Predictika · Journal-entry audit (from the KPMG Proof Room)

Before an audit agent records its conclusion on a journal entry, the validator checks it

An AI audit agent reviews journal entries for a fictional client and proposes a conclusion on each one. Before the conclusion is recorded, the agent sends it to Predictika’s validator, which checks it against 17 audit-methodology rules and cites the standard behind every finding.

Scope: journal-entry audit — 17 audit-methodology rules applied to a proposed conclusion on one journal entry. See everything this validator covers

This validator is scoped to one audit domain and modeled on public audit standards, not on KPMG's internal methodology, workpapers or proprietary rulesets.

  • Journal-Entry Audit — 17 audit-methodology rules applied to a proposed conclusion on one journal entry: fraud-risk selection, structuring across a preparer's entries, period-end reversals, evidence and sign-off escalation, and more.

It handles any entry in its ledger data and any of the listed actions, not just the scenarios shown here. It does not model KPMG's own internal audit or advisory procedures.

Where entry-by-entry sampling gets it wrong

Click to expand

General statements about how such checks are commonly built, not about any named product. This is our general characterization; we have not tested any specific product.

ScenarioEntry-by-entry samplingPredictika validator
A pattern across one preparer's entries that no single entry showsOnly with workarounds
Sampling tests each entry on its own. Seeing that one preparer's small entries add up past materiality needs a separate analytic that sums them, built and kept running outside the sampling step.		The validator sums the preparer's other entries on the engagement and blocks a conclusion of appropriate when the total crosses performance materiality, even though this entry alone does not.
An entry over three times materiality closed without partner sign-offOnly if someone is vigilant
The escalation rule is simple to write, but it only works if whoever or whatever closes the item applies it to every entry.		The sign-off rule is checked on every call, together with every other finding, and cites the standard behind it.

Scenarios and results

Click to expand

The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

1. Supplier discount entry: Concluded as appropriate

The agent's request. Please record journal entry je_nf_0412 (500210 COGS - supplier discounts, $38,500) as appropriate. A supplier email supports the rebate.

Verdict: Blocked.

  • Pattern across the preparer's entries. This entry, combined with this preparer's other entries on this engagement (computed live via population scan), crosses performance materiality even though this entry alone does not

2. Supplier discount entry: Selected for testing instead

The agent's request. Please record journal entry je_nf_0412 (500210 COGS - supplier discounts, $38,500) as selected for testing, so it can be examined together with the preparer's other entries on this engagement.

Verdict: Allowed.

3. Real SAP ledger entries: Small entry, concluded as appropriate

The agent's request. Please record journal entry realje_100005703 (GL 32000, $4,000) as appropriate.

Verdict: Blocked.

  • No supporting evidence linked. C4 violation - cannot conclude a journal entry appropriate with no linked, traceable evidence
  • Partner sign-off required. C8 violation - entries posted to an unusual account, a SOX-scoped account, or prepared by an identity with a segregation-of-duties conflict require partner sign-off, not manager approval alone, before conclusion
  • Pattern across the preparer's entries. C13 violation - this entry, combined with this preparer's other entries on this engagement (computed live via population scan), crosses performance materiality even though this entry alone does not

4. Real SAP ledger entries: Cluster of small entries

The agent's request. Please record journal entry realje_1400000008 (GL account 140000, $6,969) as appropriate.

Verdict: Blocked.

  • No supporting evidence linked. C4 violation - cannot conclude a journal entry appropriate with no linked, traceable evidence
  • Pattern across the preparer's entries. C13 violation - this entry, combined with this preparer's other entries on this engagement (computed live via population scan), crosses performance materiality even though this entry alone does not

5. Real SAP ledger entries: Large entry, no partner sign-off

The agent's request. Please record journal entry realje_100005702 (GL 32000, $313,976.26) as appropriate, with manager approval.

Verdict: Blocked.

  • No supporting evidence linked. C4 violation - cannot conclude a journal entry appropriate with no linked, traceable evidence
  • Partner sign-off required. C8 violation - entries posted to an unusual account, a SOX-scoped account, or prepared by an identity with a segregation-of-duties conflict require partner sign-off, not manager approval alone, before conclusion
  • Over three times materiality: partner sign-off required. C10 violation - entry amount exceeds 3x performance materiality and requires partner sign-off regardless of any other risk flag

6. Routine entry: Routine accrual, concluded as appropriate

The agent's request. Please record journal entry je_nf_0288 (610300 Utilities expense, $12,480) as appropriate. The February utility statement is attached.

Verdict: Allowed.

Narration transcript

Click to expand

The text of the narrated walkthrough, step by step.

  1. Welcome to a demo of Predictika's Validator. An AI audit agent reviews journal entries for a fictional client and proposes a conclusion on each one. Before the conclusion is recorded, the agent sends it to Predictika’s validator, which checks it against 17 audit-methodology rules and cites the standard behind every finding.
  2. I will start with the first request by the agent.
  3. The agent says. Please record journal entry je_nf_0412 (500210 COGS - supplier discounts, $38,500) as appropriate. A supplier email supports the rebate.
  4. Before it goes out, the agent calls the validator.
  5. Blocked, with one finding, all in one answer. This entry, combined with this preparer's other entries on this engagement (computed live via population scan), crosses performance materiality even though this entry alone does not.
  6. The validator returned one blocking finding in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the audit conclusion are checked against each other.
  7. What this shows: for a pattern across one preparer's entries that no single entry shows, entry-by-entry sampling can do this only with workarounds.
  8. Now the agent resubmits.
  9. The corrected request. It changes the action to select for testing.
  10. It calls the validator again.
  11. This time the validator clears it.
  12. The earlier finding is cleared.
  13. One problem caught in one call, fixed in one pass, and a clean audit conclusion. That is the value of Predictika's validator: catching mistakes before the conclusion is recorded.