The AI review signals accounting firms actually need

Signal versus score

The right-hand examples allow the reviewer to verify, disagree, request evidence, or change the result. The left-hand labels create work without explaining it.

The ten fields of a review-ready signal

1. Object: the exact document, transaction, proposal, payout, or period.
2. Source: original evidence and relevant location or row.
3. Observed fact: what the source actually shows.
4. Inference: what the system suspects and why.
5. Conflict or uncertainty: what does not fit or remains unknown.
6. Possible consequence: which part of preparation could change.
7. Next check: the smallest evidence or question that can resolve it.
8. Suggested action: review, request evidence, hold, change, or dismiss with reason.
9. History: prior signals, owner answers, and proposal changes relevant to the case.
10. Decision record: reviewer identity, action, time, and reason.

Not every signal needs a long form. The design should be compact enough for a queue but complete enough to open the case without reconstructing the reason from scratch.

Synthetic example 1: duplicate evidence, different payment

The example is invented.

Two PDFs from Nordlicht Office GmbH have the same invoice number NL-882, invoice date 3 July, and gross amount €714.00. One was uploaded by email and one with a bank statement. There are two bank debits: €714.00 on 10 July and a reversed €714.00 debit on 11 July.

A weak signal says: “Duplicate invoice.” A useful signal says:

Observed: Documents D-41 and D-58 share supplier, invoice number, date, currency, and amount. Payment context: one debit and one next-day reversal exist. Uncertainty: the documents may be duplicate evidence for one purchase, while the two bank rows may be a failed payment cycle. Next check: compare file fingerprint and reversal reference; do not delete or post a second expense until the relationship is confirmed.

The signal separates document duplication from payment duplication. It does not assume that two bank rows mean two purchases.

Synthetic example 2: VAT ambiguity without a tax guess

An invoice for €1,190.00 shows a foreign supplier, a German VAT amount, and a German delivery address. The customer VAT ID field is blank, while the description refers to a digital service.

The safe signal records the contradictory facts, requests supplier-establishment and service/customer-status evidence, and routes the case for professional review. It does not decide “reverse charge” or “foreign VAT” from a country name, and it does not expose an internal tax rule or threshold.

The purpose of a signal is to preserve the decision boundary, not to disguise a tax conclusion as an alert.

How to order a review queue

Priority should reflect consequence and workflow dependency, not a secret numerical threshold. Useful public factors include:

• whether the source affects money already settled;
• whether downstream proposals or close readiness depend on it;
• whether the same unresolved fact affects many child rows, as in a payout;
• whether the evidence conflicts with an approved or recurring pattern;
• whether a deadline or period lock makes delay consequential;
• whether the case can be resolved by the owner, bookkeeper, or only a qualified reviewer.

The firm should define and document its own policy. The article does not prescribe materiality or professional sampling rules.

Reviewer actions need meaning

“Dismiss” is not enough. The reviewer should be able to:

• confirm that no change is required and state why;
• change the prepared fact or proposal with evidence;
• ask the owner or bookkeeper a bounded question;
• hold the case pending an external document;
• return the workflow to an earlier step when source truth changed;
• escalate according to the firm's engagement and authority.

The decision record must distinguish the review of a signal from approval of a booking or filing. Those acts can have different reviewers and permissions.

What DeinHans prepares

In supported workflows, DeinHans can connect an exception to source evidence, show why the prepared result needs attention, formulate a targeted next question, and carry that context into a booking-proposal or accountant-review case. This can reduce the time spent rediscovering why something is in the queue.

The product boundary remains explicit. Visible signals vary by workflow; a signal can be wrong or incomplete; and DeinHans does not publish private flag names, score bands, prompts, mappings, or routing thresholds. The accountant reviews the evidence and decides the professional consequence.

Firm implementation checklist

• Every signal links to the exact source object.
• Observed fact and inference are visually distinct.
• The affected field and possible downstream consequence are named.
• The next check is specific and proportionate.
• The reviewer can view evidence without leaving the decision context.
• Action and reason are attributable.
• Material changes invalidate dependent approval.
• Dismissed signals remain traceable without cluttering the active queue.
• Queue policy is documented and does not pretend that confidence equals materiality.
• Professional and filing permissions remain separate from signal review.

When these conditions hold, AI review signals become a map of meaningful exceptions rather than a second opaque inbox.