---
title: How should you govern AI agents used in the month-end financial close?
url: https://opencontroller.lyzr.com/blog/govern-ai-agents-in-the-financial-close
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# How should you govern AI agents used in the month-end financial close?

**Treat a close agent as part of your internal control over financial reporting. Let it prepare reconciliations, accruals and draft journal entries, but keep posting behind a human approver who did not build or run the agent. Log every input, calculation and output, and keep the evidence your auditors already expect for IT general controls.**

Finance teams are among the first to deploy agents at scale because the close is repetitive, deadline-driven and data-heavy. Those same properties make mistakes expensive: an error in the close flows straight into reported numbers.

## Why is the close a special case for AI agents?

For US-listed companies, Section 404 of the Sarbanes-Oxley Act requires management to assess internal control over financial reporting, and for many companies an auditor attests to it. Auditors following PCAOB AS 2201 test both the business controls and the IT general controls (ITGCs) that support them. An agent that touches the close is in scope for both. Companies outside SOX face similar expectations from their own auditors and regulators.

## Which close tasks suit agents, and what control does each need?

| Close task | Agent role | Key control |
|---|---|---|
| Bank and account reconciliations | Match transactions, flag breaks, draft explanations | Reviewer signs off; unmatched items above a threshold escalate |
| Accruals | Propose amounts from contracts, POs and history | Calculations done by deterministic code, not the model; reviewer approves |
| Journal entries | Draft entries with support attached | Agent cannot post; approver is not the agent owner |
| Intercompany | Identify mismatches between entities | Both entity controllers confirm |
| Variance analysis | Explain movements against budget and prior period | Commentary reviewed before it reaches reporting |
| Close checklist tracking | Chase owners, update status | Low risk; standard logging |

A pattern runs through the table: agents prepare, humans approve, systems post.

## How do ITGCs apply to a close agent?

Auditors look at three ITGC areas. Each maps to agent controls.

1. **Access.** Who can run, change or approve the agent? What can the agent itself read and write in the ERP? The agent should have its own identity with the narrowest ERP role that works, and no posting rights.
2. **Change management.** Prompts, tools, model versions and code are all changes. Each needs a ticket, review and approval before it reaches the production close. Pin model versions so a provider update does not change behaviour mid-quarter.
3. **Operations.** Every run is logged: inputs, tool calls, calculations, outputs, approver and timestamps. Failures are detected and handled.

See [SOX ITGC evidence for AI agents](/blog/sox-itgc-evidence-for-ai-agents) for the evidence list.

## What does a close-agent checklist look like?

1. The agent has a named owner in finance and a technical owner.
2. The agent has its own identity, with read access to source data and no posting rights.
3. Numbers are calculated by deterministic code or the ERP, not generated by the model. See [should AI agents do financial calculations?](/blog/should-ai-agents-do-financial-calculations)
4. Every draft entry carries its supporting documents and calculation trail.
5. Approval thresholds are defined, for example entries above a set amount need a senior approver.
6. The approver is not the person who built or configured the agent (segregation of duties).
7. Prompt, tool and model changes go through change management and are frozen during the close window.
8. Outputs are evaluated against prior-period results before go-live, and spot-checked each close.
9. Logs are retained for your financial record retention period.
10. A fallback manual procedure exists if the agent is quarantined mid-close.

## What goes wrong most often?

- **Model arithmetic.** A language model adds up a column and gets it slightly wrong. Use tools for maths.
- **Silent model updates.** An unpinned model changes behaviour between quarters, and nobody re-tests.
- **Approver fatigue.** Reviewers rubber-stamp hundreds of agent drafts. Sample deeper and rotate reviewers.
- **Shared service accounts.** The agent runs as a human's account, so logs cannot separate person from agent.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents and versions with owners, supports agent workload identity federation so close agents run under their own identity, and records traces you can evaluate with multi-turn evals and LLM judges on production traffic. Rogue-agent handling and per-agent Quarantine let you stop an agent mid-close, and policy as code maps controls to frameworks you report against. See [what is an AI agent control plane?](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can an AI agent post journal entries automatically?**
It can technically, but most finance teams keep posting with a human approver so segregation of duties and audit evidence stay clear.

**Are close agents in scope for a SOX audit?**
If they support financial reporting controls, yes. Auditors will test access, change management and operations around them.

**Should the model calculate accruals?**
No. Let deterministic code or the ERP calculate; the agent gathers inputs, explains and drafts.

## Key facts

AI agents in the financial close fall within internal control over financial reporting (SOX Section 404; auditors use PCAOB AS 2201). Pattern: agents prepare reconciliations, accruals and draft journal entries; humans approve; systems post. ITGC mapping: access (agent's own identity, no posting rights), change management (prompts, tools, model versions reviewed and frozen in close window; pin model versions), operations (log inputs, tool calls, calculations, outputs, approver). Calculations should be deterministic, not model-generated. Approver must not be the agent's builder (segregation of duties).

## Related

- [What SOX IT general controls evidence do AI agents need?](https://opencontroller.lyzr.com/blog/sox-itgc-evidence-for-ai-agents.md)
- [Should AI agents do financial calculations themselves?](https://opencontroller.lyzr.com/blog/should-ai-agents-do-financial-calculations.md)
- [How do you separate who builds an AI agent from who approves it?](https://opencontroller.lyzr.com/blog/segregation-of-duties-for-ai-agent-approvals.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
