The trust gap: how a fund knows the numbers it is given are right

Start with three packs. The company is the invented one used throughout this series. Figures are monthly EBITDA in thousands of euros.

The January pack reported January at 450. A month later January had become 400. The February pack reported February at 520. A month later February had become 470. On that pattern, March's 540 will be 490 by April.

Nothing here is dishonest. Supplier invoices for January arrived in February and were posted to January. The pack for each month was built before that month's costs were complete.

The effect is that the board never sees a correct trend. In February it saw January at 400 and February at 520, a rise of 120. The settled figures are 400 and 470, a rise of 70. Every month the newest figure flatters, and every month the comparison is between one finished month and one unfinished month.

The board also never discussed the 50. It appeared in no variance commentary, because by the time it existed the meeting had moved on.

The person who assembles the number

A fund owns companies it does not run. Each month it reads figures prepared by the people those figures measure. The finance director who builds the EBITDA line is often paid a bonus on it.

This need not produce dishonesty. It does mean the number passes through one pair of hands, under time pressure, with a known preference for one direction. A late cost that would lower this month's figure is not chased. A judgement call falls the comfortable way. Over time the method leans, though each person in it is acting honestly.

Spreadsheets add their own share. Raymond Panko collected the field audits of operational spreadsheets. In his 2008 revision, 88 percent of the 113 spreadsheets examined in seven studies since 1995 contained errors. The samples are old and the studies counted errors of every size, so the figure is a warning and not a measurement of any one company.

Four properties of a number that can be relied on

It does not depend on who prepared it

The number comes from a fixed calculation applied to the records. Anyone who runs it gets the same result. The person it measures does not assemble it by hand. We put this first because the other three follow more easily once it holds.

It reconciles

The pack agrees with the ledger. The ledger agrees with something outside the company: bank statements for cash, confirmations for debt, a count for stock.

This only works if the ledger itself is closed for the month. A ledger that is missing its accruals is not the truth, and a pack that ties to it is no better. In the example, the fault lies in the close, and the spreadsheet only passes it on. Had the unbilled January costs been accrued in January, in the ledger, January would have been 400 the first time.

It stays where it was put

A reported month stays as reported. If it has to change, the change is listed with its amount and reason, and the board is told. A definition is written down, and a change of definition is dated.

It opens

Any figure in the pack can be opened to the entries behind it. When a board member asks why other expenses rose, the answer is a list of transactions.

What this looks like in a Greek company

A portfolio company of this size has its own accounting department and its own system. The difficulty is in how the month is closed.

The statutory accounts are closed properly once a year. That is when the stock is counted, the provisions are reviewed and the auditors arrive. Law 4308/2014 ties the stock count to the balance sheet date. It allows a rolling count where the company keeps reliable records of goods in and out. Where it does not, stock is known in December and estimated for eleven months, and so is gross margin.

Monthly, the ledger receives invoices as they come. Costs incurred and not yet billed are often left out until the invoice arrives. The holiday bonuses are often expensed when paid. The pack is then built beside the ledger, in a spreadsheet, and the corrections that a proper close would post are made there, or not made.

It is tempting to think that electronic reporting to the tax authority has fixed this. It has not, and it was not designed to. The myDATA platform receives every sales invoice, and it also receives summary entries for payroll each month and for depreciation each year. Other adjusting entries may be sent up to the deadline for the income tax return. It is a tax record on a tax timetable. It tells the State what was invoiced. It does not tell a board whether July is complete.

The remedy is a real monthly close in the ledger, with accruals, payroll spread across the year and a stock figure that can be defended. After that the pack only reports what the ledger holds.

A layout: the tie-out sheet

One page, each month, before the pack is sent. Here is March again, in thousands of euros, after the close has been repaired. The costs incurred in March and not yet invoiced were accrued in the ledger, so the pack reports 490 and not 540.

An unexplained difference goes back to the finance team, and the pack waits. A restatement of an earlier month goes to the board, listed, with the pack. From April the second row should read zero.

Where artificial intelligence fits, and where our product stands

Language models now write commentary and answer questions about performance. We build one into our own product, BPM Analyst, so we have an interest here and should say so.

The risk is easy to state. A model writes fluently, and fluent text is easy to believe. In November 2023 the FinanceBench study put 150 questions about published company accounts to the leading model of the time. When the model had to find the relevant pages itself, it answered wrongly or declined in 81 percent of cases. Given the whole filing, it failed in about 21 percent. Given exactly the right pages, 15 percent. That study is three years old and the models have changed. The lesson of the spread has not: the answer depended on what the model was looking at, and the reader could not see what that was.

So the question to ask of any commentary written by a model is where each number came from. BPM Analyst is built so that the model does not produce figures. Every figure is meant to come from a tested calculation in the data warehouse, run on the company's own records. The model is given those results and explains them. Where a figure it would need has not been computed, the design is that it says so and stops. Each number in the text is meant to be traceable to the calculation that made it.

Two limits should be stated. This is how the product is designed, and a design is not a guarantee. And an explanation is still a judgement. A correct figure with the wrong reason attached is a misleading board paper, so a person should read the commentary before it reaches a board.

A model added to a pack that does not reconcile makes things worse. It gives an unreliable number a more persuasive voice.

Four questions about a monthly pack

1. Are the comparatives in this month's pack the same as the figures reported last month? If not, is there a list?
2. Does reported EBITDA agree with a ledger that was closed for the month, accruals included?
3. Can the finance team open any figure to its transactions while the board waits?
4. If a model wrote any of the commentary, where did each number in it come from?

The first question costs nothing. Anyone on the board can lay two packs side by side.

A company that passes all four has no trust gap worth an outside fee. Most of the work lies in the monthly close, and a capable finance director can run that without us.

Sources

• Panko, R.R. What We Know About Spreadsheet Errors. Web version revised May 2008; originally published in the Journal of End User Computing, 10(2), Spring 1998, 15–21. The 88 percent of 113 figure is in the 2008 revision.
• Islam, P., Kannappan, A., Kiela, D., Qian, R., Scherrer, N. & Vidgen, B. (2023). FinanceBench: A New Benchmark for Financial Question Answering. arXiv:2311.11944. Sixteen configurations tested on a sample of 150 questions.
• Law 4308/2014, article 4 paragraph 4 (stock count), and the Accounting Guidance on that law, paragraph 4.4.3 (rolling counts).
• Independent Authority for Public Revenue. myDATA technical specification, version 2.0.1, March 2026: entry types 17.1 (payroll), 17.2 (depreciation) and 17.3 to 17.6 (other adjusting entries).