AI Agent for Slack — the definitive guide.

An AI agent for Slack is a bot or app that operates inside a Slack workspace using natural-language commands (typically via @mention) and can perform actions beyond simple message replies. The strongest AI agents for Slack connect to external systems (CRMs, ad platforms, payment processors) and can both answer questions from live data AND take action on the user's behalf — with explicit human approval for anything that mutates state.

A Slack bot is any automation installed via the Slack App Directory — could be a webhook poster, a poll, or a reminder. A Slack AI assistant uses an LLM to answer questions in Slack but typically doesn't take action on external systems. An AI agent for Slack does both: it understands intent, retrieves data from connected accounts, proposes actions, and (with approval) executes them. Mavrick falls into the third category and goes further — we call it an AI coworker because of the collaborative approval pattern.

The fastest path: install via the Slack App Directory if the agent is listed there (one-click OAuth). For custom agents, you'd build a Slack app manifest, configure OAuth scopes (typically chat:write, channels:read, im:write, app_mentions:read at minimum), deploy a public webhook endpoint to receive events, and submit for App Directory approval. Mavrick installs in 60 seconds via the App Directory — full setup at /install.

Depends on what it's connected to. For marketing-focused agents like Mavrick: pull ROAS across Meta + Google + TikTok in one Slack message, pause underperforming campaigns (with approval), reconcile Meta-reported revenue against Stripe actuals, draft outreach emails, qualify inbound leads via voice (60-second response), update HubSpot records, schedule executive briefings. The capability ceiling is set by how many tools the agent connects to — Mavrick's managed connector layer covers 3,200+.

The honest answer: depends on the agent. Things to evaluate: (1) Does it have an approval gate on actions that mutate data? (2) Is the OAuth-token storage encrypted at rest? (3) Does the AI model see your tokens directly, or does the agent inject them at a tool-call boundary? (4) Is there a public Constitution / system prompt you can read? Mavrick's architecture: cleared-hot approval is contractual (Privacy Charter Rule 3), credentials never reach the model (Rule 4), and the system prompt is public at /system-prompt. See /trust for the full posture.

@-mention an AI agent in any Slack channel or DM, give it a natural-language command, and the agent responds in the same thread. For example: @Mavrick what's our Meta ROAS this week vs Stripe revenue? The agent reads the message, decides what data to fetch, calls the right APIs, composes a response, and posts back in thread. If the request requires a mutation (e.g., 'pause campaign X'), the agent first proposes the action and waits for your explicit approval before executing.

Cleared-hot is the explicit one-click approval an AI agent waits for before taking any action that mutates customer data — pausing a campaign, sending an email, updating a CRM record, transferring budget. It matters because AI agents that take action without approval will eventually take a wrong action, and recovering from that wrong action is expensive (financial loss, customer trust damage, audit-trail gaps). Mavrick's cleared-hot is architectural, not a setting: there's no way for an operator to disable the approval gate.

Yes — and we walk through what's involved in /blog/build-ai-agent-for-slack. Short version: it's a 6-12 month engineering project to build something equivalent to a managed AI agent (OAuth flows, tool gateway, approval architecture, audit logging, vendor integrations, monitoring, on-call rotation). For most marketing teams, installing a managed agent is the better trade. For platform teams with unique tools or strict compliance requirements, building is sometimes the right call.