Should an AI agent ask for approval before it acts?
Set AI agent approval rules for sends, record changes and recurring work. Learn how Hops permissions differ and what to check after an action.

An AI agent should ask before a consequential action that hasn’t already been authorized within a clear scope. Sending an email, publishing a post, changing a customer record or spending money deserves a different rule from reading an authorized source or preparing a private draft.
You can get a customer reply drafted from the relevant history without approving every research step, while keeping control over the message that leaves your account.
“Help with customer follow-up” is a goal. “Send this reply to this recipient” authorizes a particular action. A useful approval makes that scope clear.
Set approval rules by what the action changes
Start with the effect of a mistake and how easily someone could correct it.
| Action | Sensible starting policy |
|---|---|
| Read an authorized source or prepare a private draft | Proceed within the granted access and task scope |
| Send an email or publish a post | Review the exact content and destination before sending, unless a narrow recurring action is already authorized |
| Update customer records or shared operational data | Review the affected records and changed fields; define limits for recurring updates |
| Spend money, change permissions or delete records | Require explicit authorization, with stronger controls for bulk or hard-to-reverse changes |
| Do something outside the request | Establish scope before proceeding |
These are policy recommendations. Check how your chosen tool enforces them.
Choose a reviewer with authority to decide. A marketer can approve an email’s wording without being authorized to promise a refund. Put that decision with the person responsible for it.
Separate one-time approval from standing permission
One-time approval suits a customer reply, public announcement or sensitive record update. Show the reviewer the exact content and proposed changes.
Scoped permission allows an action within stated limits. A team might authorize an internal report to be posted regularly to one channel. That doesn’t authorize publishing it on the company website.
For recurring work, record the allowed action, account, destination and limits. Decide what should cause the agent to stop, such as an unexpected recipient. Repetition shouldn’t expand the permission.
An agent choosing its steps and a fixed automation following a sequence both need authorization. Our AI agents vs. workflow automation guide explains the distinction.
How approvals work in Hops
Hops is the collaborative AI workspace that helps your whole team move faster. Your team and its AI agent work live in the same docs, slides and dashboards, and what one person teaches it, everyone can use. You only pay for the AI’s work.
Hops’ Collaborative AI™ capabilities let colleagues review the draft reply in the same workspace.
By default, Hops checks with you before it sends an email or updates a record in another tool.
Sharing a reply has its own check
When AI works in a team thread, the hard question is what it should share with whom. Hops uses only what the person asking has been granted. If an answer draws on something others in the thread can’t see, Hops holds it until the person who asked approves, so private data never reaches the team by accident.
Using a connected app and changing it are separate permissions
In Settings > Connections, your personal usage setting determines whether Hops may use a connection for you. The connection’s owner and editors can also set each enabled write action to Always ask or Never ask.
Enabling a connection doesn’t approve every write. Never ask removes that action’s write-approval prompt for authorized work on an otherwise usable connection; it doesn’t grant unrelated access.
Reads don’t require a write-action approval, but access and connection-use requirements still apply.
Read the scope on the write-approval card
A write-action card may offer Approve, or Allow once and Allow for me in this thread. Allow once permits one execution of the named action. Allow for me in this thread permits that same action on that connection, for that person, in that thread, for up to 24 hours.
Use the thread option only when you intend to allow repeated executions of that action, not just one message.
Published automations have bounded standing permission
An automation’s owner reviews its instructions and access before publishing. An allowed write can then run without a fresh write-approval card each time, within that automation’s published access. This doesn’t change the connection’s settings everywhere else.
Missing connection consent, disabled actions or revoked access can still block a run. Publishing an automation also doesn’t prove its next run succeeded.
A real example: checking a site before updating the tracker
On October 6, 2026, a teammate asked Hops to check our website repository against an existing audit tracker. Hops reported its findings and offered to update the tracker. The teammate then explicitly requested the update and said a broken link should already be fixed in the content system.
In its completion message, Hops reported updating 17 rows and leaving 18 rows added by someone else unchanged. It also said the broken-link entry was marked Done on the teammate’s report, explicitly noting that it hadn’t opened the content system to verify the fix.
Hops offered the update instead of making it, and changed the tracker only after the teammate asked. Because it labeled the link fix as the teammate’s word, the next person knew which item still needed a check. The same boundary works for a customer reply: research and draft first, then authorize the exact message and recipient when you’re ready to send.
Require a result after the approval
Before approving, check the content, destination and extent of the change. “Update the CRM” isn’t enough if the proposal affects hundreds of records. Ask to see which records, which fields and any exclusions.
After execution, retain a useful record:
- Permission: who requested and approved the action, when, and within what scope.
- Change: what was proposed and what actually changed, including any skipped items.
- Outcome: whether it succeeded, failed or stopped after a partial change, and how to correct it.
Check which of these your tools record for you, including any retention requirements.
An approved email can fail to send. A bulk update can stop halfway through. If a service reports an uncertain result, check the destination before repeating the action; a retry could otherwise duplicate a successful first attempt. Test a denial, an unanswered request and a failure before relying on recurring work.
Read more about dividing responsibility between people and AI, or check Hops pricing for plan details.
Put the draft in front of the right reviewer
Ask Hops to prepare a customer reply from sources you can access, then review it with your team before authorizing the next action. Start with 20,000 free credits.
Start freeFAQ
Does approving an action prove it succeeded?
No. Approval grants permission. Check the execution result and, when the outcome is uncertain, the destination before retrying. A successful approval and a successful action are different events.
Does an AI agent need approval before reading data?
Reading usually doesn’t need the same approval as sending or changing something, but it still requires authorized access. In Hops, connection-use consent can apply even though reads don’t need a write-action approval.
Does Allow for me in this thread approve every Hops action?
No. On a write-action card, it permits the named action on that connection, for that person in that thread, for up to 24 hours. Other actions and connections aren’t covered.
Can Hops automations act without a fresh approval each time?
Yes, for actions included in the automation’s published access. That is standing permission for the configured work. Missing consent, disabled actions or revoked access can still prevent a run.
Author
Hops Team