Skip to content
Gathena
da

Learn · Explained

Who is accountable when an AI agent makes an entry?

Published 31 October 2026

Short answer. The company is. Management stays responsible for the bookkeeping and the controls around it, whether a person or a program made the entry. Neither the AI nor its vendor takes that over. In daily work, the person who answers for the agent’s entries is the operator: the person in the company who checks the agent’s work and signs it off.

This guide explains what the law says, where the books must be kept, who answers for what, and how the books show who or what made each entry.

Why the AI and its vendor cannot carry it

An AI is a machine. It cannot be held accountable, so the responsibility stays with people. When a company hands work to an AI agent, the responsibility stays with the company: the entries are still the company’s entries, in the company’s books.

That is why every agent has an operator. The operator gives the instructions, reviews the work, makes the judgement calls and signs off.

What the law says

Area What it says Source
Bookkeeping and controls Management remains responsible for satisfactory bookkeeping and controls, whether a person or a program made an entry. Danish Bookkeeping Act; Companies Act §§115 and 118
The bookkeeping system A company covered by the digital bookkeeping requirement keeps its books in a system registered with the Danish Business Authority, or in a system it makes sure meets the requirements itself. Bookkeeping Act §16; order BEK 205/2024
Who booked it Each booked transaction carries the initials or similar of the person or the program that booked it. Order BEK 205/2024 §3(2)(3)
Documents The documentation for each booked purchase or sale is kept in the digital bookkeeping system. Order BEK 1383/2023 §1
Procedures A company that files an annual report describes how it makes sure all transactions are registered and its records are kept safely, and who is responsible. The description is kept up to date. Bookkeeping Act §6
AI literacy Since February 2025, the EU AI Act requires businesses that use AI to take measures to support their staff’s AI literacy. EU AI Act Article 4, as amended by Regulation (EU) 2026/1744
Decisions about people When a decision is made solely by automation and significantly affects a person, GDPR requires meaningful human involvement. GDPR Article 22

This is a summary, not legal advice. For the AI Act in full, read What does the EU AI Act ask of a company that uses AI?

Where the books are kept when an AI drafts the entries

In Denmark, most companies must keep their books in a digital bookkeeping system: either one registered with the Danish Business Authority, such as those on its public list, or one the company makes sure meets the requirements itself. A bookkeeping system is the software that books the transactions and stores them with their documents. An AI system that only reads from it, or delivers data into it, is not a bookkeeping system itself.

What matters is where the books end up:

Setup Is the AI system a bookkeeping system?
The AI system reads past entries, learns the company’s rules, and creates draft entries in the registered system. A person, or the registered system itself, posts them. No. It delivers data into the bookkeeping system. The drafts become booked transactions only when they are posted in the registered system.
The AI system posts entries into the registered system through the system’s own interface, with their documents. No. It still delivers data, and the registered system records the program that booked each entry.
A data warehouse, such as Microsoft Fabric, receives a copy of the books for reporting. No. It reads a copy; the books stay in the registered system.
The AI system books entries, or keeps entries or documents, only in its own database. Yes, as part of the whole. The setup counts as an unregistered bookkeeping system, and the company itself must make sure it meets every requirement: the fields, the audit trail, backups, e-invoicing and the rest.

The rule of thumb: the books and their documents live in the registered system. Anything else may read from it or deliver into it, but must never become the place where the books are kept, or hold the only copy of a document.

What to do when an AI drafts your entries

The bookkeeping system stays the same, but the way the books are kept changes. That leaves the company five things to do:

  1. Update the procedure description. Say how entries are now drafted, who approves and posts them, and who owns the procedure. The law asks for this description to be kept up to date.
  2. Keep the document with the entry. If the AI system receives an invoice first, it delivers the invoice into the registered system with the draft, and the entry refers to it.
  3. Keep a control before posting. The approval step, the “asks first” zone, is the internal control that catches errors.
  4. Version the rules the system learns. Our advice, not a legal requirement: keep each version of the rules, so the company can explain why an entry was drafted the way it was.
  5. Handle personal data. Entries often contain personal data, such as sole-trader suppliers or employee expenses. Analysing them is processing under GDPR, so the company needs a legal basis and a data processing agreement with the provider.

Who answers for what

Role What they answer for
Management That the bookkeeping and the controls are satisfactory, including the work an agent does
The operator The steps they approve and sign off, and which steps the agent may take on its own
The employees who work with the agent Using it within the rules, and raising what looks wrong
Whoever builds or supplies the agent What their agreement with the company says; the entries in the books stay the company’s

Four things that carry the responsibility

  • Access the company controls. The agent can read and do only what its job needs, and the company can take that access back at any time.
  • Guardrails that hold. Rules and limits that a control outside the agent enforces, such as a maximum amount or accounts it may not post to. Read What are guardrails for an AI agent, and who enforces them?
  • People who understand the system. The operator and the employees know what the agent does, where it can go wrong and how to correct it.
  • A logbook of what was done. A record of every step, its source and its approval, that an auditor or the board can read.

Every entry shows who or what made it

For each entry, the books and the logbook should show:

  1. Who or what made it. The law asks for the initials or similar of the person or program that booked each transaction. The agent therefore works under its own user in the bookkeeping system, never under an employee’s login, so its entries are marked as the agent’s.
  2. The source. The invoice, receipt or file the entry is based on.
  3. The rule. Which rule or company dictionary entry the agent applied.
  4. The approval. Who approved it, or that it ran in “acts alone” under a rule the operator set.
  5. The time. When it was proposed, approved and posted.

With these five, anyone can trace an entry back to its source and to the person who answered for it.

Approving every entry is rarely required

The law rarely requires a person to approve every single output. The four things above carry the responsibility. Which steps a person approves is decided step by step with the three sign-off zones, and some steps, such as payments, always go to the operator. Read When may an AI agent act on its own?

Sources

Not legal advice.

Bring the entries an agent would make in your books. We go through who signs off on what, and what the logbook must show.

Book a conversation