Accounting
Governing AI in bookkeeping and submission preparation
AI-assisted bookkeeping where small errors could flow straight into financial and tax submissions.
Illustrative · anonymised and reconstructed from real work
The scenario
Where the decision had to hold under scrutiny
Illustrative governance work applies the same discipline used for fraud rules to AI in bookkeeping and the preparation of financial and tax submissions. Because small classification errors or unsupported assumptions can flow through into final outputs, the focus is adding review and validation controls before anything is submitted — supporting reliability and reducing the risk of avoidable reporting errors.
What the scenario turns up
The issues
Finding 01
AI was automating bookkeeping and helping prepare financial and tax submissions, where a small classification error or an unsupported assumption could flow straight through into a final submission and become a compliance problem.
Finding 02
If a submission were later questioned, the basis for AI-assisted entries had to be defensible.
What the work does about each one
The response
Response 01
Adds review and validation controls before anything is submitted, catching errors and unsupported assumptions early and improving the reliability of what is filed.
Response 02
Records the basis for AI-assisted entries so the figures can be evidenced and explained if an authority challenges them.
How the work is done
The same five steps, applied to this decision
Every piece of work runs the same path. For this one, the weight falls on Design, Build, Assure & hand over — the steps where a decision of this kind is most often challenged.
Discovery
Map where the decision is made, who could challenge it, and the regime that binds it — before any logic is written.
Design
Set the logic out in the open so it can be read and questioned: the same inputs always give the same decision, and nothing is a black box.
Build
Build it with the checks in: coverage and oversight are confirmed before anything ships, and nothing unassessed is waved through as safe.
Deploy & test
Run it against real cases behind a human checkpoint, and keep a dated record of exactly what was decided and on what basis — one that cannot be edited afterwards.
Assure & hand over
Hand over the decision log and a plain-language write-up, so your team owns the record the moment someone says prove it.
What carries across
No automation covers this case yet
58 automations are running today, each built for a rule it names. None of them covers this case, and that is all that is claimed here. What carries across is the method: every one of them refuses to write a sentence its evidence will not support, which is the discipline this kind of work once ran by hand. Any of them can be shown to you running.
Related cases
The same standard, holding elsewhere
Discuss your situation
A decision like this can be made correctly, and shown to be correct, on the record
A short discussion is usually enough to locate where the risk sits — in the logic, the controls, the evidence, or a combination — and where it is made to hold.