---
title: Should AI agents be restricted to business hours or change windows?
url: https://opencontroller.lyzr.com/blog/time-based-restrictions-for-ai-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# Should AI agents be restricted to business hours or change windows?

**Restrict by action, not by agent. Read-only and low-risk actions can run any time. Actions that change production systems, move money or contact many customers should run only when someone who can stop them is on duty, or inside an approved change window. The rule that matters is coverage: no high-impact action without a person able to respond.**

Agents do not get tired, which is why teams like running them overnight. It is also why an agent fault at 3 a.m. can run for hours before anyone notices.

## Why do time windows reduce risk?

Time windows do not make an action safer. They make the response faster.

- **Detection.** Alerts are seen and acted on during staffed hours.
- **Containment.** Someone can block or quarantine the agent within minutes.
- **Downstream teams.** Payment, infrastructure and customer teams are available to reverse effects.

Access control frameworks already allow this. NIST [SP 800-53 Rev. 5](https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final) includes control enhancement AC-2(11), Usage Conditions, which lets an organisation enforce conditions such as time of day on how accounts are used. Agent identities are accounts.

## Which actions should be time-restricted?

Use this decision table as a starting point.

| Action type | Example | Recommended window | Why |
|---|---|---|---|
| Read-only, internal | Summarise tickets, search documents | Any time | Low impact if wrong |
| Customer-facing answers | Support chat | Any time, with human handoff path staffed or a clear fallback | Customers expect 24/7, but handoff must work |
| Bulk outbound communication | Email or SMS to many customers | Business hours in the recipient's timezone | Errors reach many people; recipients' local rules may restrict contact hours |
| Production infrastructure changes | Scale, restart, modify config | Approved change windows only | Aligns with existing change management |
| Financial actions above a limit | Refunds, payments, transfers | Business hours with finance on duty | Reversal needs people |
| Data deletion or bulk updates | Purge records, mass edits | Change windows, with backup verified | Hard to undo |
| Security response actions | Disable accounts, isolate hosts | Any time, but alert on-call immediately | Speed matters more than timing |

During a declared freeze, tighten further. See [change freezes for AI agents](/blog/change-freezes-for-ai-agents).

## What timezone mistakes catch teams out?

- **Server time versus business time.** A window defined in UTC on a server may not match the business owner's hours. Define windows in a named timezone.
- **Daylight saving changes.** Fixed UTC offsets drift by an hour twice a year in many regions.
- **Recipient timezone.** For outbound contact, the recipient's local time matters, not yours.
- **Global teams.** "Business hours" for a follow-the-sun team may be 24 hours. Then the real rule is on-call coverage.
- **Public holidays.** Calendars differ by country. A window that is staffed on a weekday may not be staffed on a local holiday.

## What is the on-call coverage rule?

Write it plainly: an agent may perform a high-impact action only when a named person who can stop it is on duty and reachable.

That means:

1. Each high-impact agent has an on-call owner in your paging tool.
2. Alerts from that agent route to that rota.
3. The on-call person has the access to block or quarantine the agent, and has practised it. See [test your AI agent kill switch](/blog/test-your-ai-agent-kill-switch).
4. If coverage lapses, high-impact actions pause until it returns.

## How should exceptions work?

Some work must run out of hours, such as month-end processing.

- **Pre-approve the window.** Name the date, time, scope and owner.
- **Raise monitoring.** Lower alert thresholds during the exception.
- **Cap volume.** Set limits on records changed or value moved.
- **Record it.** Keep the approval with the agent's change history.

Risk tier decides how strict to be; see [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents). For agents that move money, also see [controlling AI agents that can move money](/blog/control-ai-agents-that-can-move-money).

## How Opencontroller does this

Opencontroller supports policy as code that is extensible, per-agent Block and Quarantine controls that act through the connected platform, and behavioural-deviation monitoring with alerts and automatic quarantine and kill. Enforcement depth depends on the connector and traffic path. See the runtime pillar on [stopping a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent). [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should all AI agents be limited to business hours?**
No. Restrict by action. Read-only and customer-answering agents can run around the clock; high-impact actions need someone on duty.

**Is a time window a substitute for approval or limits?**
No. Windows speed up response. Keep value limits, approvals and kill switches as well.

**How do you handle global teams?**
Replace business hours with an on-call coverage rule tied to a named rota and timezone.

## Key facts

AI agent time restrictions should be set by action type: read-only actions any time; bulk outbound contact, production changes, large financial actions and bulk deletions only in staffed hours or change windows. NIST SP 800-53 AC-2(11) Usage Conditions allows time-of-day conditions on account use. Opencontroller provides extensible policy as code, per-agent Block and Quarantine, and behavioural-deviation monitoring. Licensing starts from $250K per deployment per year.

## Related

- [Should AI agents follow change freezes?](https://opencontroller.lyzr.com/blog/change-freezes-for-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [How do you test that your AI agent kill switch actually works?](https://opencontroller.lyzr.com/blog/test-your-ai-agent-kill-switch.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
