1744) follow instructions for use, assign competent human oversight, monitor the system, suspend it if a risk emerges, and keep logs for at least six months.
  • Article 25(1)(c): a company that turns a general agent to an Annex III use becomes the provider of a high-risk system.

Breaching these duties can bring fines of up to EUR 15 million or 3% of turnover (Article 99(4)). A breach could also support a presumption of defect under Directive Article 10(2)(b), a point still untested. Where agents make decisions about people, GDPR Article 22 and KVKK Article 11(1)(g) also apply.

What can organisations do to manage this exposure?

Most of these routes turn on what the organisation can show it set up, so records matter most.

  • Write down authority. Define what the agent may do alone, what needs confirmation, and spending limits, and tell counterparties where that matters.
  • Keep logs. Goals, steps, tool calls and data touched support any defence under TBK Article 66(3) and any answer to a disclosure request.
  • Allocate risk in contracts, within the limits of TBK Article 115.
  • Name a human owner. Someone needs authority to stop the agent and to reject or ratify its acts quickly.

Vendors can be exposed too. In Mobley v. Workday, a pending US case, the court held in a 2024 interim ruling that a screening software vendor could be liable as an employer's agent.

Frequently asked questions

Who is liable when an AI agent makes a mistake?

Usually the business that deployed the AI agent answers first, because software has no legal personality under Turkish law. The deployer may then have recourse against the developer or vendor under contract, or under product liability rules. The exact route depends on who was harmed and how.

Is a contract signed by an AI agent binding in Türkiye?

It can be. Under the Turkish Code of Obligations, a contract forms by declarations of intent, and the extent of authority follows what was communicated to third parties (TBK Articles 1 and 41). A company that lets an agent negotiate in its name risks being bound. No published Turkish high court ruling on this point was found.

Does the EU Product Liability Directive cover AI software?

Yes. Directive (EU) 2024/2853 treats software as a product and AI system providers as manufacturers, for products placed on the market after 9 December 2026. Only natural persons can claim, for death or injury, damage to non-professional property, or loss of non-professional data. Purely commercial losses are outside it.

Sources: Laws No. 6098, 4721, 7223 and 6698 (mevzuat.gov.tr); Directive (EU) 2024/2853; Regulations (EU) 2024/1689, 2026/1744 and 2016/679; decision summaries on this site, all read on 1 November 2026. The withdrawal of the AI Liability Directive proposal is taken from Official Journal notice C/2025/5423. General information, not legal advice.