Every number a finance agent reports should link back to a ledger entry or a source document that a reviewer can open. A figure that cannot be traced is an assertion, and in finance an assertion is not evidence. Traceability is what lets a controller rely on agent-prepared work without redoing it, and what lets an auditor test it.
Why is an untraceable figure a problem even if it is right?
Because nobody can tell that it is right. A finance agent reports that accrued expenses for the month are 48,000. The figure may be correct. But if the reviewer cannot see which purchase orders, contracts and estimates make up that total, they have two choices: recalculate it themselves, in which case the agent saved nothing, or accept it on trust, in which case the control has failed.
Agents make this problem sharper than it is with people. A person who prepares a schedule usually builds it from identifiable sources and can explain each line when asked. An agent may produce a confident total from reasoning that is hard to reconstruct, especially if it combined several sources or estimated something along the way. Unless traceability is required, an agent's figures can look exactly as authoritative as a person's while resting on much less.
The same applies to figures in commentary and reports, not just schedules. "Revenue was up because of the new contract" invites the question: which contract, how much, posted where?
What does tracing look like in practice?
Every reported figure should come with a path a reviewer can follow, without asking anyone:
- A posted balance traces to the journal entries that make it up, each with its own reference.
- A journal entry traces to its supporting documents: invoice, contract, calculation, approval.
- A calculated figure, such as an accrual or allocation, shows the inputs, each traced to its source, and the method.
- An estimate shows the basis and states plainly that it is an estimate, with who or what produced it.
- A summary figure in commentary links to the balance or the analysis it summarises.
The method is set out in detail in how to trace every figure an AI agent reports back to an entry. A journal entry support template is a good way to make the support for each entry consistent, whether a person or an agent prepares it.
What about figures from outside the ledger?
Many inputs to finance work do not start in the ledger: bank statements, supplier invoices, contracts, budgets, estimates from operating teams. Tracing them means recording where each came from and how reliable that source is.
| Figure type | Trace to | Typical provenance |
|---|---|---|
| Bank balance | Bank feed from the settlement rail or bank | Sealed |
| Purchase order value | Procurement system | Sealed |
| Supplier invoice amount | Invoice received from supplier | Signed if from a verified supplier channel; external if emailed |
| Accrual estimate from a department | Email or form from the budget holder | Declared |
| Exchange rate | Rate source the firm has adopted | Sealed if from the firm's adopted rate system |
| Figure from a web page or news article | The page | External |
Provenance tells the reviewer how much weight to put on each input. An accrual built from purchase orders in the procurement system rests on firmer ground than one built from an emailed estimate. Neither is wrong to use, but the reviewer should know which is which.
The same reasoning matters most where money moves. A bank account number taken from an emailed invoice is external, and should never be the basis for a payment on its own. That is why supplier bank detail changes are where finance agents must stop.
What should the agent do when it cannot trace a figure?
Say so, and hold. If an agent cannot link a figure to a source, it should not present the figure as if it could. The right output is something like: "Accrued utilities estimated at 3,000 based on the prior three months' average; no invoice or meter reading found for this month." The reviewer can then decide whether to accept the estimate, find the missing information or ask the department.
What the agent must never do is produce a plausible number to fill a gap without flagging it. That is the most dangerous kind of error, because it looks like a sourced figure.
How does tracing help when something must be undone?
Corrections depend on knowing what was done and why. If an agent-prepared entry turns out to be wrong, tracing shows exactly which source it relied on, and therefore what else might be affected. Reversing an entry is itself an action that needs care; whether an AI agent can reverse a journal entry depends on its depth and the firm's limits, but in every case the reversal should trace to the original and to the reason for reversing it.
How do you test traceability?
Pick a sample of figures the agent has reported, including some buried in commentary, and try to trace each one using only what the agent provided. Count how many trace fully, partly or not at all. Make that a certification measure: a finance agent that cannot trace its figures has not earned the right to prepare them for others.
Repeat the test periodically. Traceability tends to decay as new sources are added, formats change or the agent is asked to do new kinds of analysis.
Where Pryme fits
In Pryme, every fact reaching a finance agent carries its provenance class through the Pryme Context Protocol, so each figure in a prepared schedule can show whether it came from a sealed ledger entry, a signed supplier document or a declared estimate. Verdicts record which facts were relied on. This is part of AI agents for finance: books you can audit.
Questions people ask
Does every figure need a link, even small ones?
Every figure that feeds the ledger or a report should be traceable. For trivial items, a reference to the source document is enough; the principle matters more than the format.
What if the source is a spreadsheet?
Trace to the specific file, version and cell, and treat the spreadsheet's own inputs as needing a trace too. Spreadsheets are a common place for traceability to break.
Can tracing be automated?
Largely, for figures built from ledger entries and system data. Estimates and judgements need the agent to record its basis explicitly, and people to check it.