ChatGPT Agent With Slack: Turn Messages Into Reviewed Work

ChatGPT Work can connect to Slack through a connector, giving the agent access to read messages in authorized channels and extract useful information — action items, decisions, or summaries — from those conversations. The key thing to get right is the review step: a Slack message is not an approved instruction, and acting on it without review can turn a casual comment into a committed task.

Getting useful work out of Slack conversations means deciding what the agent can see, what it extracts, and what happens to the extracted items before anyone acts on them. Tools like fenbs hold the team's agreed tasks and decisions in a place every assistant can reach, separate from the Slack channel where the conversation happened.

For other ChatGPT Work agent capabilities, see ChatGPT agent mode. For agent costs and limits, see ChatGPT agent kit.

What ChatGPT Work can do in Slack

Once the Slack connector is active, the agent can read messages from channels and direct messages it has been authorized to access. It can summarize threads, list mentioned items, and identify statements that sound like action items or decisions. With posting permissions, it can also reply to threads or post to channels.

What it cannot do: it does not have access to channels it was not authorized for, even if those channels exist in the same workspace. It cannot access Slack's internal metadata such as message reactions or membership data beyond what the OAuth scope covers. Permissions are specific to what you approved at connection time.

How to connect Slack to ChatGPT Work

1. In ChatGPT Work, open connector settings and find Slack in the available connectors.
2. Click Connect and complete the Slack OAuth authorization. Slack shows you which workspace and which scopes the connector is requesting before you approve.
3. Choose the smallest set of channels the agent needs. If you want it to read a specific project channel, authorize that channel. Avoid granting access to your entire workspace if it only needs a few channels.
4. Decide at this stage whether to grant posting permissions. If you only want the agent to read and extract, read-only is safer to start.
5. Test with a request to summarize a recent thread in one channel. Check that the summary is accurate and that the agent only accessed the channel you authorized.

Read access vs post access: what to grant

Read access lets the agent summarize conversations, extract action items, and answer questions about what was said. The risk is low as long as you treat the output as a draft, not a final list of tasks.

Post access lets the agent reply to threads or send messages to channels. This is useful for brief status updates, but a posted message is visible to everyone in the channel. Make sure the agent's messages are clearly marked as AI-generated and that you review any post before sending it to a public or customer-facing channel.

You can revoke either level of access at any time from the Slack workspace's third-party app settings, or from the connector settings in ChatGPT Work.

A Slack message is not an approved instruction

This is the most important principle for a Slack-connected agent. A team member writing "we should update the pricing page" in a project channel is not the same as someone opening a task and assigning it. Slack conversations are informal. Treating every mention of work as an actionable instruction leads to tasks that nobody actually agreed to.

Review every extracted action item against your team's actual task record before adding it as a task. Ask the person who made the comment whether they intended it as a request, an idea, or just thinking out loud. Record it as a task only when someone has actually confirmed it.

Similarly, a Slack message that sounds like a decision — "let's go with option B" — is not a recorded decision until someone writes it down in the right place and the right person confirms it.

Building in this review step takes a minute per conversation and prevents a lot of work being done on instructions nobody formally approved. Treat the agent's extraction as a first pass that saves you from reading every message, not as an authoritative task list.

How to do this with fenbs

1. Create a free account and set up a project for the work that comes from your Slack channels.
2. After each relevant Slack conversation, ask the agent to list what it sees as action items and decisions. Review each one before accepting it.
3. Add confirmed tasks to the kanban board in fenbs. Record confirmed decisions in Decisions and rules.
4. Connect ChatGPT Work to fenbs over MCP from /integrations. The agent reads the agreed tasks and rules from fenbs, not from Slack, before it starts work.
5. Ask the agent to save progress in fenbs before the session ends so the next assistant starts from the current state of the work.

For how ChatGPT Work compares to Claude for collaborative work, see ChatGPT Work vs Claude cowork. For Microsoft Copilot agents in a similar pattern, see Microsoft Copilot agents.