# What should an AI agent be allowed to do without asking?

Anything a person can check and undo in time. Reading, sorting and drafting need no approval, because nothing leaves the building. Acting inside your rules can run alone, with an undo window. Anything that cannot be undone waits for a person: new payees, deleted records, signed terms. The line follows the cost of a mistake.

By Surinder Thakur, AI Product Designer and Builder. Published 17 Nov 2026. Topic: [Trusting AI with money](https://www.surinder.design/work?type=trusting-ai-with-money).

## Which jobs can run without approval?

The ones where a mistake stays inside the building. Reading invoices, sorting a support inbox, drafting a reply for a person to send.

An approval step here only slows the team down. People start clicking approve without reading, so the check stops meaning anything.

## When should it act alone, with an undo?

When the action follows a rule you wrote and can be reversed in minutes. Paying an approved supplier under a limit. Booking a slot that is held, not confirmed.

The agent acts, then waits 10 minutes before it is final. The owner sees it at once and can cancel with 1 tap.

## What must always wait for a person?

Anything that cannot be taken back, or that changes the rules themselves. A new payee, changed bank details, a deleted record, a signed contract.

The agent prepares the decision and shows its reason. A person decides, because the cost of a wrong guess lands on them.

## How many approvals is too many?

More than a person reads properly. When approvals arrive every few minutes, people approve on reflex, and nothing gets checked.

Put the stops only where a mistake costs real money or trust. A few approvals that get read keep the agent safer than many that do not.

## In 1 picture

1 question sorts every action an agent takes: can it be undone in time.

Can it be undone in time?

- Yes (inside your rules): Draft a reply, Sort the inbox, Pay a known supplier. The agent acts.
- No (cannot be taken back): Add a new payee, Delete a record, Sign new terms. A person decides.

## Next questions

- [How do you stop an AI agent making a mistake with money?](https://www.surinder.design/work/ai-agent-mistake-with-money): You limit what it can do alone. The agent pays only inside the rules you set: approved payees, a limit, a monthly budget.
- [How do you design an AI agent approval screen people actually read?](https://www.surinder.design/work/how-to-design-an-ai-agent-approval-screen): Show 4 things on 1 screen: the action, the exact amount, the reason, and the rule that stopped it.
- [What happens when an AI agent fails halfway through a task?](https://www.surinder.design/work/when-an-ai-agent-fails-halfway): A well-built agent stops, keeps what it finished, and recovers in 1 of 3 ways.
- [What is agentic UX, and why does it matter?](https://www.surinder.design/work/what-is-agentic-ux): Agentic UX is design for software that acts on a person's behalf.

## Questions that lead here

- [Can an AI agent take B2B orders from WhatsApp, email and phone calls?](https://www.surinder.design/work/can-an-ai-agent-take-b2b-orders-on-whatsapp): Yes, when it reads the same live data your sales reps do.
- [What is an AI agent, in plain words?](https://www.surinder.design/work/what-is-an-ai-agent): An AI agent is software that works toward a goal on its own.
- [What is the difference between an AI copilot and an AI agent?](https://www.surinder.design/work/ai-copilot-or-ai-agent): A copilot suggests and you act. An agent acts and you check.
- [How do you stop an AI agent from being tricked by a hidden instruction?](https://www.surinder.design/work/how-to-stop-prompt-injection-in-an-ai-agent): You cannot fully stop it, so you limit the damage.

Page: https://www.surinder.design/work/what-should-an-ai-agent-do-without-asking
