SOX in two minutes
The Sarbanes-Oxley Act of 2002 set requirements for US public companies' financial reporting. Two sections matter most for finance teams:
- Section 302: the CEO and CFO certify each quarterly and annual report, including that they're responsible for internal controls.
- Section 404(a): management assesses and reports on internal control over financial reporting (ICFR) every year.
- Section 404(b): the external auditor attests to ICFR for accelerated and large accelerated filers.
Most US companies use the COSO Internal Control–Integrated Framework (2013) as the basis for that assessment. Private companies aren't subject to SOX, but PE-backed and VC-backed companies preparing for an IPO or sale often build SOX-style controls a year or two ahead, and lenders and buyers increasingly expect them.
Definitions → /glossary/audit-controls-evidence#sox-404 · /glossary/audit-controls-evidence#icfr
What "SOX compliance software" actually covers
The phrase covers several kinds of tools. Know which one you're looking for:
| Category | What it does | Examples of the job |
|---|---|---|
| Controls management / GRC | Documents risks and controls, schedules testing, tracks deficiencies | Control matrix, testing workflow |
| Close and reconciliation software | Runs key financial controls and keeps their evidence | Reconciliations, review sign-offs, JE approval |
| Access and IT controls | Manages user access, change management, logs | Quarterly access reviews |
| Audit collaboration | Manages PBC requests and evidence sharing | Request lists |
Yoraito is in the second group: it runs close controls and keeps the evidence behind them, including the SQL behind every number.
What changes when AI prepares close work
AI doesn't get a SOX exemption, and it doesn't create a new kind of control. It adds four things your existing controls must reach:
- The AI is a user. If it reads or writes in your financial systems, it belongs in access reviews and the segregation-of-duties matrix.
- The AI's method is a system change. Prompts, skills, rules and model versions that affect the numbers fall under change management.
- AI output can be evidence. Reconciliations and schedules prepared by AI and relied on in the audit are information produced by the entity (IPE), which needs evidence of completeness and accuracy.
- AI judgment needs review. Explanations and estimates drafted by AI are inputs to a management review control, not the control itself.
Key controls when AI is in the close
| Area | Control | Evidence your auditor will want |
|---|---|---|
| Access | AI connections use a dedicated, least-privilege role; included in quarterly access reviews | Role definition, review sign-off |
| Posting | Nothing AI-prepared posts without approval by a named person | Approval trail per entry |
| Method change | Changes to prompts, rules, skills or models are requested, tested and approved | Change tickets, test results |
| Completeness of inputs | Row counts and totals checked from source to output | Input profile per run |
| Accuracy of outputs | Key figures can be reperformed | Saved query or calculation per figure |
| Review | Reviewer checks AI-prepared reconciliations and explanations at a defined threshold | Review notes, dated sign-off |
| Monitoring | Periodic tests of AI output on data with known answers | Test results by version |
| Exceptions | Errors found in AI output are logged and fixed | Exceptions log |
For journal entries specifically → /blog/erp-mcp-journal-entry-controls
COSO's 2026 guidance on generative AI
In early 2026, COSO published Achieving Effective Internal Control Over Generative AI. It applies the five COSO components to eight generative AI capability types: ingestion, transformation, posting, orchestration, judgment, monitoring, regulatory intelligence and human-AI interaction. For each type it gives minimum control expectations and illustrative evidence. We map the eight types to close tasks here → /blog/coso-generative-ai-controls-month-end-close.
What auditors are asking now
The PCAOB sought public comment on data and technology as a standard-setting topic in 2026, and the large firms broadly favoured staff guidance on how existing standards apply to AI before new rules. Existing standards on audit evidence, fraud and ICFR audits already reach AI-prepared work. Expect questions like:
- Where did AI touch the financial statements?
- What could it read and change?
- Who reviewed its output, and how?
- Can the numbers be reperformed?
More → /blog/pcaob-ai-what-auditors-will-ask. For a year-end evidence pack → /blog/year-end-audit-ai-documentation.
Designing an AI-prepared reconciliation as a key control
A strong control description says who, what, when, how precisely and what evidence is kept. Example:
Each month by BD4, Yoraito prepares the reconciliation for every bank account. Each match records the rule that made it and each figure records the SQL that produced it. By BD6, the controller reviews each reconciliation, investigates every reconciling item over $5,000 and any item older than 60 days, and signs off in the system. Evidence: the reconciliation, the saved queries, the reviewer's notes and the dated sign-off.
Illustrative wording; adapt thresholds and timing to your own risk assessment.
Sample risk and control matrix rows for AI in the close
A risk and control matrix (RCM) links each risk of misstatement to the control that addresses it. These rows show how AI uses fit into one. Adapt them to your own risk assessment.
| Risk | Control | Type | Frequency | Owner | Evidence |
|---|---|---|---|---|---|
| AI-prepared bank reconciliation is incomplete or wrong | Controller reviews each reconciliation; investigates items over $5,000 or older than 60 days | Manual, detective | Monthly | Controller | Signed reconciliation, review notes, saved queries |
| AI drops rows when reading source data | Row counts and totals compared from source to output for each run | Automated, detective | Each run | System; reviewed by preparer | Input profile |
| AI-drafted entry posts without approval | System requires named approval before posting | Automated, preventive | Each entry | System | Approval log |
| Unapproved change to AI rules or prompts alters results | Changes requested, tested on known-answer data and approved | Manual, preventive | Each change | Process owner | Change tickets, test results |
| AI connection has excessive ERP access | Dedicated least-privilege role; included in quarterly access review | Manual, preventive | Quarterly | Systems owner | Role definition, access review sign-off |
| AI-drafted flux explanation is unsupported | Reviewer checks support label; "needs owner" items resolved before sign-off | Manual, detective | Monthly | Controller | Labelled flux with sign-off |
A 12-month SOX readiness path for PE- and VC-backed companies
If an IPO, sale or lender requirement is 12 to 24 months away, this sequence keeps AI from becoming a late surprise:
| Months | Focus | Output |
|---|---|---|
| 1–2 | Risk assessment; inventory every AI use in financial reporting | Scoping memo; AI use inventory |
| 3–4 | Document processes and key controls, including AI steps | Process narratives; RCM |
| 5–6 | Fix design gaps: access roles, approval enforcement, change control | Updated controls |
| 7–9 | Operate the controls and keep evidence each close | Three months of evidence |
| 10–11 | Test controls internally or with an adviser | Test results; deficiency log |
| 12 | Remediate and re-test | Readiness report for the board |
Common control deficiencies when AI is involved
- The AI shares a human's login, so the audit trail can't tell who did what.
- Review is a signature only, with no notes on what was checked in AI-prepared work.
- Rules or prompts change without approval, so last quarter's tested control isn't this quarter's control.
- Outputs can't be reproduced, so the auditor can't rely on them as evidence.
- No exceptions log, so there's no proof the review catches anything.
What to look for in software, if SOX matters to you
- Reperformance: can your auditor reproduce each number without the vendor?
- Separation: are preparer, reviewer and approver recorded separately, and can the AI ever approve?
- Access: what can the tool read and write in your ERP, under which role?
- Change control: how are rules and AI behaviour versioned, and who approves changes?
- Assurance: does the vendor have a SOC 1 or SOC 2 report?
- AI off: can you run the process with AI drafting turned off if a control fails?