The war on direct messages is a request to convert private attention into a machine-readable asset. The meter that charges for reading it does not move either way.
Executives are pushing workers to stop sending direct messages and post in group channels instead, because AI agents cannot see DMs. The stated goal is a complete corpus. The unstated cost is that somebody has to produce it, and the somebody is an employee whose time never shows up as a line item in the AI budget.
Slack is a backronym. It stands for Searchable Log of All Conversation and Knowledge, a name Stewart Butterfield attached to the product before anyone knew whether the log would matter. Thirteen years later, the log matters enormously, because a class of software has arrived that can only act on what it can read. And the discovery of 2026 is that the log was never complete. Most of what a company knows was typed into a one-to-one window, or said out loud in a meeting that nobody transcribed, or held in the head of a person who answered a question and moved on. The searchable log turns out to be the minority of the conversation.
What Happened
The Wall Street Journal’s CEO Brief on August 25, 2026 ran an interview by editor Lila MacLellan with reporter Lindsay Ellis, unpacking a story Ellis wrote with Allison Pohle about executives turning their frustration toward employee direct messages. One CEO called it a war on DMs.
The mechanics described are straightforward. Leaders embedding AI into daily work have found that agents only know what they can reach. Private meetings and one-to-one chats are the two places where information goes and does not come back out. So executives are asking staff to redirect those conversations into group threads an agent can parse, and to drop transcripts of virtual meetings into public channels.
Methods vary. Zapier chief executive Wade Foster ran a friendly competition among his executives to do it. Joe Inzerillo, president of enterprise and AI technology at Salesforce, which owns Slack, said his own team moved that way and expects adoption to spread once people feel the payoff. His example of the payoff is recall: after back-to-back meetings, he can point an agent at transcripts sitting in Slack and get an answer about who said what, instead of hunting for it himself.
Box chief executive Aaron Levie framed the underlying pressure at the WSJ Leadership Institute, arguing that any company wanting to be AI-first will be forced to start writing down decisions, best practices and processes, because agents cannot automate what they do not know. The Journal’s own reporting carried the obvious caution alongside it: several people raised the value of spitballing, and questioned whether moving that kind of talk into a permanent, public record would make people less willing to float half-formed ideas.
The Backstory
None of this is a new corporate ambition. The attempt to convert what people know into something the organization owns has a long institutional history, running through 1990s knowledge management, the intranet era, and a generation of document repositories that filled up and were never read. Each wave arrived with the same premise: the knowledge exists, it just needs capturing. Each one ran into the same wall, which is that capture is work, and the person doing the work is rarely the person who benefits.
What makes 2026 different is that the consumer of the corpus finally has an appetite worth the effort. A wiki page nobody read was a sunk cost. A wiki page an agent reads before drafting your quarterly plan has a return. That is a real change and it deserves to be taken seriously.
The uncomfortable part is where the numbers actually sit. In October 2016, Slack published a culture guide describing how its own team worked: roughly 70% of messages in public channels, 28% in private channels, and 2% in DMs. That page is still live on slack.com today. Outside the vendor, the ratios invert. Buffer published an audit of its own workspace showing under 40% of messages sent transparently, with around 65% living in private channels and DMs. The collaboration analytics firm Flowtrace has described public-channel share sitting between 10% and 30% at the companies it walks into. Worklytics has written up a client that found 75% of its messages in DMs or private channels before deciding to go public by default.
Read those together and the shape of the problem is clear. At a typical company, an agent with company-wide reach can search something in the neighborhood of a quarter of the conversation.
The same Slack culture guide contains a detail worth sitting with. Slack’s internal etiquette includes a raccoon emoji, used to signal that a discussion belongs somewhere other than the busy channel it landed in. The guide’s own example of using it is telling someone to move their questions into a DM. The vendor whose customers are now being told to abolish DMs still publishes a manual instructing people to route into them, for the entirely sensible reason that a crowded channel is a shared cost.
There is also an unresolved contradiction at the top of the company that owns Slack. In May 2026, on the All-In podcast, Salesforce chief executive Marc Benioff described using AI to see what employees were complaining about, and his phrasing swept in DMs alongside channels. Salesforce subsequently clarified that he was referring to public company-wide channels, and Slack’s stated position is that its AI features only draw on data a member can already access, excluding private channels and DMs they are not part of. Both statements cannot describe the same system. If the clarification is right, the war on DMs is necessary. If Benioff was describing what he actually does, it is theater.
The Plan
The ask being made of employees has three parts: move message volume out of DMs and into channels, push meeting transcripts into those channels, and do it because a leaderboard or a well-liked executive said so rather than because a policy requires it.
Slack’s product roadmap is moving in the same direction with more force than the culture campaigns. In August 2026 the company shipped Slack Code, which pulls agentic coding out of the terminal and into channels, along with an Agents tab, an install flow for third-party agents from platforms including Lovable, n8n, OpenAI, LangChain and Airtable, and Anthropic’s Claude Tag. Slack executives point to Shopify, where Tobi Lutke has restricted agentic coding to public channels on the argument that it disseminates what is happening across the company. Enterprise Search is sold as the layer that reads across the whole stack.
One item in that release list is worth flagging. Alongside everything else, Slack shipped agent DMs: private one-to-one conversations, with a machine. The direct message is not being abolished. It is being reassigned.
The pricing sits underneath all of it, and it has not really changed shape since the freemium days described in the Slack business model. Slack’s free tier hides messages older than 90 days, which means the corpus is literally the paywall. Pro runs $7.25 per user per month on annual billing. In 2025 Slack retired the standalone $10 per-seat AI add-on and folded the advanced AI features into Business+, raising that tier from roughly $12.50 to $15. Enterprise+ is unpublished, and procurement platform Vendr’s data from 535 verified purchases puts real deals at a median around $26.18 per user per month.
The Business Model Angle
A DM is an attention-rationing device, not a data leak. This is the part the coverage keeps missing. When you post to a channel with 200 members, you impose a read-or-skim cost on 200 people. When you send a DM, you impose it on one. The DM exists because attention is the scarcest input in a knowledge business and messaging tools do not price it. Every message relocated from a private window to a public channel moves a cost from the sender to everyone else in the room. The company books the benefit, which is a corpus an agent can query, and pays for it in a currency that never appears on the income statement. If the same work were bought honestly it would be documentation headcount, and it would be visible, and someone would have to defend the line.
The vendor gets paid the same whether the corpus gets written or not. Slack, like nearly every collaboration platform, sells seats, and seat count is the metric the whole Salesforce business model is built to expand. The value of the AI features depends entirely on the quality and completeness of the log, which is produced by the customer’s employees at the customer’s expense. The vendor prices the access. That is a rare arrangement: the input that determines whether the product works is supplied for free by the buyer, and the seller’s revenue is invariant to whether the buyer supplies it. It also runs the moat in the right direction. The richer the log, the higher the switching cost, and the log is precisely what the free tier restricts at 90 days. Every hour an employee spends moving what they know into the workspace is unpaid capital expenditure on an asset sitting on somebody else’s platform.
The observable and the objective are not the same thing. What leadership can measure is where messages are posted. What leadership wants is captured reasoning: why the decision went that way, what the tradeoff was, who owns the follow-up. Mandate the first and you will get it. Channel volume will rise, dashboards will show the shift, executives will report success. What arrives in the channel will be thinner than what used to arrive in the DM, because the parts people were willing to say when the audience was one person will now be said in a huddle instead. Then the agent trains on a padded record, which makes retrieval worse rather than better, and nobody has a metric that would catch it.
Which brings up the selection problem in the Zapier example. Foster’s competition among executives is being cited as evidence that a nudge works. Zapier has been fully distributed since 2011, has never had an office, published a 200-page guide to remote work, and runs past 800 people across more than 40 countries. Companies like that pay the documentation tax daily and have for fifteen years, because without it they do not function. The competition did not create that culture. It decorated one that already existed. A company that has never written anything down cannot buy the same result with a leaderboard, and the executives quoting Zapier are mostly running companies in the second category.
The Risk
The strongest case for the other side is that the benefit is concrete and the cost is diffuse. Inzerillo’s example is a genuine time save. Knowledge trapped in DMs produces repeated questions, slow onboarding, and real key-person risk when someone leaves. Worklytics reported that its client’s shift to public-by-default lifted engineering channel traffic 40%. Nobody who has watched a project stall because the decision lived in a departed employee’s DM history should dismiss this.
But four risks are being priced at zero.
The first is the one the Journal’s own interview raised and then left alone: early ideas need a cheap room. A half-formed suggestion in a DM costs nothing to withdraw. The same suggestion in a permanently retained, agent-queryable channel reads later like a position. You do not lose the conversation, you lose the draft, and drafts are where the useful thinking happens.
The second is legal. Everything that moves into a public channel is more retained and more discoverable. The compliance export features on the upper tiers exist because that record gets subpoenaed. Migrating informal, unguarded conversation into the formal record raises litigation exposure in a way that no one in the coverage has attempted to quantify.
The third is the Benioff contradiction, which cuts both ways. If admin-level tooling can already reach DMs, then asking employees to change behavior is a performance and the honest conversation about monitoring is being avoided. If it genuinely cannot, then a product limitation is being solved with unpaid labor. Either way the customer covers it.
The fourth is that the corpus may not be the binding constraint at all. That is the load-bearing assumption in the whole exercise: agents underperform because they lack data. If it were true, the organizations with the best documentation habits should already be posting the largest measurable AI gains, and the evidence on firm-level productivity payoffs is still thin across the board. It is at least possible that the missing input is not the corpus but the evaluation, and that the DM war is an expensive answer to the wrong question.
There is also a ceiling nobody has named. Compensation discussions, performance conversations, HR complaints and legal matters cannot go in a public channel and should not. So the target is not 100%. It is some undefined middle, which is exactly the kind of goal that gets gamed rather than met.
Quick Questions
Why can’t agents read DMs if the company owns the data? Ownership and access are different things. Slack’s stated design is that its AI features only surface data the requesting member can already see, which excludes DMs and private channels they are not in. Admin tooling and compliance exports operate under separate rules. The behavioral campaign exists precisely because the ordinary agent path stops at the channel boundary.
Is this really about AI, or about monitoring? Both, and the framing matters commercially. Sold as AI enablement, it is a productivity investment. Described as visibility into what workers say to each other, it is surveillance and it draws a different reaction from staff and from works councils in Europe. The same behavior change delivers both outcomes.
Does Slack benefit from this shift? Directly. A fuller log makes the AI tiers more valuable and makes leaving more expensive, and it costs the vendor nothing to produce. Seat pricing means Slack collects the same revenue whether or not your employees do the work that makes the product good, which is one of the hidden metrics problems in SaaS business models: net retention looks healthy while the thing driving the product’s value is being expensed to the customer off-book. It is a pattern worth watching across B2B software generally, not just chat tools.
What would actually work instead? Pay for the capture rather than exhorting it. That means resourcing the write-up as work, targeting decisions and processes rather than raw message volume, and measuring whether the record answers questions rather than whether the message count moved.
Is Microsoft doing the same thing? The dynamic is identical anywhere an assistant reads the workspace. Teams sits inside a bundle with an even larger installed base, which means the same unpriced documentation labor is being requested at a wider scale and with less discussion.
The Business Model Analyst Take
The war on direct messages is the moment enterprise AI stopped being a software purchase and started being an operating-model change with an invoice attached. The invoice is just not itemized.
Levie is right that companies will have to write more down. What follows from that is not a culture campaign. It is a cost. Somebody has to sit down, think about what actually happened, and put it somewhere useful, and that is a job, not a vibe. Companies that treat it as a job, staff it, and measure whether the resulting record answers real questions will end up with an asset. Companies that run a leaderboard will end up with a higher message count in public channels, thinner content inside those messages, and the important conversations relocated to a huddle, a phone call, or a text thread that no agent will ever index.
The tell is already in the product. In the same month executives declared war on human DMs, Slack shipped DMs for agents. The private channel is not the enemy of the machine. The unindexed one is. Whoever ends up owning the index will be selling it back to you at a seat price, and your employees will have written it.
