Use AI well in bookkeeping—with clear control points

A practical task-suitability matrix

The EU AI Act and BSI guidance provide useful control ideas such as transparency, human oversight, competence, security, and known limitations. This article does not classify a bookkeeping tool under the Act or provide compliance advice. It applies those ideas as operating discipline.

The four layers of a safe result

1. Source fact

What is directly visible in the document, report, bank statement, or verified master data? Keep the original source and point to the relevant field or row.

2. System inference

What did AI or another rule infer: supplier identity, document type, possible relationship, likely duplicate, or suggested question? Label it as prepared context, not as a new fact.

3. Business fact

What only the owner or responsible operator can explain: purpose, participants, private use, contractual relationship, cash handling, or whether two brands are the same legal supplier?

4. Professional decision

What requires accounting, tax, legal, or engagement authority: treatment, approval, period correction, export, signature, or filing?

Many failures happen when one layer is presented as another. A supplier name extracted correctly does not prove the service purpose. A recurring pattern does not prove the current invoice belongs to the same treatment.

Synthetic example: same supplier, different purchase

The example is invented.

For eleven months, Elbwinkel Beratung GmbH receives a €119 software invoice from Cloud North Ltd. In June, a new invoice from the same supplier is for €1,190 and describes a two-day staff training.

AI can safely:

• extract supplier, date, amount, currency, and description;
• show that the amount and description differ from the recurring pattern;
• link the original invoice and relevant payment candidate;
• ask whether the purchase was training and whether a separate software invoice exists;
• prepare the answered case for review.

AI should not:

• copy the previous category merely because the supplier matches;
• invent the participants or business purpose;
• decide tax treatment from the word “training” alone;
• approve or post the proposal without the required reviewer.

The valuable output is not “high confidence”. It is a compact evidence package: source invoice, detected change, owner answer, payment context, proposed result, and explicit review reason.

Designing useful control points

At intake

Verify that the file belongs to the business and period, preserve the original, detect duplicates, and route uncertain document types visibly.

After extraction

Show the extracted value beside the source. Conflicts between subtotal, tax, and total or between document and bank should create review work rather than being normalised away.

Before matching

Require compatible identity, amount, currency, timing, and direction. A candidate can be useful without being accepted automatically.

Before a booking proposal

Collect missing owner facts, keep known exceptions attached, and ensure the proposed result can be traced to source and payment context. A proposal is a review unit, not a professional conclusion.

Before approval or downstream action

The reviewer needs authority, relevant competence, source access, the proposed effect, changes since prior review, and a way to change, reject, or return the work. Approval that cannot be challenged is not meaningful control.

What DeinHans uses AI for

DeinHans can use AI within controlled workflows to help read and organise evidence, prepare comparisons, formulate bounded questions, and assemble reviewable cases and proposals. The visible outcome should let an owner, bookkeeper, or accountant see the source and the remaining decision instead of receiving an unexplained answer.

Boundaries belong next to the capability: not every source or exception is supported equally; AI output can be wrong; product signals do not create legal certainty; and DeinHans does not autonomously approve, sign, file, or replace the accountable professional. Confidential implementation details are not part of the public control model.

Privacy and security questions to ask

Before using any AI workflow with bookkeeping evidence, establish:

• which data is sent to which service and for what purpose;
• who can access the result and source;
• what retention, deletion, and contractual controls apply;
• how incorrect output is corrected and prior approval invalidated;
• whether sensitive customer, employee, bank, or tax information is necessary for the task;
• how incidents, unauthorised access, and provider changes are handled.

The right answer depends on the system and organisation. Do not assume that “AI-assisted” says anything by itself about security or compliance.

Responsibility split

A release test for any AI-assisted step

Before accepting the step, ask:

1. Can the reviewer see the original evidence?
2. Is the system's inference separated from verified fact?
3. Is missing context expressed as a question or exception?
4. Can a qualified person change or reject the result?
5. Is the reviewer and action attributable?
6. Will material source changes invalidate the approval?
7. Does downstream action require the appropriate authority?

If one answer is no, automation has probably moved faster than control. Fix the workflow before expanding the AI role.