AI agents in Slack and Microsoft Teams are now a normal part of work: both platforms ship their own assistants, host third-party agents and give developers a supported way to build their own. For questions answered by documents and messages people can already see, the built-in assistants are usually enough. Build your own internal agent when the answer lives in your systems, when the agent needs to take actions there, or when it has to follow rules only your company has.
This guide is for engineering and IT leaders making that call while Q1 plans are being set, and ends with a checklist.
What Slack and Microsoft Teams offer for agents today
Slack
- Data access for agents. On February 17, 2026, Slack announced the Slack MCP server and the Real-time Search API. The search API retrieves Slack data without storing it on external servers, and private channels and direct messages need user consent.
- Third-party agents. Slack says more than 50 partners, including Anthropic, Google, OpenAI and Perplexity, are building agents on these APIs.
- Admin view. In April, Slack described the Agent Browser and AgentExchange as generally available, alongside granular AI exclusion controls, real-time permission sync and full audit logging.
- Deploying from builders. In August, Slack launched Add to Slack, a one-click path from builders such as LangChain, n8n, Lovable and Vercel. Slack says deployed agents inherit the workspace's permissions and data boundaries.
- Pro-code. The agent quickstart uses the Slack CLI and Bolt for Python or JavaScript, with the Claude Agent SDK, OpenAI Agents SDK or Pydantic AI as the agent framework.
Microsoft Teams and Microsoft 365 Copilot
Microsoft splits agents into two kinds. Declarative agents add your instructions, knowledge and actions to Copilot's own orchestrator and models, inherit Microsoft 365 compliance and security standards, and are designed to be used by individuals. Custom engine agents bring their own orchestration and models, support group collaboration in channels and meetings, usually need hosting outside Microsoft 365, and leave compliance and security to you.
You can build custom engine agents with Copilot Studio, the Microsoft 365 Agents SDK, the Teams SDK or Microsoft Foundry.
Next question: What separates an AI agent from a chatbot?
When the built-in assistants are enough, and when to build your own
Start with what you already pay for. If the question is "summarize this thread" or "what did we decide about the vendor", the platform assistant already sees that content with the user's permissions.
Build your own internal agent when one of these is true:
- The data is yours, not the platform's. Order history, entitlements, runbooks the assistant cannot reach.
- The agent must act. Open a ticket, reset an account, approve a refund under a limit, kick off a deploy, update a CRM record.
- The rules are specific. Who can approve what, which customers are restricted, what wording legal requires.
- Several people work with it at once. An incident channel or a deal room where the agent answers for the group, not one person.
A useful test: if the agent's value disappears once you remove access to your internal systems, you are building, not buying. A help desk agent that only reads your IT policy can be declarative. One that checks a laptop's warranty status and books the replacement is a custom build. Our guide to AI agents for IT operations lists the first workflows most teams pick.
Next question: Who builds an internal agent that works with your own systems?
The architecture of an internal chat agent
Whichever platform you pick, an internal agent has five parts.
- The chat front door. A Slack app built on Bolt, or a Teams agent built on the Teams SDK, the Agents SDK or Copilot Studio. Keep this layer thin.
- Identity passed through per user. The agent should act as the person asking, not as an all-powerful service account. Check permissions in the downstream system on every call. If the user cannot see a record in the CRM, the agent cannot show it to them in a channel.
- Retrieval over your documents. Index the wiki, runbooks and policies the agent needs, keep access rules on each document, and return sources with every answer. Our RAG development work starts with this permission model, because a retrieval index that ignores permissions leaks data.
- Tool calls into your systems. Each action is a narrow, typed tool with clear inputs, such as create_ticket or lookup_order, not a generic "run any query". Write actions get a confirmation step in the chat and limits on amounts or scope.
- Logs and evaluation. Record the prompt, retrieved sources, tool calls and result for every turn, keyed to the user. You need this for audit, debugging and the measurements below.
The choice between platforms changes the front door more than the rest:
One detail matters for Teams group use. Microsoft documents that in Teams group chats and channels, Copilot Studio agents cannot use knowledge sources that require end-user authentication, such as SharePoint, and are supported for those sources only in one-to-one chats. If your agent needs permissioned documents in a shared channel, plan for a pro-code build.
Next question: How do you build retrieval that respects document permissions?
Adoption and channel etiquette
Most internal agents fail quietly: people try them twice, get a vague answer and go back to asking a colleague.
- Start in one channel with one job. An IT help channel, an on-call channel or a sales ops channel.
- Answer in threads. An agent that posts long replies into the main channel gets muted. Keep the first message short.
- Respond only when asked. Mention or direct message, not every message in the channel. Proactive posts are for alerts the team asked for.
- Show sources and say when it does not know. A cited answer or a clean handoff to a named human queue builds more trust than a confident guess.
- Give people a way to flag a bad answer. A reaction or a button that writes to your log, reviewed weekly by the team that owns the agent.
Plan for platform limits too. Microsoft notes that Teams applies rate limiting to agents, so keep replies concise and batch updates.
Next question: Which IT workflows should an agent take on first?
Security review for agents in chat
A chat agent sits where people paste everything, including customer details and the odd credential. Review it like any internal app with production access, plus these checks.
- Data flow. Map where messages go. Microsoft notes that when you add an agent to Teams, agent and user chat content is shared with Teams and might flow outside your compliance or regional boundaries. Know the same path for your model provider and logs.
- Scopes. Request only what the job needs, especially private channel and direct message access.
- Prompt injection. Anything the agent reads, from a pasted email to a document in the index, can carry instructions. Tools that write or send need confirmation and hard limits that the model cannot talk its way past.
- Secrets. Tokens for downstream systems live in a secrets manager, scoped per tool, never in prompts.
- Kill switch and ownership. One named owner, a way to disable the agent in minutes, and an admin path to remove it from channels.
Next question: Who can review an agent's OAuth scopes and access before launch?
Measuring whether the agent is used and useful
Installs are not adoption. Measure what changed in the work the agent was built for.
- Weekly active users in the target group, and how many come back the following week.
- Resolution rate: the share of conversations that end without a handoff to a human, checked against a sample of transcripts so "resolved" means correct.
- Time saved on the target task, compared with the month before launch.
- Answer quality: a fixed set of real questions with expected answers, run on every change to prompts, models or the index.
- Cost per resolved conversation, including model calls, hosting and the human review time.
Set the targets before launch and review them after 30 days. If the numbers do not move, change the job before you add features.
Next question: How do you test an agent before it reaches every employee?
Checklist before you ship an internal chat agent
- One named job, one pilot channel and one owner.
- Confirmed the built-in assistant cannot already do the job.
- Platform path chosen (low-code or pro-code) with the group-chat and authentication limits checked.
- Identity passed through per user, with permissions checked in each downstream system.
- Retrieval index keeps document permissions and returns sources.
- Tools are narrow, typed and limited; write actions ask for confirmation.
- Scopes minimized, data flow documented and approved by IT.
- Prompt injection tests run against documents and pasted content.
- Every turn logged with user, sources, tool calls and outcome.
- Evaluation set of real questions passing before each release.
- Thread-only replies, mention-only triggers and a clear handoff to humans.
- Kill switch tested; admin removal path confirmed.
- Success metrics and a 30-day review date agreed.
Next question: What do AI agent developers cost per month?
How Ryz can help
Ryz AI pod teams are dedicated senior engineers who build agents inside your cloud and repos, working on your team and reporting to your leads, on US business hours. Pods have shipped production agents such as an AI driver-support agent covering about 218,000 driver calls a year in three languages. If you are planning an internal agent for Slack or Teams in Q1, our AI agent development engineers can scope the identity, retrieval and tool layers with you, and most of our engineers ship code in their first week.
FAQ
Can one agent run in both Slack and Microsoft Teams?
Yes, if you keep the chat layer thin. Put retrieval, tools, permissions and logging in a shared backend, then build a Slack app and a Teams agent as two front doors. Microsoft's Agents SDK is designed for multi-channel agents, and Slack's Bolt frameworks handle the Slack side.
Should we use Copilot Studio or the Microsoft 365 Agents SDK?
Use Copilot Studio when the agent mostly answers from Microsoft 365 content and connectors and you want fast, managed deployment. Use the Agents SDK or Teams SDK when you need your own models or orchestration, complex actions in internal systems, or permissioned documents in shared channels.
Do we need a vector database to build an internal agent?
Only if the agent answers from a large body of documents. Many first agents mostly call tools, such as looking up a ticket or an order, and need little or no index.
How long does a first internal chat agent take?
A narrow pilot in one channel with a few tools is weeks of work for an experienced team, not months. The time goes into identity, permissions and evaluation, not the chat interface.

