/
/
Slack AI Governance & Implementation
Expert Service
Shivanath Devinarayanan
By 2026, Slack's AI surface spans Workflow Builder AI steps, context-aware agents built on Agent Kit, admin-approved MCP integrations, and Agentforce agents in channels and DMs. Correct implementation means governing all four before they ship.
On this page
How does Slack's AI surface break down?
Workflow Builder's AI step
Slack's AI step, shipped in May 2026, drops natural-language-prompted summarizing, translating, and drafting into Slack's existing no-code workflows. It stays deterministic until the AI step fires, and stays grounded in the channels, canvases, lists, and files you point it at. Good for onboarding and simple approvals, it breaks down fast on anything that needs outside context.
Context-aware agents (Agent Kit)
Slack's own context-aware custom agents, built with Agent Kit (the slack create agent CLI, enhanced Bolt SDKs, and bring-your-own-LLM support), are goal-directed and tool-using, not single-step. They read across channels and take multi-step action, exactly why they need scoped permissions before they ship.
MCP server, MCP client, and the RTS API
Slack's hosted MCP server (mcp.slack.com) reached general availability on February 17, 2026, letting external AI clients like Claude and Perplexity search messages, files, and canvases, and take actions via admin-approved OAuth. Slack also became an MCP client, not just a server, so it can reach outside tools too. The Real-Time Search API enforces permissions live: revoke someone's access and it is reflected on the very next query.
Agentforce in Slack
Agentforce 2.0 puts employee-facing agents directly in channels and DMs via @mention, plus purpose-built apps for Sales, Service, IT Service, and Tableau Next. It runs on a separate Salesforce license with its own Agent Builder configuration, a native record-plus-conversation surface, not a bolt-on chatbot.
Choosing your Slack runtime
Slack gives you five ways to run logic: Workflow Builder, the Deno SDK, a self-hosted service you run yourself (Bolt or otherwise), a durable orchestrator outside Slack, or an agent working through MCP. A deployed Deno SDK function has a hard 60-second execution ceiling in production, 15 seconds locally, 10 seconds behind a button or modal, with no documented retry or rollback if it times out mid-run. Pick the runtime by what the outcome demands, not by which one your team already knows.
Slack as your system of context
Home · About · Services · Portfolio · Success Stories · Writing · News · Events · Contact · Media kit · Privacy Policy