← All posts

Who pays when AI gets it wrong? Contracts, liability and insurance for AI projects

An AI system will produce wrong output. Not might: will. So the contract for an AI project has to say in advance what happens when it does, and four clauses do most of the work: acceptance measured on an agreed test set, a stated error rate with a remedy, a named person who reviews output before it is used, and a liability cap both sides can survive. Two outside rules now shape those clauses. The EU AI Act decides who counts as the provider of a system, and the revised EU Product Liability Directive treats software, AI systems included, as a product for anything placed on the market after 9 December 2026.

Why is an AI error different from a software bug?

An AI error is different because a bug is a defect that can be found, reproduced and fixed, while a language model produces a spread of answers of which some fraction is wrong, and that fraction goes down with work but never reaches zero. A contract written for ordinary software promises to fix defects. Applied to AI, that promise becomes infinite. It helps to separate three situations, because different clauses handle them:

  • An ordinary error within the agreed error rate. Expected behaviour, nobody is in breach.
  • Errors above the agreed rate. The system performs worse than what was accepted. A service problem with a remedy, not a liability event.
  • An error that causes loss. A wrong output reaches a customer or drives a decision and someone loses money. This is where the liability cap, indemnities and insurance come in.

What should acceptance and the error rate say?

Acceptance should say that the system is accepted when it reaches an agreed score on an agreed metric, measured on a test set frozen before the build and shared with the client. "To the client's reasonable satisfaction" does not work for AI, because there is always one more example that fails, and acceptance never arrives. How to build that test set is in LLM evals. Next to it, the contract should state the expected error rate as a number, say that errors within it are expected behaviour and not defects, and set a remedy ladder when the measured rate is worse: re-tuning within a set period first, then a fee reduction, then termination. Every engagement I take starts with a discovery call, a written blueprint and a pilot on the client's own data, and the pilot is what makes this clause writable: nobody can commit to a number they have never measured.

Who reviews the output before it is used?

The contract should name who reviews output, which outputs, at what rate, and what happens if the review is skipped. If the client's staff review, skipping it is their decision and their risk, not the supplier's defect. Human review is the main control on model error, and its absence is behind most losses that become claims. For high-risk systems the AI Act turns this into law: Article 26 requires deployers to assign human oversight to people with the competence, training and authority to exercise it (AI Act, Article 26). Which actions need that approval in practice is covered in human in the loop.

How much liability should an AI supplier accept?

An AI supplier should accept liability capped at a figure it can pay, with indirect losses, lost profit and business interruption excluded. The market norm is the fees paid in the previous twelve months, the same cap model providers such as Anthropic give their own customers: Anthropic's commercial terms limit its liability to the fees paid for its services in the previous twelve months (Anthropic, commercial terms). A supplier that accepts more downstream carries the whole gap alone. Fraud, wilful misconduct, death and personal injury usually cannot be capped in Europe, so expect those outside the cap. Indemnities deserve the same care: an intellectual property indemnity should cover the supplier's own original work, not output generated by a third-party model or material the client supplied, because those are exactly what the providers exclude from their own indemnities.

What is the warranty nobody can give?

Nobody can warrant that AI output is accurate. Anthropic's terms state, in capitals, that it does not warrant that its services or outputs are accurate, complete or error-free. A supplier that promises accuracy to its client is making an unfunded promise its own supplier refused to make. What a supplier can warrant is what it controls: reasonable skill and care, compliance with the law, and a system that met the acceptance criteria at the moment of acceptance. If a client needs a guaranteed accuracy, the honest route is an insured performance warranty, and its cost belongs in the price.

Who is the provider under the AI Act?

Under the AI Act the provider is whoever places a system on the market or puts it into service under its own name, and the deployer is whoever uses it under its authority. Three situations cover most projects:

  1. A bespoke system that goes live under the client's brand. The client is the provider and the deployer; the developer supplies a service.
  2. A system sold under the supplier's own name. The supplier is the provider and carries the heavier obligations.
  3. Someone else's high-risk system under your logo. Article 25 makes whoever puts their name or trademark on it the provider (AI Act, Article 25). Changing a general system's purpose into a high-risk one, such as screening job applicants, does the same.

The contract should name the roles, state the intended purpose, and record who classified the risk and when. The dates by which each duty applies are in the EU AI Act for Greek businesses.

What changes on 9 December 2026?

On 9 December 2026 the revised EU Product Liability Directive starts to apply to products placed on the market after that date, and it treats software as a product, including AI systems and software supplied as a service (Directive (EU) 2024/2853). It compensates death, personal injury, damage to property, and destruction or corruption of data not used exclusively for professional purposes, and it does not cover pure economic loss. It also lets courts order disclosure of evidence and presume a product defective where technical complexity makes proof excessively difficult. The practical reading: an AI error that costs one business money in dealing with another stays a contract question, while AI that touches people's safety, health, devices or personal files now also sits under strict liability.

Does insurance cover AI mistakes?

Insurance may or may not cover AI mistakes, and the policy wording decides. A standard technology professional indemnity policy answers claims of negligence, and a correctly built system that produces a wrong answer is arguably not negligent. A leading UK technology policy page read in September 2026 did not mention artificial intelligence at all. Specialist AI cover exists, such as Munich Re's aiSure (Munich Re), but it is written for vendors at scale, and no Greek-market AI policy turned up in that research. Ask your insurer in writing whether any exclusion mentions artificial intelligence or automated decisions, whether a wrong output from an integrated third-party model is covered, and whether "professional services" includes designing and running AI systems.

What else belongs in an AI contract?

Two clauses are missed most often. First, an express permission to send the client's data to named model providers: without it, a standard confidentiality clause makes every API call a breach, as covered in GDPR and AI. Second, a model change clause that lets the supplier switch models as long as the agreed acceptance score still holds, because providers retire models on their own schedule, as explained in how to choose an AI model. Checking a contract like this against your own templates is one of the systems under AI for legal and compliance.

Frequently asked questions

Can I make my AI supplier liable for every wrong answer?

Not in a contract a serious supplier will sign. A fair contract treats errors within the agreed rate as expected, errors above it as a service issue with a remedy, and loss-causing errors as liability up to a cap.

Who is responsible if our staff act on a wrong AI answer?

That depends on the review clause. If reviewing the output was your team's obligation and it was skipped, the risk sits with you. If the supplier took on the review, it is theirs.

Does the new product liability law cover business software?

It covers software as a product, but it compensates specific kinds of damage, such as injury, property damage and loss of non-professional data. Money one business loses because of another's software remains a matter for their contract.

This is practical guidance, not legal advice, and a lawyer should read any contract before you sign it. If you are about to sign or send a contract for AI work, send me the scope and the acceptance terms, without confidential data. Whether acceptance, the error rate and the review duty can actually be measured is usually visible on one page.