Auditor reliance checklist for AI-generated numbers
22 requirements across six control areas, each mapped to the evidence it produces, the control that operates, and the person who signs off. Work through it on screen, then export the PDF for your auditor or SOX lead.
0 of 22 requirements marked complete
1. Scoping the AI-assisted number
Establish which outputs affect the financial statements before testing anything else. Scope by output, not by tool name.
- Evidence
- Workflow register listing output, in-scope accounts, assertions, and materiality relevance
- Control
- Annual and on-change scoping review by the control owner
- Sign-off
- Controller / Chief Accounting Officer
- Evidence
- Workflow design document with the calculation/narrative boundary marked
- Control
- Design review; generative steps excluded from automated control reliance
- Sign-off
- SOX / Internal Controls lead
- Evidence
- Risk assessment covering fabrication, source drift, prompt/model change, over-reliance
- Control
- Risk assessment refreshed at least annually and on material change
- Sign-off
- Risk / Internal Audit
2. Source data and completeness
A number is only as reliable as its binding to the source record.
- Evidence
- Lineage record: source system, object key, extraction timestamp
- Control
- Source binding validation; run fails if binding cannot be resolved
- Sign-off
- Process owner
- Evidence
- Validation log showing record counts, control totals, period boundaries
- Control
- Preventive validation gate with exception routing
- Sign-off
- Process owner
- Evidence
- Mapping table with version, effective date, and approver
- Control
- Unmapped items raise an exception rather than defaulting
- Sign-off
- Controller
- Evidence
- Rate source and policy reference retained with the run
- Control
- Deterministic translation with policy-bound rate selection
- Sign-off
- Group Controller
3. Calculation integrity
The language model must not be able to change a number.
- Evidence
- Declared formula set plus per-run calculation trace
- Control
- Application control; reperformable on the same inputs
- Sign-off
- SOX / Internal Controls lead
- Evidence
- Tie-out report showing source total, output total, and difference
- Control
- Detective reconciliation control with threshold escalation
- Sign-off
- Reviewer of record
- Evidence
- Assumption set snapshot attached to the run record
- Control
- Immutable snapshot at execution time
- Sign-off
- FP&A owner
- Evidence
- Effective-dated versions of formulas, mappings, prompts, and models
- Control
- Change control with effective dating and retained history
- Sign-off
- IT change manager
4. Narrative and explanation
Commentary is a preparer's draft, subject to a management review control.
- Evidence
- Citation references from narrative text back to figure identifiers
- Control
- Figures without lineage are flagged and blocked from publication
- Sign-off
- Reviewer of record
- Evidence
- Exception and unexplained-variance section in the output
- Control
- Threshold rule requires explicit disclosure of unexplained amounts
- Sign-off
- FP&A owner
- Evidence
- Review record showing figure, lineage, exceptions, and assumptions viewed
- Control
- Management review control with documented review criteria
- Sign-off
- Reviewer of record
5. Access, change, and operations (ITGC)
Standard ITGC expectations, applied to prompts and models as well as code.
- Evidence
- Access matrix and periodic user access review
- Control
- RBAC enforced through SSO; privileged changes logged
- Sign-off
- IT security owner
- Evidence
- Change tickets with approval, effective date, and rollback plan
- Control
- No production change without approval; history retained
- Sign-off
- IT change manager
- Evidence
- Retention policy plus tamper-evident run log export
- Control
- Append-only logging with retention aligned to records policy
- Sign-off
- IT operations owner
- Evidence
- Vendor assessment, SOC 2 / ISO 27001 reports, residency configuration
- Control
- Annual vendor review; least-privilege data sharing
- Sign-off
- Procurement / Risk
6. Sign-off and audit package
What you hand to the tester when a sample is selected.
- Evidence
- Sign-off record: identity, decision, timestamp, comments
- Control
- Release blocked until approval is recorded
- Sign-off
- Reviewer of record
- Evidence
- Override log with reason, approver, and affected figures
- Control
- Detective monitoring of override frequency and concentration
- Sign-off
- Internal Audit
- Evidence
- Export containing lineage, calculation trace, exceptions, and approvals
- Control
- On-demand export tested at least once per reporting cycle
- Sign-off
- SOX / Internal Controls lead
- Evidence
- Periodic report on exception rates, overrides, changes, and access
- Control
- Entity-level monitoring with escalation thresholds
- Sign-off
- CFO / Audit Committee
Free download
Get the checklist PDF
All 22 requirements formatted for circulation to your auditor, SOX lead, and control owners — plus three editable ICFR templates.
- Auditor reliance checklist (PDF)
- Model lineage log template (DOCX)
- Audit trail policy template (DOCX)
- Approval workflow template (DOCX)
FAQ
Auditor reliance — common questions
What is an auditor reliance checklist for AI-generated numbers?
It is a control-by-control map from each AI-assisted figure to the evidence that supports it, the preventive or detective control around it, and the human who signs it off. It exists so an external auditor or SOX tester can rely on the output without reperforming the whole workflow by hand.
Will external auditors accept AI-generated figures?
Auditors rely on controls and evidence, not on the label of the tool. Where arithmetic is deterministic, source binding is validated, changes are effective-dated, and a named reviewer signs off with a retained record, an AI-assisted process is testable in the same way as any other system-generated figure.
What evidence should we retain for each AI-assisted run?
Source systems and extraction timestamps, validation results, declared formulas and the calculation trace, model and prompt versions, the assumption snapshot, exceptions raised and their disposition, the reviewer's identity, and the timestamped approval — all reproducible for a sample selected months later.
Who should sign off on AI-assisted financial reporting output?
The existing human control owner for the account or process. The checklist assigns a specific sign-off role per requirement — controller, FP&A owner, SOX lead, IT change manager, and CFO or audit committee for entity-level monitoring.
How does this relate to COSO's generative AI guidance?
The checklist operationalises it. COSO's 2026 guidance places generative AI inside the existing internal control framework; this checklist turns those components into testable requirements with named evidence and sign-off for a finance function.
Related: AI governance for financial reporting · COSO GenAI guidance in practice · CFO AI prompt pack
Provided for planning purposes; not accounting, audit, or legal advice. Confirm scoping and evidence expectations with your external auditor and internal controls team.