glad.Add to Slack

Product

AI agent approval before action: a yes in the Slack thread

AI agent approval before action is how Glad works by default. Glad asks for a yes in the thread before any action that changes data or sends something, except private changes that only the requester sees, such as a draft. Reads never ask. The approval card shows exactly what will happen, the person who asked clicks Approve or Deny, and every decision is logged. It works on every plan.

This is human in the loop for AI in Slack: a named person, in the thread where the work was asked for, deciding on the exact action before it runs.

Last updated 9 October 2026

Example

ThreadExample
S
Sam
@Glad email Dana at Northwind the renewal quote from this thread.
GladAPP
Approval needed
Gmail, send email
To
dana@northwind.example
Subject
Your renewal quote
Text
Hi Dana, as promised, here is the renewal quote for next year...
GladAPP
Approved by Sam. Glad sends the email and replies in the thread.
Example only: an approval card as it could look in a Slack thread.

How it works

  1. 1

    You ask in a thread.

    For example: "@Glad email Dana the renewal quote from this thread."
  2. 2

    Glad reads and prepares.

    Searching, looking things up and reading a thread run without asking. Glad then prepares the action.
  3. 3

    Glad posts an approval card in the same thread.

    The card shows what will be sent or changed: the recipient, every Cc and Bcc, the subject, the start of the text, and other details such as the repository or channel. Attached files show by name, type and size. The card is posted in the thread, so people in the channel see what will be sent. That is the point: the team sees it too.
  4. 4

    The right person decides.

    The person who asked approves actions on their own connections. For a connection an admin shared with the team, an admin approves. Anyone else who clicks gets a private note: "You are not allowed to approve this."
  5. 5

    Glad runs exactly that action, once.

    The card turns into one line, such as "Approved by Sam.", and Glad reports back in the thread. If you click Deny, nothing runs. If nobody answers, the approval expires and nothing runs.

What you can do

  • Approve or Deny on the card in Slack. Approving or denying costs no task.
  • Approve and always allow (admins). This button appears when the approver or the requester is an admin. That tool then runs without asking.
  • Manage the always allowed list in Settings, Approval rules. Each tool shows who added it and when. Remove it and the tool asks first again.
  • Pick which tools Glad may use, per app. Glad can only use tools that are on. Each tool is tagged read, write or risky ("Can't be undone"), and tools that ask first are marked.
  • Review the record on the Approvals page in the console: tabs for All, Waiting and Decided, with the action (for example "Gmail: Send email"), the status, who asked, and who decided and when.
  • Find it in the audit log. Each request for approval, each yes or no, each expiry and each tool run is logged.
  • Get asked before bigger jobs. A /deep run counts as 5 tasks, and Glad asks before it starts one. Admins also choose how far a bigger job may run before Glad asks the requester to continue.

Who controls what

The person who asked approves or denies actions on their own connections, and sees their own approvals, with the details, on the Approvals page.

Admins approve actions on shared connections, set a tool to always allowed and take that back, turn tools on and off, and see every approval in the workspace. Each always-allow change is in the audit log.

Rules nobody can change:

  • Glad cannot approve its own action or add a tool to the always allowed list.
  • A message from Glad to anyone but you, or a post in a channel, asks first every time. Always allow never applies to it. A direct message to yourself runs without asking.
  • An approval covers the action on the card, once. It cannot be reused, and only the first click counts. A late click gets: "This approval was already decided or has expired."
  • In a channel shared with another company, only your own workspace's members can ask or approve.
  • A deactivated member's decision no longer counts.

How other tools handle it

All facts below come from each vendor's own documentation, checked on 2026-10-09.
GladWhen it asks before actingAny action that changes data or sends something, except private changes; reads never askWho approvesThe person who asked; an admin for a shared connectionAdmin choices per toolAlways allow; tools on or offPlanEvery plan
ViktorWhen it asks before actingA fixed list of actions, plus any tool an admin sets to askWho approvesNot stated beyond "someone"Admin choices per toolOff, Run automatically, Ask for confirmationPlanEvery plan
LindyWhen it asks before actingWrites on integrations set to Require approval, in shared Slack threadsWho approvesThe person who asked, or any workspace owner or adminAdmin choices per toolAlways allow, Require approval, Don't offerPlanEvery plan
SlackbotWhen it asks before actingActions you set to Needs approvalWho approvesNot statedAdmin choices per toolAlways allow, Needs approval, Blocked, CustomPlanNot stated
@ChatGPTWhen it asks before acting"asks for your approval before taking action in connected tools"Who approvesNot statedAdmin choices per toolNot statedPlanBusiness, Edu, Enterprise
Where others are better today: Viktor lets an admin set any tool to "Ask for confirmation", including one that would otherwise run, and its approvals also work in Microsoft Teams; Glad has no setting to make a tool ask. Lindy lets any workspace owner or admin approve on someone else's request, where Glad lets an admin approve only on shared connections. Slackbot lets you block a single tool inside Slack's own settings.

Why it matters

The EU AI Act describes human oversight for high-risk AI systems. Article 14 says they "shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which they are in use." It lists the ability "to decide, in any particular situation, not to use the high-risk AI system or to otherwise disregard, override or reverse the output", and warns about "automatically relying or over-relying on the output" (automation bias).

This page does not say Glad is a high-risk system, and it is not a compliance claim. The rule and the approval card point the same way: a person sees the action before it takes effect and can stop it, and the Approvals page keeps the record.

Limits

  • No "force ask" setting. An admin can let a tool run without asking, but cannot make a tool ask that would otherwise run, such as a read.
  • One yes is enough. There is no second approver.
  • Approve or Deny only. There is no edit-and-approve button. To change the action, deny it and ask again.
  • Slack only. You cannot approve in the console or by email. Microsoft Teams is planned.
  • No standing approval per channel.
  • The card is visible in the thread. Everyone in the channel sees what will be sent. Keep that in mind before asking for a sensitive send in a busy channel.
  • Approvals take attention. Many cards in a day can tire people. Today the only way to ask less is for an admin to always allow a tool.

Questions

What is human in the loop for an AI agent in Slack?

It means a person decides before the AI acts. In Glad, the AI agent prepares the action, posts it as a card in the Slack thread, and runs it only after the right person clicks Approve.

Which actions need approval in Glad?

Any action that changes data or sends something, such as sending an email, updating a deal or filing a ticket. Private changes that only the requester sees, such as a draft, run without asking, and reads never ask. An admin can set a tool to always allowed.

Who can approve an AI action in Glad?

The person who asked, for actions on their own connections. An admin, for actions on a connection shared with the team. Anyone else gets a private note that they are not allowed to approve.

What happens if nobody approves?

The approval expires and nothing runs. The run is free. Ask Glad again when you are ready.

Can an admin turn approvals off?

Only per tool, with "always allow", and an admin can undo it at any time. It never applies to messages Glad sends to other people or posts in a channel. Glad cannot allow itself anything.

Does approving cost a task?

No. Approving and denying are free. A task counts only when Glad uses a connected tool or makes a document.

Keep reading

See it in a real task

Try Glad in your Slack.

50 free tasks. No card needed.

Add to Slack