Can an AI agent run the month-end close?

It can run much of the work, not own the outcome. An AI agent can coordinate the close checklist, chase task owners, prepare reconciliations, accruals and flux reviews, and report what is outstanding. The financial controller owns the close: they approve judgemental entries, review the reconciliations, lock the period and sign off the result.

What does "running the close" actually mean?

The month-end close is a sequence of dependent tasks: cut-off for sales and purchases, accruals and prepayments, payroll posting, depreciation, intercompany, reconciliations, revaluation, review, lock, report. Each has an owner and a deadline. Much of the controller's effort goes into tracking who has done what and chasing what is late.

An agent can take two roles. As a preparer, it does individual tasks: the reconciliations, the standard journals, the flux review. As a coordinator, it tracks the checklist, starts tasks when their inputs are ready, and reports progress. Coordinating other agents is depth 5 work, reserved for orchestrators, and it does not change what each task agent may do. A coordinating agent cannot give a task agent more authority than that agent has earned.

What conditions decide how much it takes on?

• A written checklist. Every task, owner, dependency and deadline, as in the month-end close checklist with agent tasks.
• Each task's own rules. Posting accruals, reconciling cash and running depreciation each have their own conditions and depths. The close does not override them.
• A clear split between preparation and review. Every task the agent prepares has a named reviewer.
• Period controls in the ledger. The agent cannot lock or reopen a period.

What does the agent prepare, and where does the controller sign off?

What the agent prepares is set out more fully in month-end close: what the agent prepares. The sign-off rows at the bottom are the ones that never move. Locking the period ends the controller's chance to correct it, and signing the close says the numbers are ready to report.

An illustrative example

Take a software company with a five-day close. On day one the agent checks the payroll file has posted, runs the month-end accruals it is allowed to post and drafts the three judgemental ones, and starts bank reconciliations as the bank feed closes.

On day two it reports that revenue cut-off is incomplete: the billing team has not confirmed three contracts that started on the last day of the month. It chases. Billing confirms two and corrects the third. On day three it finishes all 180 reconciliations; reviewers sign them according to risk rating, and the controller signs the six high-risk ones. Its flux review explains every movement above threshold except one, a jump in travel costs, which the controller traces to a sales conference.

On day four the controller approves the remaining entries, locks the period and signs the close.

What controls make it safe?

• Task rules unchanged by the close. Each task's depth applies whoever starts it.
• Coordinator authority limited. It can start and chase tasks; it cannot approve them.
• Period lock with a person.
• Close summary as evidence. Every task, owner, reviewer and time, in one record.
• Overruns reported, not hidden. A task late against the timetable is shown as late.

Where Pryme fits

In Pryme, the coordinating agent is an orchestrator certified to depth 5, the depth coordination needs, and each task agent keeps its own certificate and depth. The minimum chain applies to every action, so a coordinator's request never lifts a task agent above its own bounds. Every task verdict is filed.

Payroll often sits at the start of the close and has its own rules, covered in whether an AI agent can run payroll. The wider finance controls are in AI agents for finance: books you can audit.