If your deals already live in Slack, the question is not whether to use AI but how to use ChatGPT in Slack without copying a deal into a second window and losing half the context on the way. There are two ways to do it. They need different plans, different people switch them on, and only one puts the answer in front of your whole deal room.
Two different things are both called ChatGPT in Slack
The first is the ChatGPT app for Slack, a dedicated sidebar reachable from your app list or the navigation bar. It is yours alone: OpenAI's help documentation states that chats you start there "cannot be viewed, modified, or deleted by other users on Slack." You use it to search channels, messages and files, summarise long conversations into notes with owners and action items, and draft replies without leaving the window.
The second is @ChatGPT in Slack and Microsoft Teams, announced at DevDay on 29 September 2026 and covered in our AI News item on it. This one is a participant: you tag it in a channel or thread and the reply lands where everyone can see it and build on it.
Private research versus a shared answer. That decides which prompts you write.
Which plan you need, and who has to turn it on
For the sidebar app, OpenAI's documentation says you need "a Plus, Pro, Business, Enterprise, or Edu ChatGPT account." Free is not on that list. It also states that "Slack is enabled by default for Business workspaces," while Enterprise and Edu accounts need Slack enabled and access granted first.
For @ChatGPT in channels the setup documentation is narrower: "This feature is available to Business, Edu, and Enterprise accounts." A ChatGPT workspace admin or owner starts setup in the admin console, and a Slack workspace owner or authorised app manager completes app approval. On Teams, an admin connects the tenant then installs the app through the Teams admin center. You cannot self-serve this one before a call.
One more note that matters. Admins can limit @ChatGPT to "Selected channels only" but, in OpenAI's words, "Channel restrictions apply to Slack channels. They do not restrict direct messages."
The workflow that pays for the setup: rebuilding deal state
A deal has run four months across a channel, a Slack Connect thread and six DMs, and you have to walk into a forecast call and say what is actually true. The sidebar app searches channels, messages and files you already have access to, and channel-name matching is fuzzy, so you can name a channel without the # symbol:
Search the acme-rollout channel and my DMs with anyone from Acme over the last 60 days. Build a deal state summary with four labelled sections: 1. What the customer has said, quoted in their own words. 2. What we have committed to, with who committed it and the date. 3. Open questions nobody has answered. 4. What you are inferring rather than reading. Do not move anything from sections 1 to 3 into section 4. If a date, a number or a name does not appear in the messages, write "not stated" instead of your best guess.
That fourth section is the point. A summary that blends what your buyer said with what the model assumes is worse than no summary, because you will quote it. The same split runs through our prompts for sales call notes. The pre-call version is shorter and gets used more often:
Search the Acme channels for everything since our last call on [date]. I speak to [name, title] tomorrow. Give me: - The three things that changed since that call, each with the message it came from. - Anything we promised that is still outstanding. - The two questions this person asked that we answered badly. - One question I should open with. Label every line as either read from a message or inferred by you.
What changes when the answer is public
@ChatGPT in a channel is a different instrument. The output is visible to the whole room, which makes it right when several people need the same version of events: a handoff, an escalation, a deal desk thread.
@ChatGPT Read this thread and write the handoff brief the implementation team needs: the customer's stated problem in their own words, what we sold, what we promised that is not in the order form, the two risks most likely to cause an escalation in the first 60 days, and the named owner for each open item. Where the thread does not answer something, say so rather than filling it in.
Keep it to internal rooms. A customer-facing Slack Connect channel is not where you want to learn what your assistant says about a commitment you made in week three.
What it cannot see, and what it cannot do
Permissions are inherited, not expanded. The documentation is explicit: "ChatGPT respects Slack's existing permissions - only the messages and files you already have access to in Slack are searchable." If procurement talked in a private channel you are not in, your summary is missing it and will not say so. The shared assistant, the setup docs add, "does not automatically inherit each requester's access to apps or private files."
Capabilities are reduced. In the Slack app, "deep research, agent mode, and model switching are not available." Semantic search also requires Slack AI Search on a supported Slack Business+ or Enterprise+ plan; without it you get keyword retrieval, which changes how specific your prompts must be.
There is no shared memory. Chats from the Slack app do not write memories, though replies may draw on memories you saved previously if that setting is on, and continuing a conversation on the web does not sync back to Slack.
Then the quota: heavy use can exhaust your per-user Slack API quota even when workspace-wide limits are nowhere near. And Slack is a third party, so your messages sit under its retention and residency rules. Our guide to what not to paste into chatbots covers that line.
Common mistakes
Reading a channel summary as the deal. It summarises the rooms you are in. The CRM and the order form remain the record.
Assuming the admin restriction covers DMs. It does not, and OpenAI's documentation says so directly.
Tagging it in a Slack Connect channel. Shared channels include the customer.
Expecting continuity with the web app. Conversations do not sync. Finish work where you started it.
Writing prompts that assume semantic search. Without Slack AI Search, "find the pricing concern" may return nothing where "find messages about discount, list price or budget" returns the thread.
Letting it name a stakeholder you have never spoken to. A model reading a channel lists whoever was copied in; coverage is a different question, handled in our guide to mapping the buying committee.
Frequently asked questions
Which ChatGPT plan do I need to use ChatGPT in Slack?
The ChatGPT app for Slack works with a Plus, Pro, Business, Enterprise or Edu account, and Slack is enabled by default for Business workspaces. Tagging @ChatGPT in Slack or Microsoft Teams channels is narrower: OpenAI's documentation states it is available to Business, Edu and Enterprise accounts, and an admin has to set it up.
Can ChatGPT read private Slack channels I am not in?
No. OpenAI's documentation states that only the messages and files you already have access to in Slack are searchable, and the shared assistant does not automatically inherit each requester's access to apps or private files. The risk is not exposure, it is a confident summary that silently omits the conversation you needed.
Does ChatGPT remember my Slack conversations?
Chats from the ChatGPT app for Slack do not write memories, although responses may incorporate memories you saved earlier if that setting is on. Continuing the same conversation on the web does not sync back to Slack, so each surface should be treated as its own workspace.
Where to start on Monday
Pick one live deal, run the deal state prompt against its channel, and read section four first. If the inferred list is longer than the quoted list, the channel is not where your deal is being decided, which is worth knowing before the forecast call. For a prompt built for your role, the free prompt builder takes ten seconds and the free plan includes 100 prompts.