Skip to content
All articles
Law & regulation 8 min read· by Lars Zimmermann

AI decisions: who answers for them in your company

Everyone argues about the liability of AI makers. Who in your company answers for today’s decision, and how to assign that responsibility so it holds.

In short

Internally, management answers for AI decisions until it has clearly assigned the responsibility. Germany’s new Product Liability Act, passed by the Bundestag on 8 October 2026, protects natural persons and does not cover your company’s pure financial losses. Assigning means: a named person, a clear boundary, authority to halt, a reporting route for concerns. Supervision stays with you.

Auf Deutsch lesen: deutsche Fassung

Update of 28 October 2026. At the time of writing, the Act to Modernise Product Liability Law was still before the Bundesrat, the second chamber of the German parliament. In the meantime, on 16 October 2026, the Bundesrat did not call for the mediation committee, and the Act has been promulgated (Federal Law Gazette, BGBl. 2026 I No. 321 of 27 October 2026). It enters into force as planned on 9 December 2026, Articles 2 and 5 to 16 already on the day after promulgation. That changes nothing about what the text says, but it does change the procedural status. I am leaving the text as it is and putting the reference above it, instead of correcting it silently.

On 17 September 2026, Palantir chief executive Alex Karp sat at CNBC and said a sentence that has been running through the whole AI debate ever since: whoever builds AI and causes harm with it should answer for it under civil and criminal law. The first line of defence, he said, is liability: companies and their bosses should stand up for their own actions. Sensible guidelines would be fine with him, but such things cannot be enforced. He accuses the labs calling for regulation of wanting to use it to protect themselves from their own customers’ lawsuits.

Since then three camps have faced each other. Some want to slow development down. Others consider the market oversight enough. And Karp wants to see the makers in court. All three are talking about the same question: what do those who build models owe?

The question that a managing director of a mid-sized company has to answer today is asked by nobody in this debate. This morning your AI calculated a quotation, pre-sorted a complaint, triggered a purchase order. Who in your company answers for that? Not before a US court. At your place, on Tuesday, when the customer calls.

Who is liable externally, I have taken apart in a separate text. This one is about the inside: who is responsible within the company, how to assign that so it holds, and what an auditor wants to see when asking about it. This is my assessment as an auditor, not legal advice.

Who is liable towards customers and injured parties, what the new product liability law changes and what insurers are currently rewriting is covered in the article on the outside.

To external liability

Manufacturer liability ends at your office door

In Germany of all places, the manufacturer side has just become very concrete. On 8 October 2026 the Bundestag passed the Act to Modernise Product Liability Law, which can be read as Bundesrat printed paper 584/26. It transposes Directive (EU) 2024/2853, expressly counts software among products and is due to enter into force on 9 December 2026. The Bundesrat is currently deliberating on it in the second round.

Anyone who concludes from this that the liability question has thereby landed with the provider should read Section 1(1) of the new German Product Liability Act (ProdHaftG). There, compensation goes to a natural person, and for three kinds of damage: to life and limb, to an item of property that is not used exclusively for professional purposes, and to data that do not serve professional purposes.

Put next to that what your AI actually decides in the business. A quotation with the wrong material price. A justified complaint that is sorted out as unfounded. A supplier who slips in the rating although he delivers on time. All of these are financial losses in business dealings. For none of them does the new Act provide a claim, neither for your company nor for your business customer.

Even if Karp got every maker into court tomorrow: the decision your AI took in quotations today would not be affected by it.

Whether you can hold the provider to account for such a loss is decided solely by your contract with it. Read the section on limitation of liability before you rely on it. A second passage in the new Act belongs in every rollout plan. Anyone who substantially modifies a product without the manufacturer’s consent and then puts it into service is, under Section 5, themselves the manufacturer of the modified product. Retraining, deep rebuilding, passing it on to customers under your own name. That is exactly the edge at which a great deal of tinkering is going on in mid-sized companies right now.

For products placed on the market or put into service up to and including 8 December 2026, the old Act of 1989 continues to apply under Section 23. Anyone introducing a system in these weeks should therefore record the date in the rollout record. Otherwise someone will ask about it in three years, and nobody will remember.

Inside there is no gap, only a blur

Internally the legal position is simpler than most people think. Section 43(1) of the German Limited Liability Companies Act (GmbHG) requires managing directors to exercise the diligence of a prudent businessman. Subsection 2 makes them liable to the company for a breach of duty. No separate rule for AI is needed for that. What the company uses falls within the diligence of those who run it, for as long as they have not recognisably distributed the task.

The everyday problem is therefore rarely a gap in the law. It is a blur within the company. The conversation after an error goes like this or something like it:

  • Purchasing says: IT set up the system.
  • IT says: the business department specified the threshold.
  • The business department says: the model decided, we only approved what it proposed.
  • The management says: I thought IT was taking care of it.

Each of these sentences is true. Together they add up to nobody. Ill will is rarely behind it. An AI system works right across three departments and belongs to none of them.

How to distribute something like this cleanly was described by the German Federal Court of Justice (BGH) for a related case back in 2018. The judgment of 6 November 2018 (II ZR 11/17) concerned the division of tasks between several managing directors. Such an allocation of responsibilities only provides relief if it delimits the tasks clearly and unambiguously, is supported by all, and ensures that they are performed by professionally and personally suitable persons. What cannot be delegated remains with the board as a whole. For the other managing directors the duty turns into a duty of supervision. The court does not strictly require written documentation.

The judgment concerns the level of the management. The court did not rule there on delegation to employees. It nevertheless serves as a test standard for any assignment within the company, because it names the three things on which responsibility fails in an emergency: a blurred boundary, an unsuitable person and oversight that falls asleep after the assignment. That no written form is required should not reassure you, by the way. Anyone who has to prove in a dispute how the tasks were distributed is better off with a sheet of paper than with a memory.

Four words from Article 26 that work for any AI

For the same question the EU AI Act has found a wording that I consider the most usable in the whole text. Article 26(2) obliges deployers of high-risk AI systems to assign human oversight to natural persons who have the necessary competence, training and authority, and to give them the necessary support.

Legally this applies only to high-risk systems, and even there only from 2 December 2027 for the cases under Annex III and from 2 August 2028 for Annex I. That is what Article 113 says as amended by Regulation (EU) 2026/1744. Most AI applications in mid-sized companies do not fall under it at all. As a test for each individual one, the four words work all the same:

  • Competence: does the person understand what the system does, and do they recognise a wrong proposal before it goes out?
  • Training: have they learned how this specific system fails, or only how to operate it?
  • Authority: may they halt the system without asking anyone first?
  • Support: do they have time for it in the work schedule, a deputy during holidays and backing when halting costs an order?

In practice it tips over at the third word. Article 26(5) requires deployers to inform the provider and the market surveillance authority and to suspend use as soon as they have reason to consider that the system, when used as intended, presents a risk to health, safety or fundamental rights (risk within the meaning of Article 79(1)). But only someone who is allowed to can suspend. The best test question on this fits into one line: who may switch off your AI today at 2 pm without calling anyone first?

If the answer comes with a name and without hesitation, the responsibility is assigned. If the answer contains an ‘actually’, it is not. And if the answer is ‘the management’, you have handed the responsibility back to where it lay anyway. Nothing is gained by that, because at 2 pm the management is rarely sitting in front of the screen on which the error is just arising.

What an auditor wants to see when asking about responsibility

ISO/IEC 42001 turns the same question into four auditable points. Clause 5.3 places the assignment of roles, responsibilities and authorities expressly in the hands of top management. Anyone can prepare the assignment. Under the standard, ensuring that it is in place and known is for the management.

The three other points are in Annex A, and for each there is a typical nonconformity:

  • A.3.2, AI roles and responsibilities. Typical is a matrix that names an AI officer and the head of IT and nothing else. Who exercises human oversight, who is responsible for data quality and who manages the provider is stated nowhere. As a result, half a dozen further controls lack a responsible person.
  • A.10.2, allocating responsibilities between you, your suppliers and your customers. Typical is a contract with the model provider that regulates availability and price, but not who answers for model quality and bias. At the first misbehaviour of the system, the responsibility is then in dispute.
  • A.3.3, reporting of concerns. Typical is a reference to the existing whistleblowing system. That is designed for breaches of law. An employee who notices that the complaints AI has been rejecting every case from a particular customer for two weeks finds no category for it there.

I consider the third point the most underestimated. Responsibility also means that the people who see an error first can get it off their chest. Those are rarely the named persons. It is the clerk, the foreman, the working student who have the outputs in front of them every day. Without a reporting route their observation remains a feeling, and the named person only learns of the problem from the customer. The guidance to the standard requires for this route, among other things, confidential reporting, qualified staff to handle reports, protection from reprisals and escalation to management within an appropriate time.

In the audit, the answers to the question about responsibility usually differ in the very first sentence:

  • Weak: IT is responsible for AI here. Robust: the head of internal sales is responsible for the quotation module, her deputy stands in for her, and both may switch it off.
  • Weak: we have an AI officer. Robust: the AI officer keeps the overview, oversight of the three systems lies with the three business departments, and here is the list of names.
  • Weak: anyone can raise concerns here at any time. Robust: here is the reporting route, here are this year’s reports, and here is what became of each one.

What a clean assignment does in your favour

Up to here it all sounds like obligation. The other side is at least as important. An effective assignment protects the company, and it protects you. By the logic of the BGH judgment, where there is an effective division the duty of the other managing directors shrinks to supervision. Anyone who has assigned cleanly and demonstrably exercises oversight stands differently in a dispute from someone who carried everything themselves and documented none of it.

The second gain is speed. A system that someone may halt is halted earlier, and a system halted early costs less than one halted late. Anyone who first has to reach the management loses, if in doubt, an afternoon. During that afternoon the AI keeps working.

The third gain is the quietest. Employees who know that they may stop look more closely. Those who believe that someone else is responsible anyway stop reading the outputs. That is exactly where the error arises that half the world is currently arguing about at maker level, only without a television studio and with your name underneath.

What you do with this on Monday fits on half a page. Take the one AI system that makes the most decisions in your company and write down five lines:

  • Which decision does the system take or prepare, and where does it end?
  • Which person is responsible for it, by name and not by department?
  • May this person halt the system, and who deputises for them?
  • Where does someone report who finds something odd, and who reads it?
  • What does the contract with the provider say about errors of the model?

Then go to the person from the second line and ask them whether they would be allowed to switch the system off today. If they hesitate, the assignment is still paper. If they answer at once, you have clarified more in half an hour than the whole debate has since September.

Karp has a point when he calls liability the first line of defence. In your own company it still does not begin with the maker, but with a question you can answer today: who at your company may halt the AI, and does that person know it?

Share: LinkedIn E-Mail

Frequently asked questions

Is the AI provider liable if its AI decides wrongly in my company?+

Under the new German Product Liability Act, which the Bundestag passed on 8 October 2026, only within narrow limits. There, compensation goes to natural persons for damage to life and limb, to property that is not used exclusively for professional purposes, and to data that do not serve professional purposes. A pure financial loss of your company is not included. Whether the provider answers for it is governed by your contract with it.

Can I hand over responsibility for the AI to an employee?+

You can hand over the task; you keep the supervision of it. An assignment holds if the boundary is clear, the person is suitable and has the authority to halt the system. With such an assignment, your own duty narrows to supervision. You should still record the assignment in writing, because in a dispute what counts is what you can prove.

Does the law prescribe an AI officer?+

The EU AI Act does not prescribe an AI officer for deployers. For high-risk systems, Article 26(2) requires human oversight to be assigned to persons with competence, training and authority. ISO/IEC 42001 requires roles according to the needs of the organisation in A.3.2. A single title without oversight roles in the business departments is rarely enough for that.

From when does Article 26 apply to my company?+

Only if you use a high-risk AI system as a deployer. Under Article 113 as amended by Regulation (EU) 2026/1744, the chapter applies to systems under Annex III from 2 December 2027 and to systems under Annex I from 2 August 2028. The four test words competence, training, authority and support already work today for any system.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.

Questions about your own case?

In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.

Continue reading