I have spent the better part of a year wiring agents into every surface I can get my hands on. .NET 10 services calling Claude through Microsoft Agent Framework. Ollama running locally and pretending it’s a senior engineer. Custom MCP servers exposed over Streamable HTTP. Semantic Kernel orchestrators that talk to SQL Server 2025 vector indexes and pull back grounded answers. I have built agents that write fiction, agents that triage tickets, and agents that channel twenty different leadership voices on demand.
And after all of that, here is the unglamorous truth I keep arriving at: the hardest problem in agentic AI is not the model. It is not the framework. It is not even the data. It is the surface. It is the place where a human and an agent actually meet, exchange context, and get something done together without one of them feeling like they wandered into the wrong building.
That surface, for better or worse, has a winner. And it is not Microsoft Teams. It is not Zoom. It is not Telegram. It is not Discord. It is Slack.
I want to be careful here, because I have spent thirty years on the Microsoft stack and I am not in the business of doing Salesforce free PR. But credit where credit is due: Slack is currently doing a better job than anyone else of making humans and agents work together inside the same conversation. And the lead is not small.
The Substrate Argument
Start with a basic architectural question: where does collaborative work actually live?
If you are honest about it, work lives in conversation. Meetings produce conversation. Documents are downstream artifacts of conversation. Tickets are conversation that got formalized. Email is conversation that lost its rhythm. The collaborative substrate of a modern company is a stream of messages, and everything else is metadata.
Once you accept that, the design question for agentic AI becomes obvious. Your agent needs to live where the conversation already happens. Not in a sidebar. Not in a separate app you Alt-Tab to. Not as a chatbot widget pinned to the bottom right of a website. In the conversation. As a participant. With the same affordances a human teammate has — @mention, thread, DM, channel context, the ability to react, follow, archive, and search.
Slack is the platform that took this seriously. Everyone else is still treating agents as features grafted onto a product that was built for something else.
What Slack Actually Built
Let me give the receipts, because hand-waving about “platform thinking” only gets you so far.
Salesforce reports that Slack now hosts one million businesses and processes one billion messages per day. AI-enabled apps built for Slack grew 690% year over year. Those are not vanity metrics. They are the training ground and the distribution channel for everything that comes next.
Agentforce, formerly Einstein Copilot, was rolled out inside Slack starting in early 2025 and went broad in 2026. The mechanic is exactly what you would design if you were starting from a clean sheet: agents show up as members. You @mention them. They respond in thread. They have profiles, skills, and knowledge. Admins manage them through an Agentforce Hub. Three editions ship out of the box — Sales, IT Service, and HR Service — and you can build your own in Agent Builder.
Then, at Dreamforce in October 2025 and again at TrailblazerDX in early 2026, Slack opened the platform further. The new generation of Slackbot is an MCP client. It can route work or prompt questions to Agentforce, to any agent in your org, or to any third-party app that speaks MCP. Slack also stood up an MCP server in February 2026, which means external AI clients like Claude and Perplexity can now read Slack data and act inside Slack without anyone writing custom glue code.
Read the partner list and tell me this is not the open ecosystem we have been demanding for a decade: OpenAI, Anthropic, Google, Perplexity, Writer, Dropbox, Notion, Cognition, Vercel, Cursor. Adobe and Cohere too. That is not a walled garden. That is the agentic equivalent of the early App Store, and Salesforce is the one running the front door.
Slack did not win because its AI is better. It won because its product shape was already correct.
The architectural elegance is that none of this required Slack to become an AI company. Slack was already the right shape for an agent runtime. Channels are bounded context windows. Threads are conversation memory. @mentions are tool calls. Reactions are lightweight feedback. The DM is a private session. Enterprise Grid is a multi-tenant boundary. Every primitive an agent needs was already in the product before LLMs existed. Slack did not pivot. It just plugged in.
Microsoft Teams: The Strongest Competitor, But Carrying Too Much
This is where I have to be careful, because Microsoft is the platform I have built on for thirty years, and Microsoft is also the company most likely to catch up. Adoption numbers favor them: Futurum Group’s CIO Insights Survey, Q4 2025 wave, put Microsoft Copilot at 40.2% of CIO AI platform adoption, ahead of Gemini at 26.2% and ChatGPT/Azure OpenAI at 24.2%. That is a lead earned by Office 365 sales motion, not by product superiority.
Microsoft 365 E7, the Frontier Suite, bundles M365 E5, the Entra Suite, Microsoft 365 Copilot, and Agent 365 as the governance control plane. General availability lands May 1, 2026, which is two weeks from now as I write this. Agent 365 is a serious offering — it is a unified observability and policy plane for agents across the org, whether you built them on Microsoft, brought them in from a partner, or grafted them in from somewhere else. Copilot Studio added agent-to-agent communication, Work IQ API for shared context, and MCP support in April 2026. Copilot Cowork, which runs on Anthropic’s agentic models inside Microsoft 365, hit the Frontier program on March 30, 2026.
So why am I not declaring Microsoft the winner?
Because Teams is not the substrate. Teams is the meeting layer. The Microsoft agent story is fragmented across Outlook, Word, Excel, PowerPoint, Teams chat, Teams meetings, Copilot Chat on the web, Copilot Studio, Azure AI Foundry, and now the Microsoft 365 app. Microsoft has the most agents, the most governance, and the most distribution. What it does not have is one place where humans and agents talk to each other as peers in an ongoing conversation.
Watch what people actually do in Teams. They join a meeting. The Copilot summarizes it. They get a transcript. They get action items. They get a recap. All of that is post-hoc on a discrete event. The chat function inside Teams exists but is treated as a courtesy — most Teams chat threads die after a meeting ends. Compare that to Slack, where the channel keeps living, the agent stays at the table, and the conversation is the artifact. That is the structural difference. Microsoft is shipping incredible AI plumbing, but the conversational locus of work is still missing.
Keith Kirkpatrick at Futurum flagged the harder version of this problem in his April analysis of Copilot Cowork: when a shared agent contributes to a group decision in real time, who owns the output? That is not a philosophical question. It is a compliance and audit question, and Microsoft is going to have to answer it before regulated industries — financial services, healthcare, the buyers that pay enterprise prices — will move past pilots.
Zoom: A Meeting App Pretending to Be a Work OS
Zoom has the most aggressive AI rollout of any vendor on a per-employee basis. AI Companion is free with paid Zoom seats, AI Companion 3.0 went GA in December 2025, and at Enterprise Connect 2026 Eric Yuan and Velchamy Sankarlingam announced workflow orchestration across Workplace, Phone, and CX. They have a federated model approach that mixes Zoom’s own LLMs with OpenAI, Anthropic, and Google Gemini. They acquired Bonsai. They acquired BrightHire. They are not asleep.
Here is the problem. Zoom is fundamentally a meeting platform that has spent the last three years pretending to be a collaboration platform. Zoom Team Chat exists. Almost nobody uses it as their primary messaging tool. Zoom Docs exists. Almost nobody uses it as their primary document tool. The center of gravity in Zoom is the meeting, and meetings are bounded events, not continuous workspaces. An agent that lives in a Zoom meeting is an agent that wakes up for forty minutes and goes back to sleep.
I have also written before — and I will not relitigate the whole argument here — about the privacy posture Zoom took with AI Companion in earlier iterations. They have improved their language since then, and credit to them for the federated approach and the explicit “no training on customer content” stance. But the fundamental product shape has not changed. Zoom is a brilliant meeting app. It is not where work happens between meetings. And the gap between conversation and action — the exact gap Sankarlingam said the platform was designed to close — is structurally hard to close when your product was designed around the conversation ending.
Telegram: The Most Innovative Bot Platform Nobody at the Office Will Use
I have to give Telegram its due. The Bot API 9.5 update on March 31, 2026 — bots that can create and manage other bots, programmatically, without manual BotFather steps — is genuinely the most interesting agentic primitive any messaging platform has shipped this year. Combined with bot-to-bot communication, streamed responses, and the May 2026 update that lets any user tag an AI bot by @username in any chat, Telegram is doing things that look ten years ahead of the WhatsApp Business API.
Then there is TON’s Agentic Wallets, launched in May 2026, which let agents on Telegram hold funds and execute on-chain transactions within hard spending limits. Telegram bots are not just talking now. They are spending. That is a frontier capability and it is shipping today.
So why isn’t this the answer?
Because Telegram is consumer-first and crypto-adjacent, and no enterprise compliance officer is going to sign off on it as the system of record for company conversation. There is no SSO story. No data residency. No SOC 2. No audit trail enterprises will accept. The platform has 900 million monthly active users, but those users are individuals, traders, and communities. Telegram is the most interesting playground for agent builders and the least likely place your CFO will let agents touch a financial workflow. That is not a knock on Pavel Durov’s product. It is just a different game.
Discord: Wrong Audience, Wrong Architecture
I will keep this short because the comparison is uneven. Discord is excellent at what it does — community engagement, voice-first collaboration, gaming and developer communities. But Discord caps integrations at 50 per server, has no enterprise pricing tier, no comprehensive compliance story, and a thread model that auto-archives. Slack offers 2,600+ native integrations and indefinite searchable history on paid tiers. The Clyde chatbot was discontinued in 2023. AutoMod AI and conversation summaries remain in limited testing. Discord is not trying to be the agent runtime for enterprise work, and that is fine. It is a category error to put it in the comparison except as a control variable.
The Receipts — And the Caveats
I am not in the habit of repeating vendor marketing, so let me give you the numbers I have actually been able to verify, with the caveats they deserve.
Production figures from live Agentforce deployments inside Slack: one enterprise pipeline-management agent saved 125,000 hours in two months. Agentforce-powered sales agents reported a 60% improvement in deal cycle speed. Those are real numbers from real customers, surfaced in Salesforce’s own communications and corroborated by third-party coverage.
Now the caveat. Futurum Research has also reported that 87% of Agentforce deployments stall before generating measurable ROI. That number deserves to live next to the others. The agent platform exists. The wins exist. The mass deployment failure rate also exists. Anyone selling you frictionless agentic transformation is selling you a slide deck. What Slack has actually built is the runtime where success becomes possible, not guaranteed.
This pattern, by the way, is the same one we saw with cloud, with mobile, and with every prior platform shift. The early winners do not win by being correct on every deployment. They win by being the place where the next generation of work gets attempted.
Why This Matters for Architects
If you are designing an agentic strategy for your company in 2026 — and if you are not, you are already late — here is the architectural lesson Slack has handed everyone for free:
The collaboration platform is the agent runtime. Pick one, commit to it, and stop pretending your agents will live everywhere equally well.
Conversational data is the most valuable training and grounding signal you have. Whichever platform owns the conversation owns the agents.
@mention as a tool-call primitive is underrated. Humans already know how to use it. Agents inherit a UX humans understand without training.
MCP is the right abstraction layer. Whichever platform exposes MCP-server and MCP-client capability natively will eat the integration market.
Governance is table stakes. Adoption fails without it. Agent 365 and Slack admin controls are not differentiators, they are minimum viable product.
Slack got the substrate right ten years before it mattered. Microsoft will close the gap with sheer distribution and the Agent 365 governance story. Zoom will keep optimizing the meeting. Telegram will keep building wild agentic primitives for a consumer audience. Discord will keep being Discord. But for now, in May 2026, if you ask me where humans and agents are actually collaborating well together inside an enterprise, the answer is one product, and it is owned by a company that paid $27.7 billion for it five years ago and is finally getting paid back.
Marc Benioff bet the farm on a thesis that channels would become the operating system of work. The receipt arrived in the form of agents that actually fit through the door.
Credit where it is due.
— Gal Ratner
Solution Architect, Prana Entertainment & Inverted Software