Evidence by Construction: Make the Review File a Workflow Output
A good AI-generated finance memo is not enough if the next reviewer cannot rebuild the path that produced it.
The better operating pattern is evidence by construction: make the review file a required output of the workflow itself. This does not mean turning every finance workflow into an audit. It means designing the workflow so the evidence needed for review is emitted while the work happens.
What “reviewer-ready” should mean
A useful benchmark comes from audit documentation, with an important boundary: the benchmark is about reconstructability, not about claiming audit status. PCAOB AS 1215 requires audit documentation detailed enough for an experienced auditor with no prior connection to understand the procedures performed, evidence obtained and conclusions reached, and to identify who performed and reviewed the work.
Provenance standards point in the same direction. W3C PROV models entities, activities, agents, time and derivation. OpenLineage represents run events with timestamps, producers, schema versions, inputs and outputs. These support an event-based design in which provenance is captured as part of the run rather than relying on memory-based reconstruction later. Whether that materially reduces review effort must be measured in the target workflow.
NIST’s AI RMF Core adds another piece: human-oversight processes, roles and scope should be defined and documented. For a finance workflow, that means the file should say where preparation ends and where a human-owned decision begins.
The failure mode: evidence assembled after the answer
Post-hoc evidence assembly can make simple but material mistakes harder to detect and reconstruct. A number may be real but from the wrong period. A transformation may be correct but undocumented. A source may have changed since the run. An exception may have been discussed but never carried into the final package. A derivative may point to an older draft. A reviewer may see a recommended disposition and mistake it for an approval.
Design goal, not a guarantee. A review file built during the run is designed to make these failures easier to detect before the workflow advances. It does not guarantee prevention.
The evidence chain
claim/number → source → exact location → retrieval time → transformation → validation → exception → proposed disposition → reviewer action
The exact implementation can vary. A spreadsheet may use source-row IDs and formulas. A research workflow may use URLs, section anchors and extraction timestamps. A reconciliation may use account IDs, trade dates, balance snapshots and break IDs. The important part is that the chain is carried forward into the review file instead of being recreated from prose.
What belongs in the file
| Layer | Minimum evidence | Reviewer question |
|---|---|---|
| Scope | Run ID, entity, period/cutoff, scope version | What exactly was reviewed? |
| Inputs | Source ID/location, retrieval time, version/hash | Which evidence existed at the time? |
| Claims & numbers | Exact location, entity, period, units, transformation | Can I rebuild the result? |
| Validation | Observable checks and results | What was independently checked? |
| Exceptions | Evidence, unresolved gap, proposed disposition, owner | What is still open? |
| Output identity | Artifact version/hash | Am I reviewing the right version? |
| Human handoff | Exact decision, options, remaining checks, sign-off status | What decision is still mine? |
Download / copy the Reviewer-Ready Handoff Record template (YAML + synthetic example + reviewer checklist).
Preparation is not approval, and approval is not execution
This boundary becomes concrete in finance-agent reference designs. Anthropic’s public GL Reconciler traces breaks to transaction-level evidence, re-verifies them, and formats an exception report for controller sign-off; it explicitly does not post to the ledger. Its Month-End Closer stages accruals, roll-forwards, variance commentary and journal-entry drafts for controller review, while posting remains outside the agent. Its KYC Screener packages rule outcomes and gaps but reserves the decision for a compliance officer.
These are provider examples, not an industry standard. But they illustrate a useful authority split:
Preparation ≠ recommendation ≠ approval ≠ execution.
A synthetic reconciliation example
In a synthetic GL-to-subledger case, a frozen GL balance of USD 12,450,000 and a frozen subledger balance of USD 12,425,000 produce a +USD 25,000 break. The review file records the subtraction, source rows, period, entity, units and an independent recomputation.
A synthetic candidate timing item, FEE-09130, appears in the GL extract but not in the bounded subledger extract. The record keeps the item open, carries the unresolved gap and asks the reviewer to accept after support, request more evidence, or reject and reclassify. The workflow prepares the case; it does not approve or post anything.
What to measure
If the design goal is to reduce review friction, measure that outcome instead of assuming it. Direct integrity tests can include material-claim provenance coverage, material-number provenance coverage, independent reconstruction success, exception completeness and version/hash match. Workflow outcomes such as reviewer clarification requests, median review time and reopen rate should be measured on comparable cases before claiming savings.
What this pattern does not prove
A reviewer-ready record does not make the underlying data correct. It does not prove that a model is accurate, a control is sufficient, a workflow is compliant, or a human decision is right. It does not prove lower risk, lower cost, faster close, fewer people or better audit outcomes.
What it does is simpler: it turns the evidence path into part of the product of the workflow.
A practical next step
Take one existing AI-assisted finance handoff and ask: could a second reviewer reproduce the material claims, numbers, exceptions and decision boundary without asking the model or preparer what happened? If not, the missing fields belong in the workflow before the next run.
Sources and notes
- PCAOB — AS 1215: Audit Documentation
- W3C — PROV-DM
- OpenLineage — API / run-event model
- NIST — AI RMF Core
- Federal Reserve — Supervisory Guidance on Model Risk Management
- U.S. GAO — 2025 Green Book
- Anthropic financial-services — GL Reconciler
- Anthropic financial-services — Month-End Closer
- Anthropic financial-services — KYC Screener
Research cutoff: 2026-10-04. The proposed Review File Contract and acceptance tests are Treasury Desk author synthesis, not an industry standard. CSS class rendering depends on the Treasury Desk site stylesheet; this fragment includes no CSS, scripts, tracking or embeds.