On March 31, Salesforce announced an AI-heavy overhaul of Slack: 30 new features, agents in the sidebar, enterprise search, summaries, and a Slackbot that finally does more than remind you about a meeting. A few weeks later, on April 30, they announced they would crowdsource the next phase of the AI roadmap directly with customers. Two announcements, a month apart, both framed as Slack getting serious about AI.
I read the first one and counted to thirty. Then I read the second one and knew exactly how this story ends.
A few weeks ago I wrote about what I called the comfortable trap: established companies bolting AI onto a legacy product as a retention play, while the actual category-defining work goes unfunded. Slack is now the cleanest example of that trap I have seen this year, and not because Salesforce is badly run. It is doing the perfectly rational thing, and the rational thing is the trap.
Thirty Is the Tell
When a product ships one feature that changes how you work, that is a bet. Shipping thirty at once is usually a roadmap getting cleared so a slide can say “AI-first,” and you announce breadth when you have no depth to announce.
I have shipped features in batches like this. Early at Drift, when conversational AI was the whole pitch, we learned fast that the count of AI capabilities was inversely correlated with how much any single one mattered. Thirty things that each shave a few seconds off an existing workflow give you a longer release note, not a new way of working.
Every one of those thirty features also gets built on the architecture Slack already has. Channels, threads, the message-as-atomic-unit model, the workspace boundaries, the search index that was designed for keyword retrieval and is now being asked to feed a retrieval-augmented agent. You can put an agent in the sidebar, but it still reasons over a data model designed in 2013 for humans typing to humans. The AI inherits every constraint of the substrate underneath it. That is the cost of bolting on. Every bit of capability you add drags weight along with it, and that weight compounds.
Asking Customers Is How You Get Disrupted
The April announcement is the part that should worry anyone at Salesforce. Crowdsourcing the AI roadmap with existing Slack customers sounds like good product discipline. It is the exact mechanism Christensen documented in The Innovator’s Dilemma.
The disk drive makers asked their best customers what they wanted. The customers said: more capacity, more speed, better cost on the metrics we already care about. Nobody asked for the smaller, worse, cheaper drive that eventually ate the category, because no current customer experiences a category shift as a feature request. They experience it as their current tool, but a little better.
Slack’s existing customers will ask for better search, smarter summaries, fewer notifications, an agent that can file a ticket without leaving the channel. All reasonable. All sustaining improvements to the thing they already use. None of them is “replace the channel-and-thread model with something built for agents talking to agents and humans dropping in to supervise.” Your customers cannot ask for the thing that redefines the category, because if they could imagine it clearly enough to request it, it would already exist. A roadmap built from customer requests is, by construction, a roadmap of sustaining improvements. That is the definition of the comfortable trap with a survey attached.
What Net-New Would Actually Look Like
The interesting question is not “what AI features belong in Slack.” It is “what does team communication look like if you design it now, assuming agents are first-class participants from day one.”
That product probably does not have channels as its primary structure. Work, not conversation, is the unit. An agent picks up a task, does part of it, and the human sees the result and the reasoning, not a transcript of messages. Memory is durable and structured, not a search index bolted onto a chat log. Permissions are scoped to what an agent is allowed to do, not just which rooms a person can read. The interface is built around supervising autonomous work, with conversation as one mode among several rather than the whole frame.
I think about this constantly building at Vestmark. When we built Vestmark Pulse, the temptation was to add AI to the existing workflows that 400-plus employees and clients managing roughly two trillion in assets already knew. Some of that is right; you do not throw away workflows people depend on. But the parts that matter most were the ones we designed around what AI could now do, on a foundation that assumed it, rather than the ones we retrofitted onto what was already there. The retrofit ships faster and demos fine, but it is the net-new build that actually moves the numbers.
The Part Salesforce Can Still Do
Salesforce is not a small company with no options. It has the distribution, the enterprise trust, the data, and the capital to build the net-new version of team collaboration and put it in front of every customer it already owns. That is an advantage no startup has. It is also a depreciating one.
The mistake would be to treat the 30 features as the strategy. They are a retention play that keeps Slack from churning while the real work, hopefully, gets funded somewhere else. If that somewhere else does not exist, if the entire AI investment is thirty features and a customer survey, then a year from now someone will demo a collaboration product designed for agents from the ground up, and the response inside Salesforce will be the most dangerous sentence in business: we’ll get to that after we finish the current roadmap.
If you run a product like this, do one thing this quarter. Take the smartest people you have, give them a mandate to build the agent-native version of your product on a clean foundation, and explicitly forbid them from reading the customer feature-request backlog. Judge them on whether they built something that could replace you, not on whether they shipped what your current users asked for. The survey will tell you how to keep this year’s customers; it will not tell you anything about who eventually replaces you.