Skip to content
be-human.ai
← Back to the journal

Tools & integrations

Slack AI Agent Permissions: Channel Scopes and Named Approvers

Learn how to limit which Slack channels an AI agent reads and writes, which actions need a named approver, and how to set both up step by step.

Be Human typographic cover with the title: Slack AI Agent Permissions: Channel Scopes and Named Approvers
Conceptual illustration created with AI · Be Human

To limit which Slack channels an AI agent reads and writes, grant its Slack app only the channel memberships and OAuth scopes the job needs, keep read permissions separate from write permissions, and send every action that changes a record, posts to a customer or spends money to a named human approver before it runs. Most problems come from one broad permission that nobody revisits. The sections below show how to set these layers in order and how to review them later.

Key takeaways

  • Start with one channel and one job. An agent that can read every channel can summarize, leak or act on information it was never meant to touch.
  • Treat reading, drafting and writing as three separate permissions, each granted and revoked on its own.
  • Name a specific person as approver for each write action, with one backup, and record every decision in the thread or a linked log.
  • Check the Slack app's scopes and channel memberships on a fixed schedule, not only when something breaks.
  • Write down what the agent must never do in each channel, so the rule exists outside the prompt.

Why channel access comes first

A Slack agent does not browse your workspace the way a person does. It reads what its app is allowed to see, responds where it has been invited, and acts through the permissions its app received at installation. That makes channel membership and app scopes the real control surface. A prompt that says "do not share pricing" is a request. A missing scope, or a channel the app was never added to, is a boundary the model cannot argue its way past. Authorization belongs in the systems around the model, a principle set out in the [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html). For the broader idea, see Least privilege for AI agents: start with the smallest access.

This matters because Slack workspaces accumulate history. Support threads, finance discussions, hiring notes and customer escalations often sit in the same workspace, sometimes in channels whose names say little about what is inside. Before you decide what the agent should do, decide what it is allowed to see. The rest of the design follows from that decision.

Sort channels into three zones

Before you change any setting, list every channel the agent might connect to and sort each one into a zone.

  • Zone A, read-only context: channels whose content helps the agent answer questions, such as an approved internal FAQ or a product announcements channel. The agent may read these and never post in them.
  • Zone B, draft and suggest: channels where the agent may post drafts, summaries or proposed replies for staff to review. It does not send anything to customers or change external systems from here.
  • Zone C, act with approval: channels where a specific action is allowed only after a named approver confirms it. These are usually the channels that connect to a record, a payment, a refund or a customer-facing message.

Keep sensitive channels out of all three zones unless someone has written down a reason. Leadership, legal, payroll and HR conversations rarely belong in an agent's context, even as read-only input. If a channel's purpose is unclear, leave it out and revisit it at the first review.

Scope the Slack app before the agent sees a message

Access in Slack comes from three places: the app you install, the scopes that app requests, and the channels where it is a member. Review each one before the agent goes live.

Start with the scopes. Request the narrowest set that supports the zones you defined. A read-only zone needs history and read scopes. A draft zone needs posting permission only in designated channels. Zone C actions should go through a flow that asks a human before any write reaches an external system. Do not grant broad write scopes for convenience and plan to restrict them later, because later often does not happen.

Then control membership. An app generally sees only the channels it has been added to, and private channels require an explicit invitation. Add the agent to zone A and zone B channels by hand, one at a time, and record who did it. Do not add it to a catch-all channel or to a channel that a bulk process fills automatically.

Finally, check what the agent is allowed to produce. A draft channel should use a clear label in the message text, so nobody mistakes a suggestion for a commitment. For the difference between reading, drafting and sending, see Read, draft and send: three different AI permissions.

Decide which actions need a named approver

Not every action needs a human. Summarizing a public runbook or answering a question from an approved FAQ is low risk. Creating a refund, changing a booking, messaging an external party or updating a financial record is different, and the approval rule should say so.

Use this checklist for each proposed action:

  • Does it change data in another system, such as a CRM, billing tool, calendar or spreadsheet?
  • Does it reach someone outside the company, including customers and suppliers?
  • Does it commit money, time, inventory or a legal position?
  • Could it repeat by accident if the agent retries after a timeout?
  • Would a mistake be hard to reverse within an hour?

If you answer yes to any of these, the action needs an approver. The word "named" matters. "The manager" is not an approver if three people hold that role and nobody knows which one is on duty. Name one person and one backup, and make sure both are members of the channel where approvals happen.

The approval should attach to the exact payload. If the amount, recipient or date changes after approval, the action needs a new approval. Record who approved, what was approved and when. If the approver says no, the agent should stop, explain the decision in the thread and not retry with different wording. For Slack-specific patterns, see Slack Approval Buttons for AI Agent Actions: Decide Before Any Write.

A setup sequence you can follow

Use this order. It places the most sensitive decisions at the start, when they are easiest to make.

  • Step 1: Write one sentence describing the job the agent will do and the outcome that proves it worked.
  • Step 2: List every channel that job touches and sort each into zone A, B or C.
  • Step 3: Remove sensitive channels from the list, or document why each one stays.
  • Step 4: Name the owner of each zone, the approver for each zone C action, and a backup for each approver.
  • Step 5: Request the narrowest Slack scopes that support the zones, then install the app.
  • Step 6: Add the agent to zone A and zone B channels manually and record each addition.
  • Step 7: Run the agent in zones A and B only for a set period, reviewing its posts and the sources it read.
  • Step 8: Enable zone C actions one at a time, each with its approver, and check the first approvals against the written rule.
  • Step 9: Schedule the first permission review and write down who runs it.

Keep the sequence short enough that someone can follow it without a meeting. If a step is hard to complete, that usually means the rule is not yet decided, and the agent should wait.

Illustrative scenario: a support channel and a billing channel

This scenario is illustrative and does not describe a real company. A small services business wants an agent to help its support team answer routine questions and prepare refund requests. Support uses a public channel for questions and a private channel for escalations. Billing discussions happen in a separate private channel.

The team places the public help channel in zone A, so the agent can read approved answers. The escalation channel goes into zone B, where the agent can post a draft reply and a short summary for the support lead. The billing channel stays out of the first version, because refund decisions depend on account history that the agent cannot verify from Slack alone.

When a customer asks for a refund, the agent drafts a summary and a proposed amount in the escalation channel. It does not post in billing and does not contact the customer. A named support lead approves or rejects the draft. Only after approval does a separate, scoped step create the refund request in the payment system, using the approved amount. If the lead rejects it, the thread records the decision and the agent takes no further action.

The lesson is not the specific channels. Each channel has a purpose, each permission matches that purpose, and the write step waits for a named person.

How Be Human handles this in managed operations

Be Human designs the channel zones, Slack app scopes, approval rules and success criterion as part of the first engagement, so the agent starts with a narrow role and a written boundary. Managed operations then cover monitoring of what the agent reads and posts, recurring permission reviews, incident response for cases such as a misdirected draft or an expired connection, and controlled changes when the team wants to move a channel from zone B to zone C. Each change is recorded, so the business owner can see who approved it. This follows the governance emphasis of the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework): clear ownership, documented decisions and review over time. It does not guarantee that an agent will never make a mistake, which is exactly why the approval layer exists.

Schedule a permission review every quarter, and sooner after any staff change, incident or new integration. Check three things: who is a member of each channel the agent can read, which scopes the app currently holds, and whether each zone C approver is still with the company and still the right person. Revoke what no longer serves the job. A narrower agent that works is more useful than a broad one that someone has to watch. If you are not sure which channel should host the agent's first task, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent read private Slack channels? Only if someone with the right permission adds it to them, and the app holds scopes that allow reading that channel type. Keep private channels out of the first version unless there is a clear reason, a named owner and a written rule for what the agent may do with what it sees.

Should the agent have one approver or a group? Use one named approver for each write action, with one named backup. A group such as "the support team" is hard to act on because nobody knows who is responsible at a given moment, and an approval without a clear owner is difficult to audit later.

Does a prompt instruction count as a permission? No. A prompt guides behavior, but it is not an access control. Real limits come from Slack scopes, channel membership, the systems the agent can reach and the approval step that must complete before a write. Text inside a message can also try to override instructions, which is why the [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) recommends keeping these layers separate.

How often should we review an agent's Slack access? Review it on a fixed schedule, usually quarterly, and again after any staff change, incident or new integration. Check channel membership, app scopes, approver assignments and the log of recent write actions. Anything that no longer matches the written job should be removed rather than kept in case it is needed later.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →