Yes, the agent can; the real question is what it should do while your people are away. Overnight, it can do work that needs no approval, provided someone on call can stop it. Anything held for a person waits until morning, customer contact follows the same hours rules as staff, and some limits should tighten out of hours.
What decides what an agent should do overnight?
Three things the agent depends on, none of which is the agent itself.
People. Held actions need a named approver. If none is awake, the action waits, and a queue of held items builds up for the morning. More important, if something goes wrong, someone must be able to stop the agent within minutes.
Systems. Many core systems run batch jobs overnight: end-of-day posting, interest calculation, file exchanges. Data the agent reads may be mid-update, and some systems refuse changes during their windows.
Rules about customers. Collections contact, marketing and some service messages are restricted to reasonable hours in many markets, and your own policies usually go further. An agent that sends arrears reminders at three in the morning breaches them just as a person would.
What works overnight, and what waits?
| Overnight, the agent may | Waits for the morning |
|---|---|
| Reconcile the day's transactions and prepare breaks for review | Clearing breaks that need a person's judgement |
| Classify and route overnight emails and form submissions | Replies that would be held for review |
| Gather documents and pre-fill cases | Decisions on those cases |
| Prepare reports for the start of the day | Anything with a customer contact restriction |
| Execute routine work it is certified for, within lower out-of-hours limits | Actions above the out-of-hours limit |
The last row is a choice many teams make. An agent certified to issue refunds up to a set amount during the day might be limited to a lower amount overnight, because nobody can catch an error until morning. Depth is still earned on evidence; the out-of-hours limit is a tighter limit set by the business on top.
Quiet hours are a good fit for the boring work your first agent should do: reconciliations, matching, preparing the day's queues.
Who answers for the result?
The process owner, as during the day. Running around the clock adds two responsibilities: naming who is on call to stop the agent, and deciding which capabilities run at which hours. An agent running unattended with no named person able to stop it is not a decision anyone should sign off.
What does this look like in practice?
Take a payments firm's operations team that works 8 a.m. to 6 p.m. Overnight, its agent reconciles the day's settlement files against the ledger, prepares a list of breaks with the likely cause for each, and classifies the overnight inbox. It also processes routine card refunds already approved by the merchant, up to a lower overnight limit.
At 2 a.m. a settlement file arrives with half its usual records. The agent does not reconcile against a partial file. It holds that reconciliation, pages the on-call operations analyst as its brief requires for incomplete files, and carries on with other work. The analyst confirms the file is incomplete and asks the sending bank for a resend. By 8 a.m. the team has a prepared break list, a classified inbox and a note on the file problem.
The morning hand-over matters as much as the night. Staff need to know what was done, what is waiting, and what went wrong; our guide on training staff to work alongside an AI agent covers building that routine, and allocating work across the team covers who picks up what.
What controls make overnight running safe?
- An on-call person with authority and the means to stop the agent.
- Out-of-hours limits tighter than daytime ones where errors would go uncaught.
- Contact-hour rules enforced at the send action, not only in the brief.
- Freshness and completeness checks on files and batch data.
- System windows respected. The agent knows when target systems are mid-batch and waits.
- A morning summary of everything done, held and flagged overnight.
Overnight running usually comes at a later stage of deploying an AI agent, once daytime running has shown how the agent behaves.
Questions people ask
Does running overnight need its own certification?
Not separately, but test cases should include overnight conditions such as partial files and batch windows, so the agent's handling of them is proven.
What about agents covering several time zones?
Then "overnight" differs by region. Set contact hours and approver coverage per region, not per agent.
Should the agent wake people up?
Only for the events its brief lists, such as incomplete critical files or a stop condition. Everything else goes in the morning summary.