What auditable does not mean
It does not mean the decision was right. Auditable is about reconstruction, not quality: a decision that turned out badly can be perfectly auditable, and that is exactly when the record earns its cost, because it is the only way to tell a bad decision from bad luck.
It also does not mean more paperwork. Most of what makes a decision unauditable is not missing documents, it is a single number nobody can source and a spreadsheet that gives a different answer when someone opens it a year later.
The five records
- The options. Including the ones rejected, and including waiting, because a decision between one option and nothing is not a decision anyone can review.
- The inputs, with a source and an owner for each. Entered by whom, from what, on what date. A figure with no owner is the first thing an auditor pulls.
- The method. Not a description of software, but what was computed from what: this revenue over this period, less this capital, discounted at this rate.
- The result, including the bad case and the odds, not only the headline figure.
- The date and the decision maker. Who decided, when, and on which version of the figures.
The test that matters: can it be run again?
Reproducibility is the one property that cannot be faked with documentation. If the same inputs produce the same answer every time, then the record of the inputs is a record of the decision. If the answer moves, the record proves nothing and the discussion in two years will be about whose version was right.
This is where ordinary tooling tends to fail. A spreadsheet that has been edited by four people cannot be rerun as it stood; a chat assistant asked the same question twice gives two different paragraphs, and neither is a computation. Insist on the property, whatever you use.
Separate what was computed from what was said
Keep the computed figures and the commentary apart, in the record and on the page. The figures are either reproducible or they are not. The commentary is interpretation, and interpretation is where a decision is persuaded rather than computed. Where any part of a figure came from a judgment rather than a calculation, the record should say so in those words.
Write the bad case into the record
A record that only holds the expected outcome cannot support a review, because the question after a bad year is always whether the bad year was foreseen. If the low case was in the paper and was accepted, the decision was sound and the outcome was unlucky. If it was never written down, nobody can tell, and the organization learns nothing.
Doing it with BrainPower AI
The product keeps an audit trail behind every figure, holds the inputs and the result together, and reproduces a decision exactly from the same figures, which is the test above. The AI layer explains results in words and is never allowed to originate a number, which is the separation in the section before last, enforced rather than promised. See investment risk for what the record looks like on a decision with a real downside, and project approval for the version a board reads. Naming the options and owning the assumptions stay with you.
Questions people ask
Is an auditable decision the same as a documented decision?
No. Documentation describes a decision; an auditable decision can be reconstructed and rerun from its own record. Plenty of heavily documented decisions cannot be reproduced, and a few lightly documented ones can.
What do auditors and regulators usually ask for?
Who decided, on what information, when, and whether the risks were considered at the time. Each of those is one of the five records above. The order the questions arrive in varies; the records do not.
Can a decision made partly on judgment be auditable?
Yes, as long as the judgment is recorded as judgment. The failure is not using judgment, it is presenting judgment as computation. Say which figures were judged, by whom, and the record holds.
How long should we keep the record?
As long as the decision has consequences, which for capital decisions is longer than most retention policies assume. A kiln line approved today will be reviewed when it is rebuilt.
Can AI output be part of an auditable decision?
As explanation, yes. As the source of a number, no: a figure that cannot be reproduced on demand cannot be audited, and language models do not guarantee the same output twice. Keep the computation deterministic and let the AI describe it.