AI agents

How to make an AI agent ask before it acts

The short answer

Sort its actions into "does alone" and "asks first", and use a tool that holds sends, payments, deletions and access changes until you approve. Each approval should show exactly what will happen and wait for an answer.

4 min read

Should this action ask first?

Does the action change anything, or only read?
AI agent

Let it run alone

Reading email, searching a CRM or a web page breaks nothing. It shows up in the log.

Does it move money, delete, or change access?
AI agent

Always ask, whatever else you loosen

Payments, refunds, deletions, shared files and new people are hard to undo. They wait for a yes, every time.

Does it reach anyone outside your business?
AI agent

Ask first, with the full text shown

A sent email cannot be unsent. Approve it until you approve that kind unchanged for weeks.

AI agent

Ask first at the start, then loosen

Changing a record in your own apps can be undone. Once you approve it unchanged for weeks, let it run alone.

Why "ask me first" in the prompt is not enough

Tell a model to check with you before anything risky and it usually will. But prompt text guides the model and enforces nothing. For anything that moves money, a person should approve each call in the software.

The test: if the AI agent skipped the question, would the action still go through? If yes, you have a polite request, not an approval.

What a good approval shows

  • Who is asking: which AI agent, on which run.
  • Exactly what will happen: the recipient and full text, the record and new value, or the amount and who gets it.
  • Why: one line of its reasoning, so you can spot a wrong premise.
  • A way to change it, not only yes or no.
  • It stays open: an unanswered approval waits; it never expires into "yes".

How to make an AI agent ask before it acts

  1. Step 1: List every action it can take

    From its apps: read, create, update, send, delete, share. You will be surprised how many it has.

  2. Step 2: Put each one in a group

    Does alone, asks first, or always asks. Use the chooser above. When in doubt, put it in asks first; you can move it later.

  3. Step 3: Remove what it never needs

    An action it never uses should not be there at all. Fewer actions means fewer approvals and less to go wrong.

  4. Step 4: Pick one place for approvals

    One place you check a few times a day, on your phone too. Approvals scattered across email, chat and dashboards get missed.

    Watch out. Fifty approvals a day and you stop reading them. Trim what asks, and group low-risk items into one daily report.

    Do this step in 50agents

    Set up an AI agent that finds my HubSpot deals with no reply in 5 days every weekday at 9:00 and drafts a follow-up email for me to approve

    Build it

    50agents puts every approval in one place, Needs you. Anything that spends money, reaches people outside your workspace or changes access asks first, and stays open until you answer. See the ready-made version.

  5. Step 5: Watch the approve-without-edits rate

    If you approve one kind of action unchanged for a few weeks, consider letting it run alone. If you edit most, fix the instructions.

  6. Step 6: Plan for retries

    An approved action can run twice if the tool crashes in between. Ask your tool how it makes approved actions safe to repeat.

What can go wrong

  • Approving without context

    The card says "Send email?" without the text. Never approve what you cannot see in full.

  • Nobody answers

    The approval sits for a week and the moment passes. Check at fixed times, and give the AI agent a rule for stale items.

  • The AI agent decides when to ask

    If asking is up to the AI, one day it will not ask. Put the rule in the tool's settings, not only in the prompt.

  • Approved twice

    A retry after a crash sends the same email again. Prefer tools that record each approved action once.

Questions people ask

What is human in the loop for AI agents?

A person approves certain actions before they happen. It differs from reviewing a report after the fact, which some call human on the loop.

Can I approve from my phone?

Pick a tool where you can. Approvals that wait for a desk slow the AI agent down to your desk hours.

When can I let it act without asking?

When you have approved that kind of action unchanged many times, and a mistake would be cheap and easy to undo. Never for payments, deletions or access.

What if nobody answers an approval?

It should wait. An unanswered approval never expires into a yes. Pick one place to check a few times a day, and let someone else answer when you are away.

Which actions should always ask?

Anything that sends outside the business, moves money, deletes or changes who has access. Reading and drafting can run alone.

Can the AI agent decide when to ask?

It should not. If asking is written only in its instructions, it can skip the question. The approval must be a step the software enforces.

Read next

Or let Captain build it

One sentence to Captain. Free to start.

Build it with Captain
How to make an AI agent ask before it actsBuild it