Digital agencies run on Slack: one channel per client, plus internal channels for each account team. The slow part is rarely the reply itself. It is the lookup behind it: which Linear issues are still open, what the SOW in Notion actually covers, which deals are open in HubSpot. An AI agent in Slack can do that lookup where the question is asked, cite where the answer came from, and post a weekly roundup on a schedule. This guide covers three agency workflows that work today, how to set them up in Runbear, and the limits to know before you put an agent near client conversations.
What an AI agent does in an agency's Slack
A chatbot answers from what it was trained on. An agent in Slack can also call the tools your team already uses and read the documents you connect to it. In Runbear those are two separate pieces:
- Knowledge sources such as Notion, Confluence, Google Drive, SharePoint, selected Slack channels, websites, and uploaded files are synced and indexed ahead of time, so the agent searches them when a question comes in. They are not read live; sync is incremental. See Knowledge Base overview.
- Tools (MCP connectors) such as Linear, Jira, Asana, ClickUp, HubSpot, and Google Calendar are called in real time, for live lookups and actions. Many tools let you choose, per connection, between Per-User, where every teammate signs in with their own account, and Shared, where the agent uses one account for everyone. Some tools, such as HubSpot, always use one shared token. See Tools overview.
For an agency, the practical result is one agent that knows your internal playbooks and can check the live state of each client project, without anyone leaving the channel.
Workflow 1: answer "where are we?" in the client channel
The most common interruption at an agency is a status question. An account lead is asked for an ETA, opens three tools, and pieces together a reply. With an agent deployed to that channel, the account lead @mentions it and gets a draft built from the project tracker and the scope document, with the sources listed.
How to set it up:
- Create an agent and connect the tools it should read, for example your project tracker as a tool and your SOW and playbook folder as a knowledge source.
- In the agent's instructions, say who it is talking to and what it must not do, for example: answer status questions from Linear and the SOW only, always list the documents or projects you used as sources, never quote internal pricing, and say "I'll check with the team" when the data is missing.
- Open the agent's Chatbots tab and add a Slack connection for the client channel. The agent always replies to @mentions and direct messages; other triggers are switched on per connection.
Start with @mention only. On agents that run on the Claude Agent SDK runtime, the Slack triggers available are: always reply to @mentions and DMs, continue in threads after the first reply (AI-powered smart replies), reply without @mention, and trigger by keyword. Reply without @mention answers every new message in the channel, so in a client channel it is usually better left off. See Chatbot settings.
Workflow 2: internal answers from your own playbooks
New account managers ask the same questions in their first months: how the agency handles change requests, what the approval process is for ad spend, where the brand guidelines for a client live. Connect those documents as knowledge sources and deploy the same agent, or a second one, to an internal channel such as #ask-ops.
Because answers come from synced documents, keep the source of truth in one place and keep it current. If a policy changes in Notion, the agent picks it up on the next sync, not the moment the page is edited. If you also sync selected Slack channels as a knowledge source, past decisions discussed in those threads become searchable too.
Workflow 3: a weekly client digest on a schedule
An agent does not have to wait to be asked. On the agent's Event Triggers tab you can add a scheduled job (On Schedule or One-time) with its own prompt. With the job's Notification toggle on, the agent posts its result to a Slack or Microsoft Teams destination you pick each time the job runs. See Triggers.
A good first prompt for an agency: "Every Monday, for each client project in Linear, list what closed last week, what is still open, and anything unassigned. Then list open client deals and their stages from HubSpot." Post it to an internal channel first, and only move a version of it into client channels once the team trusts the output.
What to check before you put an agent in client channels
- Who can trigger it. By default, everyone in the workspace can use the agent in the channels it has access to. Runbear has a Restrict to specific users option on a Slack connection, but it is in limited availability and enabled per organization on request. On Claude Agent SDK agents, that list does not cover replies in threads the agent is already part of.
- Whose access it uses. A Shared tool connection uses one account for every request, so anyone who can talk to the agent can reach what that account can see. For client data, check per tool whether Per-User or Shared is available and right. HubSpot, for example, always uses one shared token.
- What it can see. Separate agents per client, or at least separate knowledge sources, can help keep one client's documents out of another client's channel.
- Verification. Ask for answers with sources and keep a human in the loop for anything sent to a client. A draft the account lead reviews is a safer default than an agent replying on its own.
- Security review. Runbear is SOC 2 Type II in production. Data is encrypted at rest and in transit. Check your client contracts for any rules on AI tools or data processing before connecting client systems.
What it costs
Runbear charges by usage, not by seat. One subscription covers every teammate who uses your agents in Slack or Microsoft Teams. Plans include a set number of builders, the members of your Runbear organization who create and manage agents in the dashboard: 5 on Team and 10 on Business. Usage is metered in credits, based on the tokens each model call uses and the model you choose for each agent. Team is $99 per month with 20,000 credits, and Business is $399 per month with 100,000 credits. See pricing.
Where to go next
For a broader look at how agencies use agents across client operations, see how agencies use AI agents to scale client operations. To compare options, read our list of the best AI agents for Slack. The agency-specific setup is on the Runbear for agencies page, and the general no-code AI agent for Slack page explains deployment in more detail.
FAQ
Can an AI agent answer client questions in a shared Slack channel?
It can reply in any Slack channel where you deploy it, and it always replies to @mentions. For client-facing channels, the safer setup is to have the agent draft answers that an account lead reviews, keep reply without @mention off, and limit which tools and documents that agent can reach.
Does the agent read Notion or Google Drive in real time?
No. Knowledge sources are synced and indexed, and sync is incremental, so edits appear after the next sync. Tools such as Linear or HubSpot are called in real time.
Can the agent post a weekly status update without being asked?
Yes. A scheduled job on the agent's Event Triggers tab runs on a repeating schedule or once, and with Notification on it posts the result to a Slack or Microsoft Teams destination you pick.
Is Runbear priced per seat?
No. Runbear charges by usage, not by seat. Team is $99 per month with 20,000 credits and 5 builders; Business is $399 per month with 100,000 credits and 10 builders.
