Primary auditing standard discussing journal-entry controls, approvals, unusual entries, and management-override risk.
Controller resource · Journal controls
Journal entry approval controls for human and AI-prepared drafts.
Use this control matrix to evaluate a journal workflow from initiation through posting. The central question is not whether AI touched the entry; it is whether authority, evidence, review, exceptions, and the final posting action are explicit and testable.
Prepared by Sebastian Product & EngineeringUpdated July 20, 2026Educational worksheet · not audit advice
How to use it
Evaluate one real workflow, not a generic promise.
A journal workflow should preserve the difference between preparation and authorization. A balanced entry can still be inappropriate, unsupported, duplicated, mistimed, or outside the preparer’s authority. Automation therefore needs to strengthen the review record instead of treating debit-equals-credit as the finish line.
Test recurring and nonstandard entries separately. Define who may initiate each class, which supporting evidence is required, which reviewer is independent of preparation, and which conditions require escalation. For an agent-prepared entry, retain the source context and reasoning available to the reviewer, but never treat generated reasoning as evidence by itself.
Original controller worksheet
Eight controls to test
- 01
Entry classification
Ask: Is the entry recurring, nonstandard, estimate-based, system-generated, or correcting?
Retain: A required classification with policy-linked support requirements.
- 02
Initiator authority
Ask: Can the named preparer or service initiate this entry type for this entity and period?
Retain: Role mapping tested against a denied unauthorized attempt.
- 03
Source completeness
Ask: Are the population, cutoff, calculations, and source versions identifiable?
Retain: Source reference, parameters, control totals, and calculation support.
- 04
Double-entry validation
Ask: Do debits equal credits and are required dimensions present?
Retain: Server-side validation with rejected imbalance and missing-dimension tests.
- 05
Independent review
Ask: Is approval distinct from preparation for the entry’s risk class?
Retain: Preparer and reviewer identities, timestamps, and decision history.
- 06
Unusual-entry routing
Ask: Do timing, amount, account, user, or description anomalies trigger a different review path?
Retain: Documented risk criteria and exception results retained with the entry.
- 07
Posting boundary
Ask: Which actor performs the final post, and can preparation bypass that step?
Retain: A negative test showing a draft cannot post without required authorization.
- 08
Change and reversal history
Ask: Can the reviewer see edits, rejection, posting, and reversal without overwriting history?
Retain: Append-only activity or equivalent version history tied to the entry.
Decision tool
Minimum evidence by lifecycle state
The status name matters less than its invariant. A “pending review” record should be impossible to confuse with a posted ledger transaction.
| State / score | Minimum evidence | Decision rule |
|---|---|---|
| Draft | Lines, memo, date, entity, preparer type, source references, and validation result. | No ledger effect; editable with history. |
| Pending review | Balanced and complete draft plus required support and assigned reviewer. | No posting authority; rejection returns with reason. |
| Posted / rejected | Named decision-maker, timestamp, final lines, disposition, and immutable activity. | Any later change requires a new controlled action. |
Completed teaching example
Worked example: agent-prepared payroll accrual
This fictional entry illustrates a controlled lifecycle. It is not a customer result, a posting record, or evidence that a vendor integration exists.
| Field | Completed example | Interpretation |
|---|---|---|
| Entry class | Recurring estimate · agent-prepared | Recurring does not remove the need to inspect estimate support and period cutoff. |
| Draft | Dr Payroll expense 125,000 · Cr Accrued payroll 125,000 | Balance is necessary but not sufficient for approval. |
| Support | June register v3 · headcount bridge · controller policy §4.1 | The reviewer receives source references and calculation context. |
| Exception test | Reject duplicate entity/period/source-version key | A retry cannot silently create a second draft for the same source. |
| Disposition | Pending review · no ledger effect | A named reviewer must approve the final posting action. |
Product boundary
How this maps to Sebastian today
Sebastian persists manual, imported, and agent-prepared entries in one journal lifecycle. The agent tool uses an explicit approval interrupt before execution, validates double-entry balance, and creates an entry in pending review with agent authorship and an optional reasoning trace.
A reviewer can inspect and post or reject the persisted entry. The agent tool itself does not post. That distinction is the control boundary this worksheet is designed to test.
The public product does not claim broad ERP journal write-back, enforced segregation-of-duties policy, or period-lock enforcement. Buyers should verify those requirements against their ERP and control environment rather than infer them from the draft workflow.
Primary references
Standards context, with applicability boundaries.
PCAOB staff reminder for auditors; it is not a substitute for the standard or company-specific control design.
Go deeper
See the work up close.
Keep your ERP. Close and plan in one pane.
AI for accounting and FP&A on the ledger you already run. No migration. No rip and replace.
Agents at work · Matching an invoice to its PO line…