A finance agent can run on a healthy connection and still answer from incomplete history. Before you rely on its answer for one duty, confirm that the records able to change the conclusion arrived, through the right cutoff, and that your reviewer can open them.
A supplier credit appears in the first review. The document shows $20,000, the connection is running and the records look plausible. Then the Controller asks whether the import includes the month when the company applied its oldest credits, and you have not checked yet.
The credit note proves a credit existed. Working out what remains takes the later applications and settlements, and without them the available amount is unresolved. A finance AI data readiness check starts there: name the conclusion the duty supports and the records that would change it.
Write the question before checking the connection
For a supplier-credit review, the question is whether a credit remains available for a defined supplier and legal entity at a stated cutoff.
That takes more than credit-note headers. Your reviewer needs the applications, settlements and any terms that affect entitlement. The source owner then names the reports or objects that hold those records, with their date basis and filters.
Write the dependency in accounting language first. “We need all applications through the cutoff” is clearer than a report named “transactions” that nobody has opened.
Some missing fields make preparation harder. Other gaps block a conclusion, and missing allocation history is one of them.
Finance data completeness: compare against a source reference
Keep a source-generated reference for the agreed population. Record the entity, currency, period, filters, extraction time and the unit each row represents, and compare the import on the same basis.
In an illustrative example, the reference holds 1,240 allocation records totalling $3,100,000. The first import holds 1,180 records totalling $2,940,000. The gap of 60 records and $160,000 is a missing December.
Treat the import as short of the history the duty needs, and restrict conclusions on the affected credits while the source owner investigates the omission.
Counts and totals locate a problem without proving the whole population correct, because a duplicate and an omission of equal size cancel out. Keep source identities and check that each application connects to the right credit and entity.
Use a known historical event to challenge the answer
Choose a case whose outcome your accounting team can verify on its own.
Continue the example with credit CN-410. Its original value is $20,000. A restored December allocation shows $14,000 applied earlier, which leaves a possible remainder of $6,000, subject to further checks.
Before that history arrived, the remaining amount had no support. After the restore, the team can explain the prior application. It still has to check whether later activity or contract terms affect the remainder.
Ask the reviewer to show the record that changed the answer, and confirm its source identity, entity and date. If a backfill overlaps existing records, reconcile the ERP data backfill by verified identity before you rely on the repaired population.
That test asks more of the agent than a successful connection does: it has to retrieve the right historical event and apply it.
Check access and currentness as well as history
Your intended reviewer has to be able to open the supporting evidence. An agreement the implementation administrator can see might be closed to the person reviewing entitlement. Record that limit and the question it blocks.
Historical completeness and freshness are separate checks. A restored December says nothing about whether yesterday’s applications arrived. Record the latest covered event and the latest successful retrieval, and confirm that those dates support the duty’s cutoff. Finance data freshness vs completeness works through both.
Keep these checks tied to the conclusion. Your team might explain an earlier application while a claim about today’s available balance stays restricted.
Record what the evidence now permits
When the data changes, keep the earlier output and explain the revision. Revisit any AP packet or close review that used the incomplete history, because a corrected dataset does not reach someone already working from the old explanation.
Name the person who owns the source repair and the person who owns the accounting conclusion. Record the supported population, the remaining limitation and the evidence the next decision needs.
The readiness result is a plain statement of what your team can rely on for this duty, with open questions still visible. Once the inputs hold, evaluate the finance AI agent on that duty. Rever applies the same discipline in its zero-day scan, where missing sources and blocked procedures remain visible beside the findings; the Explore Rever page on the zero-day scan describes it.
The Check the History worksheet gives the Controller and source owner the worked case and a short blank review sheet for one duty.
To work through one duty on screen, What Does This Import Support? helps you define it, check its required history and record the evidence behind your source-readiness decision, in a printable matrix.
