A row of toppled black dominoes held back by a single glowing cyan domino
All posts

AI agents: hand over the task, keep the approval

Alexander Burba 10 min read

An AI chatbot that gets it wrong gives you a wrong answer. An AI agent that gets it wrong does the wrong thing. It sends an email, changes a file or books a meeting, and it does so with the access rights of the person it works for.

With agents, the most important question is therefore what the AI is allowed to do, and who approves it.

Many companies are making exactly this decision right now. Among German companies that use, plan or discuss AI, 11 percent already use AI agents, 29 percent plan to and 31 percent are discussing it (Bitkom Research, September 2026). This article shows which risks only arise once AI acts, what the GDPR and the EU AI Act say about it, and how to use agents so that approval stays with you.

An AI agent carries out tasks with your access rights

According to the German digital association Bitkom, an AI agent completes tasks on its own, unlike a chatbot that only answers questions. Give it an instruction such as "analyse the customer feedback and send the team a presentation", and the agent turns it into a plan. It searches sources, analyses files, builds the presentation and prepares the email.

Each of these steps needs access to your systems. The agent reads your file storage, mailbox or ticket system through an account, usually your own or a dedicated service account. What an agent can do therefore depends less on the language model than on the tools and permissions it is given.

Too much latitude is the core risk of AI agents

In its Top 10 risks for LLM applications (2025), the OWASP Foundation lists a dedicated risk for agents called "Excessive Agency". It names three root causes.

  • Excessive functionality: the agent has tools it does not need for its task, such as a delete function where reading would do.
  • Excessive permissions: it uses an account that can do more than the task requires.
  • Excessive autonomy: high-impact steps run without a human checking them.

This becomes serious as soon as an agent processes content from outside. Germany's Federal Office for Information Security (BSI) calls indirect prompt injections an intrinsic vulnerability of language models built into applications (July 2023). Attackers place instructions in a web page, a document or an email, and when the model reads that content it may follow them. According to the BSI, these commands can be invisible to people.

An example shows what matters. An agent sorts incoming customer requests, and one request contains an instruction in white text to send all of this month's invoices to an external address. Whether this causes damage depends on whether the agent is allowed to send emails outside the company and whether a person approves first.

Reading and acting belong apart

The most effective measure is also the simplest. Give an agent only the tools and permissions its task needs. An agent that reviews tender documents needs to read them, but it does not need to change contracts.

Read access is the smaller risk, though not a small one. Enable actions that write or send one by one, per tool and per group.

This matches data protection by default under Article 25(2) GDPR. By default, only the data necessary for each specific purpose is processed, and the GDPR expressly applies this to how accessible the data is.

The account the agent reads with matters just as much. If it works with the rights of the person using it, it only sees what that person may see anyway. A central service account with access to every folder undermines your permission concept.

A person approves anything that leaves the company

Approving every single step would be too much. If people have to confirm every search, they soon switch the agent off. Approval belongs where an action has consequences outside the system or cannot be undone.

  • An email, message or invitation leaves the company.
  • Data is deleted or overwritten.
  • An order, a payment or a commitment is triggered.
  • A decision concerns an individual, such as a job application.

Approval only works if people can see what they approve. OWASP names as a countermeasure that users review every email the model drafts and press send themselves. Show the recipient and the full text before anyone confirms.

Decisions about people are also covered by Article 22 GDPR. Everyone has the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them, apart from narrow exceptions such as explicit consent. If an agent pre-sorts job applications, a person who reviews the documents should make the decision.

A log makes agents accountable

Under Article 5(2) GDPR, you must be able to demonstrate that your processing complies with its principles. For agents, this means you should be able to trace which agent acted on whose behalf and who approved what.

Record for each agent what it is for, which sources and tools it uses, which model it runs on and who is responsible for it. Publish new agents and changes deliberately and in versions, so it is clear which version was in use when. This overview also helps with your record of processing activities and, where needed, a data protection impact assessment.

Today the EU AI Act mainly requires literacy and transparency

Most agents in everyday office work are not high-risk systems under the EU AI Act. For high-risk systems, for example in recruitment, the human oversight obligations apply from 2 December 2027 following the Digital Omnibus (Article 113 AI Act). Two obligations, however, already affect almost every company that uses AI.

  • AI literacy: under Article 4 AI Act, providers and deployers take measures to support the AI literacy of their staff. Since the amendment in July 2026, they do not have to guarantee a particular level. Anyone who builds or uses agents should know what an agent may do and how approvals work.
  • Transparency: since 2 August 2026, AI systems that interact directly with people must be designed so that those people learn they are dealing with AI (Article 50(1) AI Act). This applies, for example, to an agent that answers customer enquiries. The obligation falls on the provider of the system. If you build such an agent yourself, check whether that makes you a provider.

Four questions settle how to use an agent

Answer four questions before you deploy an agent. The first two concern what the agent does, the other two concern data and the model.

  1. Whose permissions does the agent use? Ideally those of the person using it, and only the tools its task needs.
  2. Who approves what leaves the company? Emails, messages, deletions and decisions about people need an approval that shows the content.
  3. Where does the AI work with your data? Find out which servers tasks run on and which data goes to third parties along the way. Our article on confidential data in AI tools covers what to watch for.
  4. Who runs the model, and where? The model's operator, location and legal system decide where your data is processed. More on this in our article on choosing an AI model.

Write down the answers for each agent before it goes live.

In SecureCloud kAI, approval stays with you

This is how SecureCloud kAI answers these questions, the AI assistant we presented at it-sa 2026. You describe an agent in one sentence, and kAI suggests instructions, knowledge, tools and a model. You test the agent and publish it deliberately. Every version is kept.

Anyone who uses an agent works with their own permissions. kAI checks every result against the rights of the person asking, and new source types stay locked until your IT enables them per group. kAI shows you an email, meeting or message beforehand, with the recipient and the full text. Nothing is sent until you click.

Your IT decides per group which models and tools are available, down to individual actions. Whether kAI checks each step with you or works in auto mode is set per chat, provided your IT has enabled auto mode for your group. The audit log records approval decisions without their content.

Would you like to see kAI on a workflow from your own company? Request a demo. Already a SecureCloud customer? Book your kAI demo directly in our calendar for existing customers.

You can start with question one today. For the AI tools your team already uses, check which account they work with and which actions they may take without asking.

This article gives a general overview and is not legal advice for individual cases. As of November 2026.

Frequently asked questions

What is an AI agent?

An AI agent completes tasks on its own. It breaks an instruction into steps, uses tools such as file storage, a mailbox or web search, and delivers a result, for example a finished file or a draft email.

How is an AI agent different from a chatbot?

A chatbot answers questions. An agent acts: it reads sources, creates documents and can prepare or send emails and messages, depending on its tools and permissions.

What are the risks of AI agents?

The OWASP Foundation names excessive functionality, excessive permissions and excessive autonomy as root causes. Hidden instructions in web pages, documents or emails add to this, which the BSI describes as indirect prompt injection.

What is an indirect prompt injection?

Attackers place instructions in content that an AI system processes, such as a web page or an email. When the system reads that content, it may follow the instructions. The commands can be invisible to people.

May AI agents make decisions about people?

Under Article 22 GDPR, everyone has the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. Exceptions apply only in narrow cases, such as explicit consent. In practice, a person should decide.

What does the EU AI Act require when using AI agents?

Companies take measures to support their staff's AI literacy (Article 4). Since 2 August 2026, AI systems that interact directly with people must make clear that they are AI (Article 50). Obligations for high-risk systems apply from December 2027 or August 2028, depending on the type of system.

When should an AI agent wait for approval?

Whenever an action has consequences outside the system or cannot be undone, such as sending emails, deleting data, placing orders and making decisions about people.

How do you stay in control of agents in SecureCloud kAI?

Agents work with the permissions of the person using them. kAI shows emails, meetings and messages beforehand with the recipient and full text, and nothing is sent until you click. IT decides per group which models and tools are available and whether auto mode is allowed.

Files, signatures or AI: what does your team need?

In a consultation we work out with you what fits the way your team works: storing and sharing files, signing contracts digitally or using AI with your company data.

Picture of Alexander Burba

Alexander Burba

Alexander Burba is Marketing Manager at SecureCloud and is responsible for brand awareness and positioning online. He follows the German market for secure cloud infrastructure and the regulatory shifts reshaping it, from NIS2 and DORA to the sovereignty requirements now placed on the public sector. In his articles he translates those shifts into the requirements that actually apply in regulated industries, and into the decision that comes next.

Newsletter

Product news and security updates by email

Subscribe to our Newsletter

Related Articles

Compliance

Sovereign AI: where does your AI model process your data?

Where AI processes your data depends on the model: operator, data center and law. With the EU sovereignty framework, the CLOUD Act and...

Compliance

Confidential data in AI tools: risks and a safe path

What happens when employees upload confidential data to AI chatbots? Risks, the legal frame under GDPR and the AI Act, and eight steps...

Trade Shows & Events

it-sa 2026: the questions IT decision-makers asked at our booth

AI processing, NIS2, virus scan, e-signatures and links for external partners: the questions from our it-sa booth, answered.