Why we publish no percentages until we have twenty cases

Audit marketing likes percentages: “we recover 92 % of amounts claimed”, “we find errors in 7 invoices out of 10”. The number looks like proof, but without a sample it is just a confident-sounding sentence.

Three out of three is not one hundred percent

If the first three cases ended in success, the percentage is 100. A fourth, unsuccessful case turns it into 75. A fifth, into 80. An indicator that jumps by tens of points with every new case measures nothing: it shows how many cases there were, not how well things are going.

That is exactly why a small-sample percentage is more dangerous than no percentage at all. It creates an impression of precision where there is none yet, and decisions get made on a coincidence.

Why twenty

Twenty resolved cases is not a statistical ceremony — it is the point where a single case stops moving the indicator by several points. Five cases are too few for the number to settle; waiting for a hundred would mean saying nothing for a year or more.

The threshold applies per module and per check. A sample from terminal invoices says nothing about excise refunds, and within one module different checks can perform differently.

The same threshold applies to us

It would be easy to say “the sample is too small” in public while relying on those same five cases internally when deciding which module to expand. The threshold is written into the code and shown on the internal dashboard: below twenty cases the dashboard shows not a number but an explanation of why there is none.

What to judge until there are percentages

Ask whether every finding comes with a reproducible calculation: inputs, formula and result that can be recomputed without us. A number without inputs is an opinion, however firmly it is stated.

Ask what it rests on. A tariff, a contract clause or a legal act should be cited with the date from which it applies. In our report every source carries a fingerprint of its content — if a source goes stale, the finding drops into review automatically rather than staying in an amount you would not be able to defend.

And ask whether the system tells you what it did NOT check. A report listing only findings creates the impression that everything was checked — which is more dangerous than a missed error.

You can read a sample pack — the report with its calculations and the claim template — without submitting any documents of your own.