How to run end-of-day operations checks with an AI agent

Give the agent the checklist and the read access, and keep the signature with the duty manager. Each evening it reads every control point from the systems themselves, marks each as complete, late or failed, chases owners of open items, and drafts the end-of-day sign-off with anything carried forward. It never reruns failed jobs or clears breaks; people do.

What are end-of-day checks, and who owns them?

Every operation that processes money or records has a set of things that must be true before the day can be closed. Batch jobs ran. Payment files were sent and acknowledged. Cut-offs were met. Suspense and clearing accounts are within tolerance. Queues are below their thresholds. Today's reconciliations are done or their breaks are logged. Held items are either approved or knowingly carried to tomorrow.

The duty manager or operations control lead owns the sign-off. Each control point has its own owner: the payments team for the file acknowledgements, finance for suspense balances, technology for the batch schedule. The checklist itself is owned by operations control and approved by risk.

What are the steps, and who acts at each?

1. Operations control: maintains the checklist. Each item names the system to check, the condition that counts as complete, the deadline and the owner.
2. Agent, depth 1: reads each control point from the system of record at its deadline: batch status from the scheduler, acknowledgements from the payment gateway, balances from the ledger, queue counts from the work queue system.
3. Agent, depth 1: marks each item as complete, late, failed or unable to check, with the evidence it read.
4. Agent, depth 3: chases owners of late and failed items with a templated internal message naming the item, the condition and the deadline.
5. Agent, depth 1: rechecks chased items after a set interval and updates their status.
6. Item owners: resolve failed and late items: rerun a job, resend a file, clear a break, or record why it cannot be done today.
7. Agent, depth 1: checks the held queue for its own and other agents' items still waiting, and lists them.
8. Agent, depth 2: drafts the end-of-day sign-off: every item's final status, every carried-forward item with its owner and reason, every exception.
9. Duty manager: reviews and signs the day closed, or escalates.

Why does the agent never fix what it finds?

Because the fix is usually a change to a live system. Rerunning a failed batch can post entries twice if the first run partly completed. Resending a payment file can pay twice. Clearing a suspense break means posting a correcting entry. Each of these needs a person who understands why the failure happened, and many need a second person to check: the approach in running four-eyes checks on operational changes applies.

Keeping the agent to reading and chasing also keeps its role clean. The sign-off says what was true at the end of the day, compiled by something that had no hand in making it so.

What does "unable to check" mean, and why does it matter?

Sometimes the agent cannot reach a system or cannot find the evidence it needs. That status must never be treated as complete. A missing acknowledgement file is not the same as an acknowledged one. The sign-off should show "unable to check" items prominently, and the duty manager should confirm each by another means before signing.

If the agent itself has failed that day, for example a certification check or a drift check, it drops at once to the highest depth the run cleared, or to depth 1 if it cleared none, as argued in a failed run should cost depth the same day. For the end-of-day agent, that means it continues to read and draft but stops chasing, and the duty manager knows to check its work more closely.

What should you measure?

• Items late or failed at deadline, by item and owner, over time.
• Items carried forward, and how long they stay open.
• "Unable to check" items, which show where the agent's access or the systems' evidence is weak.
• Time from last deadline to signed day.
• Problems found the next morning that the checks should have caught. The target is none.

Which controls leave evidence?

Each item read from the system of record. Evidence: the value read, the source and the time.

Chases recorded. Evidence: each message, recipient and response time.

Carried items named and owned. Evidence: the owner and reason for each.

Sign-off by a named duty manager. Evidence: the signed record, with what they saw.

The agent's own verdicts. Evidence: every action it took or attempted, with its deciding bound.

How does this relate to month-end?

Month-end close is a larger version of the same pattern: a checklist of conditions, owners, deadlines and a sign-off. Many items are the same, with more of them. The end-of-day agent's records give the month-end team a clean starting point, and the boundary between preparing and signing is the same; see whether an AI agent can run the month-end close.

Where does it usually go wrong?

The checklist is out of date. A new payment scheme goes live and no one adds its acknowledgement file. Review the checklist whenever a system or process changes.

Deadlines are set by the agent's schedule. The agent checks at 18:00 because that is when it runs, but the file is due at 18:30. Each item's check time must follow its deadline.

Chasing becomes noise. Owners receive the same reminder every night for known, accepted delays. Let the duty manager mark an item as accepted for a period, with a reason.

The sign-off is signed unread. A long, green report invites a quick signature. Put exceptions and carried items at the top. Certify the agent's reading of each control point on real past days, as train, shadow, certify, widen sets out.