How AI Supports Our Accounting
In our accounting, AI assigns documents and reports uncertainty. External payments and unclear cases require individual approval.

A quick proposal is useful in our own accounting. A quick, incorrect execution is not. The AI assigns documents, checks cases, and proposes next steps. External payments and unclear cases each require human approval. Tightly scoped rules or documented approvals cover clearly defined internal transfers and preparatory assignments.
Which tasks does the AI handle?
Incoming documents are assigned to a case. Fields and relationships are checked for contradictions. If details are missing or the content does not fit together, the workflow flags the case for review. If the case is coherent, it creates either a proposal or a tightly scoped assignment, depending on the applicable rule.
For external payments and unclear cases, the proposal has no effect until a person approves it. For clearly defined internal transfers and preparatory assignments, the workflow may act only within documented rules. It brings the available evidence together, identifies open questions, and shows the expected effect. The decision can be reviewed without repeating the preparation.
Why does uncertainty matter more than an amount threshold?
An amount threshold assesses only one number. The real risk may sit elsewhere. An assignment may be contradictory, a period may not fit, or the same case may appear again in a different form. These cases can require review regardless of their financial size.
We therefore treat uncertainty as a signal of its own. If characteristics conflict, a necessary relationship between documents is missing, or the case falls outside tested patterns, automation stops before any effect. An amount threshold can still apply as an additional rule. It does not replace a review of the substance.
How does required human approval work?
When approval is required, it applies to one visible proposal and one action. External payments and unclear cases are presented individually. Approval does not carry over to similar cases later. Documented standing approvals may apply to tightly scoped assignments.
The reviewer sees the intended step, the information needed for the decision, and any known uncertainty. Rejecting, postponing, and requesting focused corrections remain valid outcomes. This prevents approval from becoming a quick confirmation of a decision the automation has effectively made already.
Why do dry run and readback belong together?
The automated assignment paths start in dry-run mode. They calculate the intended change without writing anything to the target system. A write operation must be enabled explicitly. The reviewer can therefore compare the proposed action with its expected effect.
When an action is approved or permitted by a tightly scoped rule, the workflow executes only that action. It then reads the new state back from the target system. This readback is compared with the expected result. A successful request alone is not enough because a target system may process input differently or apply only part of it.
What must the journal record?
The journal connects the proposal, uncertainty, decision, execution, and readback in one traceable history. It records which rule applied, whether an approval or a documented rule permitted execution, and which result the target system returned. Personal and business data belongs in the journal only when it is required for that evidence.
A hash can show that two files are technically identical. It is not a complete duplicate check. The same content may produce a different hash after a new scan or a different file representation. Conversely, an identical file may belong to two distinct business cases. Duplicate detection therefore needs several characteristics, normalized content, and the previous process history.
Why can an empty queue be a warning sign?
“Nothing pending” sounds good. The queue may also be empty because the earlier document intake failed. Processing then shows no visible backlog even though cases are stuck before they reach the queue.
We therefore inspect the path leading to the queue too. When did the last input arrive? Was the source read successfully? Are there rejected or unassigned cases? Only these signals distinguish a quiet day from a silent upstream failure.
What guidance does the BSI provide for this setup?
The paper on generative AI models (opens in a new tab) from Germany’s Federal Office for Information Security (BSI) helps teams assess technical and organizational risks. Controlled data paths, verifiable results, and clearly assigned responsibility are especially relevant to our setup.
The paper and this field report do not replace a legal or professional review of a specific case. The workflow must fit the organization, its responsibilities, and the systems involved.
If you want to design an accounting workflow with a dry run, clear approval rules, readback, and a journal, talk to us. Together, we can define what AI may prepare, what tightly scoped rules may permit, and where the process must stop for a person to decide.
What Is the Next Step in Your Company?
Our page AI Consulting and AI Adoption explains how an AI rollout works, what drives the cost and how data protection and the works council come in. Get in touch if you want to plan the next step concretely.
Over 20 years of experience with development, SharePoint, and Microsoft 365. Sascha advises companies on implementing and optimizing digital workplaces – from strategy to technical implementation.
Learn more about the author →Contents
Related Content
How Do You Choose Your First AI Use Case?
22.09.2026AI Colleagues After Probation: What Really Works Today
29.09.2026Brilliant yet Clueless – The User Manual: Prompt Engineering, Model Selection, and How to Use AI Properly
26.03.2026Brilliant yet Clueless – The Fails: When the Gifted Agent Gets Spectacularly Dumb
12.03.2026