Auditors look for these things with samples. Most finance teams never look systematically at all.

A purchase invoice posted twice. A round amount booked by hand on a Sunday. An invoice that is many times what the vendor usually charges. Each of them is easy to spot once someone points at it. The trouble is that nobody points. Day to day, the team is busy closing the month, and the sample comes once a year.

So I asked Business Central: “Anything unusual?”

Why rules, not a black box

There are AI products that promise to “detect anomalies” with a model nobody can explain. In finance that is a hard sell. If a tool flags a posting, the controller has to be able to say why. If it misses one, the auditor will ask what it was checking.

So I went the other way. Simple rules, written in AL, that anyone in finance can read and challenge. The AI comes in afterwards, to connect and explain what the rules found.

The six rules

The extension checks the last twelve months:

  • Same vendor, same amount, within ten days. The classic duplicate.
  • Invoices far above that vendor’s usual amount. Compared with the vendor’s own history, not a global threshold.
  • Manual postings at the weekend. Not wrong in itself, but worth a look.
  • Round manual amounts on expense and income accounts. Real invoices rarely end in three zeros.
  • Manual postings on revenue accounts. Revenue should come from sales documents.
  • Outliers per G/L account. Each posting is compared with the mean and spread of the *other* postings on the same account.

The last rule needs a word. If you compare a posting with an average that includes the posting itself, one extreme value inflates the average and the spread, and partly hides itself. Leaving the posting out of its own comparison (a leave-one-out check) stops that. It is a small detail. It is the difference between catching the outlier and explaining it away.

How it works

Same principle as the whole “Hack Business Central” series: AL calculates, the AI explains.

  1. The AL extension runs the six rules on request and returns every hit with the rule, the posting, the amount and the reason. It writes nothing.
  2. Business Central’s MCP server exposes the result as a read-only tool for an AI assistant.
  3. The assistant reads the hits, links those that belong together, ranks them by urgency and writes a short summary.

The result in the demo

The demo runs on Microsoft’s CRONUS demo company with an invented story planted in it. No client data.

The rules found nine hits. All nine come from the planted story. There were zero false alarms in the regular CRONUS data. The assistant summed up:

Three points are worth clearing up this week. First: there are two invoices from First Up Consultants for €14,976 each, posted on 15 and 18 September, with almost identical invoice numbers (FUC-2026-0915 and FUC-2026-0915-1). This is very likely a duplicate payment. Second: an invoice from Graphic Design Institute for €48,000 is 36 times that vendor’s average. Third: on Sunday 13 September, €15,000 was posted manually as “Consulting lump sum (no invoice)”. Round amount, no document, weekend.

What I like about this answer:

  • It prioritises. Nine hits become three things to do this week.
  • It connects. The duplicate vendor shows up in two rules. The assistant reports it once, as one case.
  • It invents nothing. Every sentence traces back to a rule and a posting. If a rule did not fire, the assistant has nothing to say about it.

The €15,000 Sunday posting is a good example of why combining rules matters. Weekend, round amount, manual, no document. Each signal alone is weak. Together they are a question someone should answer.

Limits, honestly

  • Rules find what they are written for. A clever fraud that avoids all six patterns will not show up. This is a daily safety net, not a forensic audit.
  • Zero false alarms is a demo result. Real data has seasonal lump sums, legitimate weekend batches and vendors with irregular invoices. In a real company the thresholds need tuning, and some hits will be explainable.
  • A hit is a question, not a verdict. Whether an invoice is reversed or a payment stopped is decided by a person. The configuration is read-only.
  • The language model can be wrong. That is why it only reads rule results. It does not decide what counts as unusual.
  • Business Central online only. The MCP server does not exist on-premises.

Try it

The anomaly scan and the AI commentary are in the interactive demo: Open the Management Cockpit for Business Central. Demo data only.

The six rules, the architecture and all limits are described in the white paper: Ask Business Central: three AI use cases that work today.

Is your team facing a similar problem?

If your team only finds odd postings when the auditor does, a weekly scan is a good first step. Discuss your challenge. I usually reply within two working days.