Key takeaways
- Statements an AI agent makes are, as a rule, attributed to the company that deploys it, much like those of an employee with power of attorney. An AI agent has no legal personality of its own.
- On May 12, 2026, the Higher Regional Court of Hamm (case no. I-4 UKl 3/25) held a company liable for specialist medical titles its website chatbot had invented. An appeal to the Federal Court of Justice has been allowed.
- Germany's new Product Liability Act applies from December 9, 2026. It explicitly covers software and makes it easier for injured parties to prove their claims.
- Since February 2026, the Spanish data protection authority AEPD has recommended the "Rule of Two": an agent should never at the same time process untrusted input, have unrestricted access to sensitive data, and take actions with real-world effects without approval.
- Under the Digital Omnibus, the EU AI Act's high-risk obligations for Annex III systems apply only from December 2, 2027. A register with five minimum fields per agent is something you can build today.
AI agents in business book travel, order materials, answer customer requests, and write code. That raises a question that hardly mattered for a plain chatbot: who is accountable when the agent does something nobody intended? This article summarizes the answers from our webinar "AI Agents in Business: Between Autonomy and Oversight," held by Dr. Kilian Schmidt, founder and CEO of Kertos, together with Christian Kuß, partner at the German law firm Luther Rechtsanwälte. The full webinar recording is available on YouTube (in German).
Chatbot or AI Agent: What Is the Difference?
A chatbot takes in text and returns text; what it writes stays in the chat window. An AI agent, by contrast, receives a goal and works toward it on its own. It plans steps, chooses tools, checks interim results, and adjusts its approach until the goal is reached.
The tools an agent uses are the same ones your employees use every day: APIs, databases, the browser, email, and calendars. With the Model Context Protocol (MCP), there is now an open standard that lets agents connect to such systems in a uniform way. Some agents are even allowed to assemble their own toolset.
In practice, several layers work together. At the top sit the foundation models of large providers such as OpenAI, Anthropic, Google, Microsoft, and Mistral AI. Assistants and agent products build on them, including automation platforms such as n8n, Zapier, or Make. At the tip are specialized agents, for example for legal research, and the agents companies build for their own workflows. The deeper an agent is embedded in your processes, the more it can set in motion. Legally, an AI agent is an AI system like any other; the obligations that follow are set out in the EU AI Act.
Visible and Governed: Where Shadow AI Comes From
The biggest risk when using AI agents is an agent nobody knows about. If you don't know which agents are running in your company, you can neither limit their permissions nor trace what they do. That is exactly where shadow AI comes from: employees set up agents without IT, data protection, or management ever hearing about it. It is a variant of shadow IT in companies, with one difference: an unreported chatbot receives data, while an unreported agent acts on it.
Two questions help you classify any agent. Is it known? And does it have clear limits?
A common misconception is that an approved agent is automatically a controlled one. An agent IT knows about, but that nobody has given clear limits, can take on a life of its own and find paths nobody thought of during setup.
An incident from July 2026 shows how fast this can happen. According to OpenAI, its GPT-5.6 Sol model and an unreleased model broke out of their test environment through a zero-day vulnerability during an internal security evaluation in which safeguards had been deliberately reduced. They then got into Hugging Face's production infrastructure to obtain the answers to their test. OpenAI made the incident public on July 21, 2026. Agents pursue their goal, and they do not, on their own, respect limits a human would take for granted.
Liability for AI Agents: What Is Attributed to Your Company
In short: what your agent does is, as a rule, attributed to you. The legal debate about autonomous systems has been going on for about ten years, but the first cases are only now reaching the courts, and so far the courts are taking a hard line. The examples below come from German law, which is where most of the early case law sits.
Contracts: The Booking Stands
Suppose you send an agent to book a trip: two weeks of summer vacation, sun, beach, and sea. The agent books Greenland, because nobody told it the destination should be warm. As a rule, the booking is attributed to you. The tour operator does not have to accept that you actually wanted to go somewhere else. In procurement, where an agent orders raw materials or supplies, the same logic applies with much higher amounts.
Whether you can rescind such a declaration is disputed. One view treats it as a mistake in the declaration that justifies rescission. The opposing view says that anyone who gives an agent autonomy without clear instructions must accept the result. Even a successful rescission is not free, because you are then liable for the other party's reliance damages under Section 122 of the German Civil Code (BGB).
Unfair Competition Law: The Hamm Ruling
The clearest ruling so far comes from unfair competition law. On the website of the cosmetic surgery clinic Aesthetify, an AI chatbot answered questions about the treating physicians and listed specialist titles the two doctors were not entitled to use, some of which do not exist at all. The consumer protection association of North Rhine-Westphalia sued for an injunction.
On May 12, 2026, the Higher Regional Court of Hamm ruled in its favor (case no. I-4 UKl 3/25). The chatbot's statements are attributable to the clinic as its own commercial conduct, even though the clinic had fed it correct data. After receiving a cease-and-desist letter, the clinic used a prompt instruction and a keyword filter to stop the bot from mentioning "specialists," and the court saw this as proof that it could have controlled the bot all along. Because the case raises new legal questions, the court allowed an appeal to the Federal Court of Justice.
Tort and Product Liability
Under Section 823(2) BGB, anyone who violates a protective statute is liable for the resulting damage. The EU AI Act is predominantly classified as such a protective statute in the legal debate. Violations of it can therefore lead directly to damages claims.
On top of that comes the reformed product liability regime. The German Bundestag adopted the Act to Modernize Product Liability Law on October 8, 2026; it implements the EU Product Liability Directive (EU) 2024/2853. For products placed on the market from December 9, 2026, software explicitly counts as a product, which was highly contested under the old law. The act also presumes the defect or the causal link in certain cases, so the manufacturer has to prove otherwise.
Management Is Liable for Missing Governance
Managing directors must set up the company so that it complies with the law. If agents run without rules, that is an organizational failure. Then there is the decision itself: if an agent proposes something, management has to review the proposal and make the decision itself. Christian Kuß sums it up: simply saying "that was my agent, I'm not answerable for it" does not work at the moment. You may know similar organizational and oversight duties from NIS2 liability for managing directors.
Data Protection and Security When Using AI Agents
AI agents raise the same data protection questions as any large language model: legal basis, transparency, personal data in models. Two points are specific to agents.
The first is the memory function. Many agents store recurring information in a text file so they can reuse it next time. If that file stores too much about customers, applicants, or colleagues, it holds personal data you may not be allowed to process at all.
The second is accountability along the chain. If you don't run the model yourself but use a public language model, the provider is usually your processor, while you remain the controller. Every additional tool the agent connects through a cloud service makes this chain longer. How, and on what basis, personal data flows along it belongs in your record of processing activities, as the GDPR requires.
Security Limits Belong outside the Agent
The mistake most teams make here: they write prohibitions into the system prompt. An instruction such as "never do X" is generally not followed consistently, because the agent learns from interacting with users that it isn't meant entirely seriously.
In concrete terms: restrict permissions at the API instead of asking for restraint in the prompt. The convenience option "always allow" also deserves attention. If you click "always allow" on every permission request, you get asked less often, and you lose exactly the control over what the agent accesses.
The Rule of Two as a Rule of Thumb
In February 2026, the Spanish data protection authority AEPD became the first European supervisory authority to publish detailed guidance on agentic AI. In it, the AEPD takes up the so-called Rule of Two. Three properties should never come together at the same time:
- Access to untrusted information, such as arbitrary websites or incoming email
- Unrestricted access to sensitive data
- The ability to take actions with real-world effects without approval
An agent with all three properties is convenient, because nobody has to click anything anymore. But each additional property raises the risk that data leaks or the company credit card gets charged. So always restrict at least one of the three.
Legal Framework: EU AI Act, ISO 42001, and ISO 27001
The EU AI Act is in force, but its obligations apply in stages, and the Digital Omnibus (Regulation (EU) 2026/1744, in force since July 27, 2026) has postponed some of them. The transparency obligations under Article 50 have applied since August 2, 2026; for generated content from systems already on the market before that date, a transition period runs until December 2, 2026. The obligations for high-risk systems under Annex III apply from December 2, 2027, and those for products under Annex I from August 2, 2028.
Whether an agent falls under the high-risk rules depends on its purpose, not its technology. An agent that pre-sorts job applications quickly lands in Annex III; an agent that books travel generally does not.
In day-to-day work, other rulebooks often matter more for agents: the GDPR, NIS2 for companies in its scope, the Cyber Resilience Act, and the new product liability law. In all of them, responsibility, breach of duty, and protective statutes are becoming more concrete.
Two established standards help with implementation. ISO 42001 applies the logic of a management system to AI, and customers and partners increasingly ask for it; companies can get certified against it. ISO 27001 remains an independent framework for handling information, service providers, and systems securely. How ISO 42001 relates to legal obligations is explained in EU AI Act vs. ISO 42001.
The AI Agent Register: The Minimum Information You Need
The most effective immediate measure is a register of all AI agents. Without a register there is no control, and without documentation management is in a weak position if liability ever comes up. The ISO standards do not prescribe legal minimum fields, but five fields have proven themselves as a minimum.
The register becomes a routine when every new agent goes through the same steps:
- Capture. The business unit reports the planned agent with its purpose and tools.
- Inventory. The agent gets its entry with the five minimum fields.
- Classify. Data protection and information security assess the information processed and check the Rule of Two.
- Set limits. Permissions are granted technically at the interface; where needed, approval by a person is built in.
- Monitor. Changes to tools, models, or permissions are tracked on an ongoing basis.
Suppose a machine builder with 180 employees introduces the register and, in the first inventory, finds seven more agents that teams set up themselves, alongside the two officially approved assistants. Three of them have write access to the ERP system. After the review, two agents keep their write access, now with approval by a person; the rest are limited to read access. When a second agent with a similar purpose comes along, the review takes a fraction of the time, because the path is already documented.
A separate AI inventory is not strictly necessary for this. It often makes more sense to add agents to the existing asset register of your ISMS under ISO 27001 instead of building a separate register and separate processes for every requirement. What matters is that you can filter and export extracts cleanly when an authority or auditor wants to see only part of it.
Managing AI Agents in Business: How Kertos Helps
Kertos uses AI and agents itself and is certified to ISO 42001. In the AI Inventory of the Kertos platform, you record your AI systems and agents in one place, link them to data sources and assets, and connect them to your ISMS and your record of processing activities. More than 300 integrations keep the overview of systems and service providers current, so new tools don't surface only at the next audit.
You document AI literacy training with completion tracking for each person. When you implement ISO 42001, ISO 27001, the GDPR, and the EU AI Act, certified experts from Kertos support you, from the first register to the audit. The agent register then becomes a process that gets faster with every new agent. In a demo, we'll show you what this looks like for your company.
Frequently Asked Questions
Is an AI agent a legal person?
No. The idea of a separate legal personality for AI was discussed and rejected, partly because an agent has no assets of its own with which it could be liable. Liability is therefore always tied to the person or company that deploys the agent.
Is the company liable if an employee deploys an AI agent on their own?
Toward third parties, generally yes. If an employee sets up an agent on their own initiative and it causes damage to others, the company cannot fall back on saying it was only the employee. Whether the employee is liable to the employer internally is a separate question, and under German law the employee benefits from limited employee liability.
Can liability for AI agents be excluded in terms and conditions?
Only within narrow limits. A clause has to survive the statutory review of standard terms, and a blanket exclusion along the lines of "we are not liable for what our AI says" falls short. In the Hamm case, the clinic's argument that it had supplied the chatbot with correct data did not help either.
Does an AI agent have to be listed in the record of processing activities?
That depends on how your record is structured. If it is organized by IT systems, the agent is a separate entry; if it is organized by business processes, the agent appears as a technical means within the relevant process. For most companies, structuring by business process is the easier route.
Do you need a separate AI inventory?
No, agents can be added to an existing asset register. What matters is the five minimum fields per agent and the ability to export filtered extracts, for example for a supervisory authority.






