Require a citation for every number. Each figure the agent reports must carry the ledger query that produced it, the entry references it summed, and the provenance of each source. A figure without that trail is not reported; it is marked as unsupported. A reviewer then samples figures and rebuilds them from the cited entries.
Why do agent figures drift away from the books?
A language model writes numbers the way it writes words: fluently. Asked for revenue by region, it may calculate from the ledger, reuse a figure from last month's deck that sat in its context, or round an amount from an email. All three look the same in the final report.
That is why the house rule is that a finance agent's figures must trace to an entry. The rule is easy to state and only holds if the agent's tools are built so that numbers come from queries, not from prose.
What are the steps, and who acts?
- Finance systems owner: defines the sources. The ledger, the sub-ledgers and the consolidation tool are registered as sealed sources. Decks, spreadsheets and emails are registered as external. Evidence: the source register.
- Agent, depth 1: computes through queries. Every figure is produced by a query the platform runs against a sealed source. The agent never types a number it did not receive from a query. Evidence: the query text, its parameters and the time it ran.
- Agent: attaches the trail. Each figure carries the query ID and the list of journal entries it summed, or a link to that list where it is long. Evidence: the trail file stored with the report.
- Platform: blocks untraced figures. A number with no query behind it is flagged in the draft as unsupported and cannot appear in the final version. Evidence: the filed verdict.
- Reviewer: samples and rebuilds. A finance analyst picks figures at random, plus every material one, and rebuilds them from the cited entries. Evidence: the sample, the result and the reviewer's name.
- Report owner: signs off. The person who answers for the report releases it. Evidence: the sign-off.
What does a good trail look like?
| Level | What it shows | Example |
|---|---|---|
| Figure | The reported number | Q2 freight cost, 410,000 |
| Query | How it was computed | Sum of account 5410, entity UK01, April to June |
| Entries | What was summed | 312 entries, listed by reference |
| Support | Why each entry exists | The invoice or workpaper behind each entry |
| Provenance | Where each fact came from | Ledger: sealed; supplier invoice: external |
The support layer is where most trails break. An entry with no document behind it is a gap whether a person or an agent reported it. The journal entry support template shows what each entry should carry.
A worked example
Take a distributor whose agent drafts the monthly management pack. One line shows freight cost up sharply on the prior quarter.
The reviewer opens the trail. The query sums account 5410 for the quarter. The entry list shows one large item: a 60,000 accrual posted on the last day of June. Its support is an email estimate from the logistics team, provenance external. The figure is correctly computed, but its biggest component rests on an estimate. The reviewer adds a note to the pack, and the controller asks for the haulier's invoice before the accrual is reversed in July.
The trail did more than prove the arithmetic. It showed the reviewer which part of the number deserved a question.
What may the agent do, and what stays with a person?
The agent may compute, trace, compare periods and draft commentary on movements, at depth 1 for reading and depth 2 for the draft report. It may not adjust a figure to match a forecast, choose which entries to exclude, or release the report. Commentary that explains why a number moved must cite the entries too; "seasonal" is not an explanation unless the entries show it.
Board reporting and treasury reporting raise the stakes. Our answers on whether an AI agent can produce board reports and on preparing FX exposure reports with an AI agent apply the same tracing rule to those packs.
Which controls keep the trail honest?
- Numbers only from tools. The agent's brief forbids typing figures; the platform enforces it by flagging any number with no query attached.
- Provenance on every source. Sealed ledger data outranks an external spreadsheet. A figure resting on external sources is labelled as such.
- Reproducibility. Queries are stored with parameters, so the same figure can be recomputed later even after the ledger moves on.
- Sampling by a person. Tracing is only as good as someone checking it.
The topic guide AI agents for finance: books you can audit shows where this sits in the wider design.
Questions people ask
What if a figure comes from several systems?
The trail lists each query and source separately, and the reconciliation between them. If the systems disagree, the agent reports the difference rather than choosing a side.
Does tracing slow the agent down?
Very little, because the trail is a by-product of computing through queries. The real cost is the reviewer's sampling time, which is time you should be spending anyway.
Can old reports be traced after the ledger changes?
Yes, if the queries were stored with their parameters and the entry lists were kept. Rerunning the query against today's ledger will not reproduce the original if entries were later reversed, so keep the list.