We put Claude in Slack. Only the developers used it.
Two teams, the same experiment, the same silence
Two teams we met recently did the same thing. They connected Claude to Slack. Anthropic ships an official Claude app for Slack, so the hookup takes under ten minutes.
The developers used it. Nobody else did. The bot sat in a channel no one opened. Here is how one person, who runs both sales and operations, put it:
"AI is AI. Having to summon a bot every time just isn't how I work. So I don't, and I don't even open that channel. When I have something for AI, I open a chatbot app on the side."
Across both teams, usage among people in their late thirties and up was effectively zero. Not because they are lazy. For someone already losing to the volume of chat, being told to write a prompt inside that same stream is one more job, not one less.
The press release went out. The usage never started. This is not a training problem or a UI problem. It is the shape of the messenger.
An @mention is not free
The standard way to put AI into an existing messenger is to make a bot and have people summon it with @mention. Anthropic describes it exactly that way: anyone in the channel can call it with @Claude. It looks natural. It charges the user every single time.
- Decide whether this conversation is a place where AI belongs.
- Remember which bot to call, in which channel.
- Write a prompt while colleagues watch it happen live.
For a developer none of this is a cost. It is what they already do in a terminal all day. For everyone else it is a threshold.
Slack's own research shows the gap. In its June 2025 Workforce Index of 5,000 desk workers, 60 percent said they use AI. Only 23 percent had ever handed an agent work to finish on their behalf. The distance between using AI and giving it a job is most of the market.
There is a contrast worth noticing. The same people use a browser side panel AI heavily. The reason is simple. You give it a prompt and the work is done right there. It is not summoning. It is delegating.
Slack was built for people to talk to people. Bolting AI onto the side of that conversation does not make AI the point of the room. It only adds decisions the user has to make.
The layer you cannot fix: messages that flow past
You can shave down the awkwardness of summoning with better UX. Underneath it sits something you cannot shave.
Slack's unit is the message that flows past. The channel is a river and the messages are water. Letting conversation wash downstream was the right design for text ten years ago.
For AI to understand a team's work and act on it, you need the opposite. You need what stays.
- Why this decision was made
- Which conversation this to-do came out of
- What task this file belongs to
The information is not missing from Slack. It is all there. It is just scattered, with nothing connecting one piece to another. Which is why the best you can hand an AI is "read the last 500 messages in this channel and summarise." That is the same as handing a person a year of meeting recordings and asking them to find the point.
It is also why Slack search always disappoints. The problem is not the search engine. It is that the relationships between pieces of data were never defined. Once they are defined and maintained, search stops being search and becomes a question you can ask. Slack is a messenger that never built the container for those relationships.
So why is nobody leaving Slack?
Here we should be honest. The companies we met cancelled Notion to cut costs. Engineering kept Jira, legal bought its own AI tool. Through all of that pruning, Slack was the one thing nobody cut.
The reasons are not grand.
1. It integrates with everything else more easily than anything else. 2. It is the most familiar and the most known. 3. For its actual job, chat, it is perfectly fine.
Nobody stays because they love it. They stay because the reason to move is not yet big enough. That is the simplest and heaviest fact in the room.
A challenger has two options. Chase Slack's feature list one item at a time, or change what the thing is measured by.
Changing the unit, not the feature list
The work messenger of the AI era is not a channel for passing messages. It is a workroom where the context of people and agents working together stays behind as structure.
Tools that chat with AI one to one already exist in abundance. That is not a messenger, that is a very good chatbot. The change comes from inviting an agent into the room where the people are and finishing the work together. And the unit of that room cannot be text that flows past.
So we rebuilt the unit of the conversation. A single mark inside a conversation becomes a topic, a decision, and a to-do with a deadline on it. The agent is not a guest you summon with a mention. It is a member that reads that structure and does the work.
Replacing a messenger is the hardest change an organisation can make. We know. But one thing is now clear. Slack holds a world-class model and still did not change what most of its users do. The model is not the shortfall. The shortfall is a container built ten years ago with AI strapped to the outside.
The other half of what we have been learning is in They bought Salesforce. Nobody uses it. and A written SOP does not run itself.
Written from a product strategy session and customer interviews on 1 October 2026.