Agentic accounting
Agentic accounting describes the use of AI agents that carry out accounting tasks across several steps and largely on their own – for example researching a reconciliation difference, formulating a query and producing a correction proposal. The difference from classical automation is that the agent decides the route to the goal itself.
What makes the difference
Classical automation executes a predefined sequence: if A, then B. An agent is given a goal and chooses the steps itself – it can retrieve data, assess intermediate results, request further information and use several tools in sequence. That makes it usable for tasks whose sequence cannot be fully described in advance.
Where the approach holds today
Clearing reconciliation differences. Searching open items, matching several part payments to one credit note, naming the likely cause of a difference. Classical rules find the difference; the search for its cause has so far been manual work.
Preparing queries. Identifying missing documents, determining the responsible colleague, formulating a query with the right context.
Assembling documentation. Collecting and completely mapping documents for a tax audit or a due diligence request.
Close preparation. Working through checklists, checking completeness, naming open points.
Where the limit runs
The limit is not technical but a question of responsibility. An agent can propose a posting and justify it; it cannot stand behind it. Management is liable for the books, the annual accounts are signed by people, and an auditor examines the posting, not the agent.
There is also a practical problem: agents are persuasive even when they are wrong. A plausibly worded but incorrect explanation for a variance is harder to spot than an obviously missing one – particularly in investor or bank reporting.
The controls that belong with it
- Limited rights. Read access as the default, write access only with approval. No autonomous payments, no master-data changes.
- A traceable trail. Every step logged: which data, which intermediate result, which proposal.
- Amount thresholds and exception rules analogous to human approval routes.
- Process documentation. The GoBD require the procedure to be described – that applies unchanged to agent-based workflows and is harder there, because the sequence is not fixed.
How to place the term
The market currently uses the term broadly; much of what is labelled this way is rule-based automation with a natural-language interface. The useful test question: can the sequence be described completely as a rule set? If so, a rule is the better and more auditable solution. Agents pay off where the route to the goal is open – and precisely there they need the tightest controls.