Slack is Being Rebuilt as the Conversational Control Plane for Enterprise AI Agents

What happened

Salesforce unveiled more than 30 new capabilities for Slackbot on March 31 in San Francisco. Marc Benioff headlined the event himself — and the positioning was explicit. This is not Slack with AI features attached. This is Slack as the front door for enterprise AI: the single interface where employees invoke agents, route work across systems, supervise outcomes, and get answers without needing to know which backend actually did the work.

Three months after Slackbot became generally available in January, Salesforce says it's on track to be the fastest-adopted product in the company's 27-year history. According to VentureBeat, some employees at customer organizations report saving up to 90 minutes per day. Inside Salesforce itself, teams claim 20 hours per week in savings, translating to an estimated $6.4 million in productivity value (VentureBeat). Those are vendor-reported figures, but the scale signals something: adoption is real, not just trial.

Why this is a category signal, not just a product launch

The easy read on this story is product news. Slackbot got smarter. Salesforce added features. The AI feature war continues.

The more useful read is about architecture. When Salesforce frames Slack as "the new interface for the agentic enterprise" and "the operating system for work," they're staking a position in what may be the most consequential fight in enterprise software right now: who owns the layer where humans actually interact with AI agents.

For a while, that fight looked like it would go to whoever had the best model. Now it looks like it may go to whoever has the best interface — the place where employees send instructions, check agent work, and decide whether to approve or override.

That's a genuinely different bet. And Salesforce is making it clearly.

The breakdown

Reusable AI-skills: from chatbot to operating surface

The most structurally significant feature in the March 31 launch is reusable AI-skills. A user defines a repeatable task — "create a budget for an event" — and Slackbot handles it: pulling information from connected channels and apps, drafting the plan, scheduling a meeting with the relevant people. Define it once, use it anywhere.

This is the meaningful shift. A chatbot answers questions. A control plane executes repeatable workflows on command. Reusable AI-skills are the mechanism that crosses that line.

MCP support turns Slackbot into an orchestration router

Slackbot now functions as an MCP (Model Context Protocol) client. It can connect to Agentforce — Salesforce's agent development platform — or to any MCP-compatible enterprise tool, routing work accordingly. The user just asks Slack. Slack figures out which system handles it.

This matters beyond the Salesforce ecosystem. MCP support means Slackbot isn't limited to orchestrating Salesforce agents. It can route to agents built on other platforms entirely. That's a meaningful architectural claim: not just "Slack works with Agentforce," but "Slack works as a universal routing layer."

Desktop context, meeting capture, and ambient presence

The update also adds desktop context — Slackbot can now operate outside the Slack application itself, accessing what's on screen. Combine that with meeting intelligence across any video provider, plus persistent memory for user preferences and recurring workflows, and you're looking at something closer to an ambient operating layer than a messaging add-on.

This part gets less attention in the coverage than the MCP and skills announcements, but it's worth sitting with. Agents that see your screen, remember your preferences, capture your meetings, and access your files across apps are not just productivity tools. They're context machines. What they do with that context — and what knowledge they retrieve when they act on it — is what actually determines whether the results are trustworthy.

The usage numbers, with appropriate skepticism

According to CRN, citing Salesforce metrics, Slack has grown tenfold year-on-year in AI users. Salesforce also reports nearly 20 trillion tokens consumed and 2.4 billion agentic work units across Agentforce and Slack to date (CRN). Treat vendor numbers with the usual skepticism about methodology and definitions. But the order of magnitude — and the speed of adoption Salesforce describes — is consistent with what's happening in enterprise AI more broadly. Teams aren't piloting this cautiously. They're adopting it fast.

The interface war: what's actually at stake

Salesforce is not alone in this fight. Microsoft Copilot sits inside Teams, Office, and Windows. Google is running the same play through Workspace. App-native copilots built into Workday, ServiceNow, and Salesforce itself are all competing for the same surface area.

The structural question here is not which product wins. It's whether enterprises end up with one shared interface for supervising agent work, or dozens of fragmented ones — each agent accessed through a different app, with no unified view of what any of them are doing.

Fragmented interfaces mean fragmented oversight. Agents operating across ten different control surfaces, with no common point of human review, is a governance problem in the making.

Slack's bet is on consolidation: one place to invoke agents, one place to supervise them, one place to route work. Whether that bet holds against Microsoft's tighter OS-level integration and the pull of app-native copilots is genuinely unclear. What's clear is that the control plane concept — the idea of a single interface for human-agent collaboration — is a real architectural model, not just a tagline.

What the Slack overhaul doesn't fix

Here's the honest accounting.

A conversational control plane solves an interface problem. It makes agents easier to invoke, easier to route, easier for humans to supervise. That's real value, and the adoption numbers suggest enterprises can see it.

It does nothing for what happens when those agents retrieve knowledge to act on.

If an agent drafts a proposal using pricing documents from eight months ago, Slackbot won't know. If two policy documents in the same knowledge base contradict each other, the interface layer won't flag it. If institutional knowledge was never written down — or was written down and then updated without removing the old version — no amount of conversational polish will surface the gap.

We've written about why agent memory needs its own infrastructure layer, and about how AI readiness is really knowledge-base readiness. Those problems don't get easier as interfaces improve. They get more consequential. An agent that can take action through Slack can take the wrong action through Slack, at scale, because the knowledge it retrieved was stale, ungoverned, or simply wrong.

The agentic service management model that's emerging in enterprise AI assumes that agents retrieve accurate, governed knowledge. That assumption is doing a lot of work — and most enterprise knowledge bases are not built to support it.

This is where Mojar AI fits into this picture. Not as the interface layer — Slack can own that — but as the governed knowledge layer underneath. Source attribution on every retrieval, contradiction detection across documents, automatic remediation of outdated content, permission-aware access. The infrastructure that determines whether agent actions are trustworthy, not just fast and easy to trigger.

A conversational front end amplifies whatever knowledge sits beneath it. Better control planes are genuinely good news. They also make the knowledge governance problem harder to ignore.

What to watch

Whether Slack's consolidation bet holds against fragmented app-native copilots, and whether Microsoft's OS-level integration gives Copilot an insurmountable distribution advantage. Whether enterprises push back on ambient capture at scale — desktop context, continuous meeting transcription, preference memory — or accept it as the cost of the productivity gains. And whether the industry starts treating knowledge governance as infrastructure, or keeps treating it as a content problem that will somehow sort itself out.

That last one matters most. The interface war has clear frontrunners. The knowledge layer war has barely started.