# How do AI agents work in the month-end close?

- Canonical: https://hellofig.io/answers/ai-agents-for-month-end-close
- Updated: 2026-11-13
- All answers: https://hellofig.io/answers.md
- About Fig: https://hellofig.io/fig.md

**In short.** In a close run with agents, agents prepare and people post. Agents match reconciliations, propose accruals and entries from evidence, draft the flux commentary and assemble the close package, each with its support attached. People review the exceptions and approve every entry before it touches the ledger, and the audit trail records each step as it happens.

## Where agents work in the close

| Close step | What the agent prepares | What people decide |
|---|---|---|
| Reconciliations | Matches statements to the ledger on the team's rules and explains each exception from prior periods and documents. | Which exceptions are resolved and how. |
| Accruals | Proposes entries from open orders, contracts and recurring invoices, with the evidence on each line. | Whether each estimate stands. |
| Intercompany and allocations | Finds the mismatches between entities and applies the approved allocation method with its drivers. | The corrections and any change to method. |
| Journal entries | Drafts each entry with support attached and routes it for approval. | Approval of the exact entry before it posts. |
| Variance and flux | Computes variances above materiality and drafts the explanation from the ledger and operating data. | The final story leadership reads. |
| The close package | Assembles the package with every number traced to its source. | Sign-off. |

## The rule that makes it safe

An agent may read from the systems of record, prepare a proposal with its evidence and route it. It may not post to the ledger or send anything outside the company on its own. Approval is bound to the exact entry, and the approver's authority is checked again when the entry executes.

## What the audit trail has to capture

For every agent action: the inputs it read, the model and version it ran, the output it produced, the human review and an immutable history. Captured as the work happens and kept with the workpapers, that record is what lets a controller and the auditors rely on the close.

## Where to start

Start with one reconciliation that runs late every month. Write down what the agent may read and what a person must approve, run it beside the manual process for one cycle, and compare the exception lists. Then move to accruals. Fig builds this on CosmOS with the company's ERP and banks connected in place, so each step reuses the last one's connections and controls.

## Questions

### Can AI agents post journal entries?

They should prepare and route them. A person approves the exact entry before it posts, and authority is rechecked at execution.

### How much faster is the close with agents?

It depends on where the time goes today. Keeping the books current every day moves most preparation out of close week, so the close becomes review and judgment.

### Is an AI-assisted close compliant with SOX?

It can be, when the controls are designed in: approvals bound to entries, segregation of duties, and an audit trail of every agent action retained with the workpapers.

### Which close step should we start with?

The reconciliation that is late every month. It is a matching problem with exceptions, easy to check against the manual result.