Table of Contents
I increasingly hear the same frustration from customers: they’ve rolled out AI and aren’t seeing the return. The data says they’re not alone. Gartner forecasts about $2.7 trillion in worldwide AI spending this year, up nearly 50% from last year. At the same time, 56% of CEOs say that AI has delivered neither higher revenue nor lower costs over the past twelve months.
McKinsey’s survey shows the split between people and companies. Eight in ten people say that AI has made them more productive, but only 37% of companies say that it has added to their profit.
$2.7 Trillion
Global AI spend in 2026
That gap usually comes down to where the AI works. Most companies have bought into an AI product that delivers personal productivity, but the work those people do is actually shared across people, teams, and, increasingly, agents. The result is a context gap in both directions: the AI doesn't have access to what most of the team knows and the team can't see how their peers are working with the AI to get things done
Most of us use AI as a private assistant. We prompt an agent in a private space, get an answer and then pass along the output without ever sharing the reasoning. The team sees a small slice of the work and none of the insight we gained getting there.
Part 1
Why private AI looks familiar
A decade ago, most work was trapped in inboxes. A decision got made in an email chain that only three people could see, and everyone else found out later, or never. And that entire inbox was deleted when the person left the company.
Slack spent over a decade breaking down those knowledge silos and bringing context into the open, for the whole team to leverage. Today, the most innovative and agile companies across industries run on Slack because of its open channel structure: the decision, the debate, and the reasoning sit in the channel, so a new hire can read the history and get up to speed without having to find the one person on the team with tribal knowledge.
When standalone AI tools arrived, they resurfaced the legacy behaviors we left behind with the inbox: private chats, where information, learning, and context remain trapped and fragmented.
Part 2
The bookends of the modern AI stack
Companies have invested heavily in the stack that AI runs on: models, databases, retrieval pipelines, guardrails, and applications. I think of Slack as the bookends of that stack that most companies are missing: a foundation of context at the bottom, and a collaborative interaction layer at the top.
At the bottom is the foundation of context, your company’s alpha: the living record of how your company works, including what got decided last Tuesday and why, who owns what, and what’s blocked. As our CPO Jaime DeLanghe puts it, the work is the conversation. Anthropic and Jaime talked about this as turning conversation into knowledge. Many companies try to build this by hand from wikis, docs, and meeting notes, which takes sustained effort and goes stale quickly. While in Slack, the context builds itself from the conversations your teams are already having, so nobody has to maintain it. Importantly, an agent working in Slack sees what the people in that channel are allowed to see, under the permissions and admin controls the company already has in place.
Agents reach that context through the Slack MCP server and the Real-Time Search API. MCP, the Model Context Protocol, is a standard way for agents to discover information and take actions, and the Real-Time Search API lets an agent retrieve only what it needs, when it needs it, without a copy of your data stored elsewhere. Both work within the permissions your company has set.
At the top of the stack is the interaction layer, the multiplayer runtime where agents get deployed and discovered, and where teams use AI together in open channels. People use AI that’s intuitive and meets them where they work, and they learn from each other with every interaction, building AI fluency across the company. A channel is the one place where context, permissions, and the people who need the answer already exist together, which an inbox, a chat window, or a standalone agent doesn't give you. A channel can represent a team, a project, a customer shared across companies through Slack Connect, or a code channel where coding agents build alongside the whole team. Whatever the shape of the work, an agent joins the channel with the context already there, builds on what the team has decided, and acts like a good teammate, so the whole team can see what it does, build on it, and learn from it.
These two layers feed each other. Context makes the runtime smarter, and the exhaust of collaborating with agents in the open builds more context. Together, they are what we mean by multiplayer AI: people and agents working in the same channels, with shared context.
8 in 10
people say AI boosts their productivity
Part 3
How compounding context builds company intelligence
One person using AI well gets faster. In recent Salesforce research, 66% of the executives surveyed said that two colleagues solve the same problem with AI without realizing it, at least once a week. One IT director described engineers prompting AI to fix the same error because their separate chat histories kept the first solution out of reach. Bring that work into a channel and the team can react to it and sharpen it, so what one person figures out turns into an insight the whole team can use. This is where a flywheel begins to spin.
A company is a network of those teams, and that’s where the effect gets larger. What support works out on Tuesday is exactly what the product team or an account executive needs on Thursday, and when it lives in a shared channel, an agent can find it for them, within the permissions the company already has. The exhaust of that work, the decisions, the reasoning, and the corrections, becomes context that every other team and agent can draw on, which is how the whole company gets smarter and not only the team in the room.
Shopify CEO Tobi Lütke has described what's at stake when AI work stays private: “The risk is that AI does the work and we never learn from it.” When people work with their agents in public, he says, the opposite happens, and the best prompt patterns and knowledge spread.
Over time that flywheel turns company knowledge into something like a brain that improves itself. Richer context makes AI answers sharper and more specific to how the company operates, better answers get used more, and other teams see it working and do the same. That loop is hard to copy, because a competitor can license the same models but not your company's accumulated judgment.
Part 4
This is already happening in Slack
Our customers and partners are leading the way. They brought agents into their channels and put them to work next to their teams, where everyone could see the work.
At Dreamforce 2026, Boris Cherny, Creator & Head of Claude Code at Anthropic, shared that 90% of the code he writes happens in Slack. When Anthropic built Claude Tag, it brought Claude straight into Slack channels. You tag it like a teammate, and it builds context from the channel and carries the work forward where everyone can see it. Anthropic says 65% of its product team’s code is now created by its internal version of Claude Tag.
“90% of the code I write happens in Slack.”

Vercel's CTO, Malte Ubl, describes it this way: “Slack is the absolute lifeblood of Vercel. It's where all work happens.” For his team, that's literal. Much of their production software now gets written in Slack.
“Slack is the absolute lifeblood of Vercel. It’s where all work happens.”

Slack Code is the clearest example of the runtime. Coding agents like ChatGPT, Claude Code, Devin, GitHub Copilot and Vercel's agents work in a dedicated code channel next to the team. They post their plan, the exact code changes, and a live preview in the thread, and anyone can pause or redirect them. A human signs off before anything reaches production, so a product manager or designer can trigger an agent and review the output without waiting on engineering, and the context stays in one place afterward. When we introduced it, GitHub CPO Mario Rodriguez said bringing Copilot into that conversation is “nirvana,” and he sums up the model this way: “humans set direction, agents close the loop.” It’s available today on all Slack plans.
“Slack is a strategic part of a broader GitHub promise. We're bringing Copilot into the flow of conversation in code channels to help teams turn shared intent into shippable software.”

Part 5
Multiplayer AI is the new modality
Most modern platform shifts begin with individuals and mature when companies build around them. Conversational AI assistants laid the foundation for AI by getting individuals comfortable working with it, and the next modality is multiplayer, where a whole company can build on that foundation.
The companies above already work this way: their AI runs in the open, alongside the team, with shared context.
Capable models are widely available and keep improving, so the real opportunity shifts to what the model knows about your company. In Slack, that shared context is your advantage: the accumulated record of how your people work together. That includes the conversations, the decisions, every file and application, and your business data, all searchable by the people and agents who need it. It belongs to you, and it grows more valuable with use.
So stop asking which model to use. On Monday, ask three questions:
- Where does our AI actually work?
- What can it see?
- When one person gains an insight or gets the best result, who else benefits?
If the answers are “in private tools,” “a few files,” and “nobody,” you've made your people faster. You haven't yet made your company smarter.
AI works better when it works with the whole team, and that's already happening in Slack.




