AI, automation and human controlBusiness owners

Turn bookkeeping uncertainty into one answerable case

A DeinHans exception workflow from uncertainty through answer and reviewed resolution.

Summary

An AI-assisted bookkeeping workflow should be judged by what it does when it is uncertain, not only when the case is easy. Uncertainty can come from missing evidence, poor document quality, conflicting sources, a new counterparty, an unfamiliar transaction or several equally plausible matches. The safe response is to preserve the ambiguity, explain what is known, request the decisive fact and route the case to someone with authority. A forced answer makes the queue look complete while transferring hidden risk into bookings, VAT, payments or close.

Author
DeinHans Team
DeinHans Editorial Team
Reviewed and approved by
DeinHans Team
Editorially reviewed and approved for publication
Updated
21 July 2026
Published: 22 July 2026
5 min read
Troubleshooting
21 July 2026
Open table of contents

Key takeaways

  • Treat “cannot determine” as a valid controlled outcome.
  • Distinguish missing evidence, conflict, novelty and professional judgment.
  • Ask the smallest question that can resolve the case.
  • Learn from reviewed exceptions without turning one answer into a universal rule.

Four uncertainty classes

ClassExampleBest next action
MissingInvoice absent for bank paymentRequest current invoice and purpose
ConflictingInvoice total differs from payout detailReconcile sources before proposal
AmbiguousTwo invoices plausibly match one paymentShow candidates and request allocation
JudgmentCross-border service facts are complete but tax treatment is complexRoute to qualified reviewer

Giving every class the same “low confidence” label is not enough. The user needs to know which evidence or authority resolves it.

Build an exception packet

An exception should carry the transaction, linked documents, observed conflict, attempted checks, unanswered question, materiality context, responsible person and due date. If a model produced a suggestion, keep it as a proposal—not as the source of truth—and show the inputs that supported it.

Worked example: multiple plausible matches

A €2,975 bank debit could settle invoice A for €1,200 and invoice B for €1,775, or invoice C for €2,975. All are from the same supplier and open. Amount equality alone cannot decide. The workflow displays both allocation candidates, their invoice dates and references, then asks for remittance advice. If the advice confirms A+B, the approved allocation and evidence are recorded. The system does not silently prefer the single-invoice answer.

In DeinHans, the useful outcome is a bounded case rather than a low-confidence label. The bank row, three invoice candidates and their references stay together. The owner sees a short factual question—“Which invoices did the €2,975 transfer settle? Please attach the remittance advice if available.” The answer returns to that same case; the proposed allocation is recalculated; and the bookkeeper or accountant reviews it before the match is treated as resolved.

A second case: a new payout row

A previously known payout format adds a row called Deferred adjustment for €420. The net payout still reconciles mathematically, but the row's business meaning is not proven. DeinHans can retain the full payout, recognise the known components and isolate the new row. It should not inherit the treatment of a similarly named fee. The row remains an explicit shared-template question until provider evidence or an authorised reviewer establishes what it represents. Other independently supported components can continue through preparation without hiding the unresolved €420.

Exception lifecycle

  1. Detect and classify the uncertainty.
  2. Protect the case from automatic finalization.
  3. Assemble the relevant evidence and candidate explanations.
  4. Ask one precise factual or professional question.
  5. Record the answer, source and authority.
  6. Apply the decision and re-run affected controls.
  7. Monitor recurrence and improve the upstream process.

What not to learn automatically

One reviewer’s answer may depend on contract, period, materiality or local facts. Do not generalize it across all vendors because the label looks similar. A reusable treatment needs stated conditions, evidence requirements, owner, review date and a trigger for reassessment. Otherwise “learning” can automate yesterday’s exception into tomorrow’s error.

Important: A probability or confidence number does not measure legal correctness or document authenticity. Keep professional and authorization boundaries independent of model confidence.

What DeinHans can support

DeinHans can keep the exception, question, evidence and reviewer decision connected rather than dispersing them across inboxes. This supports a visible path from uncertainty to resolution. It does not remove the need for the business to supply facts or for the accountant to make qualified decisions.

Prioritize without hiding

Give exceptions a business impact, due date and resolution authority. A missing €20 receipt may be low impact but high volume; an ambiguous bank-detail change may be urgent despite a familiar supplier. Do not use model confidence as the only priority. Combine consequence, deadline, evidence gap and reversibility under a policy approved by the responsible team.

Keep deferred cases visible with a reason. “Not now” must identify who accepted the delay and when it will be reviewed. Otherwise a low-priority queue becomes a hidden write-off of unresolved evidence.

Measure resolution quality

Track time by uncertainty class, questions sent, response completeness, reviewer changes and recurrence. A quick answer is not high quality if it is later reversed. Sample resolved cases against the source and ask whether the decisive fact was truly obtained.

When one class repeats, improve the source process: require a better invoice capture, add remittance advice, clarify card policy or fix a report boundary. The goal is fewer avoidable exceptions, while preserving a safe route for genuinely novel or judgmental cases.

Sources

Sources were checked on 21 July 2026. This article provides orientation and is not tax, legal, or accounting advice.

  1. EUR-Lex – Regulation (EU) 2024/1689 on artificial intelligence
  2. German Federal Office for Information Security – artificial intelligence

Related resources

See how DeinHans supports this workflow

Explore the product flow without confusing automation with professional approval.

View product
Back to resources