Why do ERP, controlling and the board report never reconcile, and how do we get to one figure?

How do you recognise the symptom?

Revenue in the ERP, revenue in the controlling report and revenue in the board deck differ, and every department can explain its figure. Before every meeting someone builds a reconciliation. After two months the reconciliation is itself a case that only one person understands. That is the symptom, and it is common.

In a 2026 mandate a reporting package of over 20,000 values carried four control cells. When the files were linked, 12 items in the July report surfaced that still carried June values, one sheet with January values and four balancing items whose origin nobody could explain any more. No person would have seen that while copying.

Which causes sit behind it?

Three causes explain almost every case I have seen, and they usually appear together. The table sorts them by process, data and tool, with the marker by which you recognise the cause without an analysis. AI appears in no row, because no cause arises from AI and none disappears through AI alone.

The order in the table is also the order of repair. A tool that meets three spellings of the same supplier produces three suppliers. In practice, automation fails on master data, not on technology. That is why cleansing comes before any tool, and why it is the part that takes longest.

What does a good state look like, in numbers?

One data source per figure, and every report refers to it by formula. In the prototype of 2 September 2026, 11 files became one linked workbook: over 250 manual transfers are now formulas, 18 checks run automatically, both balance sheets under HGB and IFRS reconcile to zero. Before that there were four control cells.

A register in which every row carries its document. At PeriGuard each of the 239 auctioned lots sits with invoice chain, settlement stage and document group in a register with 4,153 formulas and 513 links to the document. The document index lists 275 files with a checksum. Every figure leads back to its document with one click.

An archive whose cross-references are checked by machine. At PeriGuard it holds 271 files and 374 cross-references; each was checked, none leads nowhere. An auditor, a tax adviser or a successor in finance can trace every figure of the final statement without asking a person who carries the case in their head.

A good state still has deviations, but each one is known, has a reason and a status. Every figure visibly carries how hard it is: documented, arithmetically confirmed, as stated by others, assumption, proposal, decided, open. The board then sees three figures and knows why.

What does automation contribute?

Tools find duplicates, mispostings and deviations faster than any person, and they hold the state once it is in place. They do not replace the rule for what is correct. AI does not repair a data base by itself. The rule for what is correct has to be set by someone who understands the chart of accounts.

In the reporting prototype the machine read the 11 files, recognised the data flows and generated the links, the check sheet and the log. Which inputs deliberately remain inputs and which balancing items need clarification was financial judgement. The prototype is built and reviewed but not yet in monthly operation.

In the auction the machine read around 600 pages, tied every figure to its document and recalculated every settlement line. What to claim and what to hold back, it did not decide. Negotiation path, clarification meeting and the proposal to the other side were financial judgement, and that judgement requires accounting and VAT knowledge.

What does process mining show in a finance function?

Process mining reads the timestamps from ERP and workflow system and draws from them how invoice processing, approvals and the close actually run. The result is numbers: cycle time per document, first-pass rate, rework loops, waiting time per step. The documented process and the lived process differ in every company I have seen.

For companies with EUR 4m to 100m revenue, ERP exports and lightweight tools are enough; the large suites are oversized. The precondition is a timestamp per step: receipt, capture, approval, posting, payment. Where documents travel by email and Outlook folders, the stamps are missing, and the process picture starts only at capture.

I do not yet have a mandate figure of my own from process mining; the process picture is a deliverable of stage 1 on the page AI automation in finance. What stands there is the method. As soon as a measured value exists, it appears here with a date.

What did this look like in a mandate?

The two cases since July 2026 show the same cause in two forms. In the reporting case it was 11 files without a single link, plus master data in three systems with three spellings of the same supplier. In the PeriGuard auction, the inventory list from the acquisition and the lot list shared no common identifier.

In the auction the inventory list had 239 items without location and serial numbers for 43 of them. A machine matching attempt produced 31 unambiguous matches, 40 ambiguous ones and 168 without a match. The path from the lot list into the fixed assets was rebuilt through the document chain: 239 lots, each with its invoice chain and document link.

The document base of the auction comprised 228 files with 598 pages, among them 55 buyer invoices in one collective folder and invoices in up to three versions from cancellation, credit note and reissue. Payment notes existed only in file names. Every cash flow had to be reconstructed from documents, because there was no read access to the payment account.

All figures with dates and the split between machine and financial judgement are on the PeriGuard mandate page. The reporting case stays anonymised; its operating values follow after the first month in operation. What both cases share: no tool was touched before the data base, and every figure today has a document behind it.