How to Set Up an AI Agent for Your Office: A Complete Workplace Guide

Every office runs on the same tedious patterns. Someone asks "where's the latest version of that document?" in Slack. Someone else spends 30 minutes writing the weekly status report. A third person manually checks three dashboards every morning and pastes numbers into a spreadsheet.

An AI agent eliminates these patterns. Not by replacing anyone, but by handling the repetitive coordination work that eats hours from every team member's week.

This guide covers setting up an AI agent specifically for office use: connecting it to Slack or Google Chat, giving it access to the right tools, locking it down securely for a multi-person workplace, and automating the tasks that nobody wants to do manually.

What an Office AI Agent Actually Does

Before diving into setup, understand what a workplace agent handles well and what it doesn't.

Tasks it handles well:

• Generating weekly/monthly status reports from data sources
• Answering "how do we do X?" questions about company processes
• Summarizing long documents, meeting notes, or email threads
• Drafting communications (emails, announcements, responses)
• Setting reminders and scheduling follow-ups
• Looking up information across multiple tools
• Creating first drafts of standard documents (SOPs, templates, briefs)

Tasks that need human oversight:

• Anything involving sensitive client data
• Final approvals on external communications
• Financial decisions or commitments
• HR-related conversations
• Anything that requires reading political dynamics in the room

Prerequisites

What you need before starting:

• A dedicated machine or server to run the agent (a Mac Mini, a Linux box, or a cloud VPS). It needs to stay on 24/7.
• Node.js 22+ installed on that machine (download)
• An AI model access via an Anthropic API key, a Claude Pro/Max subscription (setup-token via Claude Code CLI), or an OpenAI API key. Subscriptions provide flat monthly pricing. API keys are pay-per-token. See the OpenClaw provider docs for all options.
• Admin access to your workspace tool (Slack, Google Chat, or Microsoft Teams) to create a bot/app
• IT/security team approval if your organization has policies about third-party integrations or AI tools
• 30-60 minutes for initial setup

Step 1: Install OpenClaw on Your Office Server

Choose a machine that stays on. A Mac Mini in a closet, a Linux server under someone's desk, or a cloud VPS all work. The agent needs to be reachable by your messaging platform.

# Install OpenClaw
curl -fsSL https://openclaw.ai/install.sh | bash

# Run the setup wizard
openclaw onboard --install-daemon

The wizard walks through:

1. API key: Paste your Anthropic or OpenAI key
2. Model selection: Claude Sonnet is the recommended balance of capability and cost for office use
3. Workspace: Accept the default (~/.openclaw/workspace)
4. Gateway port: Default 18789 (fine for internal use)
5. Daemon install: Say yes so it starts automatically on reboot

Verify it's running:

openclaw gateway status

Step 2: Connect to Slack

Slack is the most common office chat platform and the easiest to connect. The setup uses Slack's Socket Mode, which means no public URL is needed.

SLACK_APP_TOKEN=xapp-your-app-token
SLACK_BOT_TOKEN=xoxb-your-bot-token

{
  "env": {
    "SLACK_APP_TOKEN": "xapp-your-app-token",
    "SLACK_BOT_TOKEN": "xoxb-your-bot-token"
  },
  "channels": {
    "slack": {
      "enabled": true
    }
  }
}

Alternative: Connect to Google Chat

For Google Workspace offices, Google Chat is the native option:

1. Create a Google Cloud project and enable the Google Chat API
2. Create a Service Account and download the JSON key
3. Configure the Chat app in the Google Cloud Console with an HTTP endpoint
4. Set the service account file path in OpenClaw config:

{
  "channels": {
    "googlechat": {
      "enabled": true,
      "serviceAccountFile": "~/.openclaw/googlechat-service-account.json"
    }
  }
}

Google Chat requires a publicly accessible webhook URL. Use a reverse proxy or tunnel service if the server is behind a firewall.

Alternative: Microsoft Teams

For Microsoft 365 environments, Teams integration is available through the Microsoft Bot Framework. The setup involves creating an Azure Bot resource, registering the app, and connecting via webhook. Check the OpenClaw docs for the detailed walkthrough.

Step 3: Configure the Office Agent Identity

The agent's workspace files define how it behaves. For an office setting, the instructions need to be professional, clear about boundaries, and focused on team productivity.

SOUL.md (Office Agent Personality)

# SOUL.md - Office Assistant

## Role
I am the office AI assistant. I help the team with research, drafting,
reporting, and coordination tasks. I'm fast, accurate, and always
available in Slack.

## Communication Style
- Professional but approachable
- Concise by default, detailed when asked
- Always cite sources when providing information
- Use bullet points and structured formatting for readability

## Boundaries
- I never share information from private channels in public channels
- I flag sensitive requests for human review
- I don't make commitments or promises on behalf of team members
- I clearly label when something is a draft versus final output
- I ask for clarification when a request is ambiguous rather than guessing

AGENTS.md (Operating Rules)

# AGENTS.md - Operating Rules

## Core Principles
1. Accuracy over speed. Never guess when you can verify.
2. Default to the most conservative interpretation of requests
   involving client data, finances, or external communications.
3. Label all AI-generated content as drafts unless explicitly
   told to finalize.

## Channel Behavior
- In public channels: only respond when @mentioned
- In DMs: respond to all messages
- In project channels: follow channel-specific instructions

## Confidentiality
- Never reference content from one person's DM in another conversation
- Never share information from private channels in public channels
- If asked about another person's work, direct them to ask that person

## Memory
- Record project decisions and outcomes
- Track recurring questions (these should become documentation)
- Remember team preferences (formatting, tone, tools)

USER.md (Office Context)

# USER.md - Office Context

## Organization
- Company: [Your Company Name]
- Industry: [Your Industry]
- Team size: [Number of people who will use the agent]

## Tools We Use
- Project management: [Jira/Asana/Linear/Notion]
- Documents: [Google Docs/Notion/Confluence]
- Communication: Slack
- Code: [GitHub/GitLab]

## Common Requests
- Weekly status reports (due Fridays)
- Meeting summary formatting
- Client communication drafts
- Process documentation lookups

Step 4: Set Up Office Automations

The real value of an office agent comes from automated routines that run without anyone asking.

Daily Standup Summary

Create a cron job that collects yesterday's activity and posts a summary:

openclaw cron add \
  --name "Daily Standup Summary" \
  --cron "0 9 * * 1-5" \
  --session isolated \
  --message "Generate a daily standup summary for the team. Check recent Slack messages in project channels for updates, blockers, and completed items. Format as a clean summary with sections for each team member or project. Post to #standup channel." \
  --delivery announce

Weekly Status Report

openclaw cron add \
  --name "Weekly Status Report" \
  --cron "0 16 * * 5" \
  --session isolated \
  --message "Generate the weekly status report. Summarize this week's accomplishments, in-progress items, and next week's priorities. Use data from project channels and any shared dashboards. Format as a professional report and post to #reports." \
  --delivery announce

Meeting Follow-Up Reminders

openclaw cron add \
  --name "Meeting Action Items Check" \
  --cron "0 14 * * 1-5" \
  --session isolated \
  --message "Check for any meeting action items from the past 48 hours that haven't been addressed. Send a polite DM reminder to the responsible person." \
  --delivery announce

Common Office Automation Ideas

Step 5: Security Configuration (Critical for Offices)

An office agent handles team data. Security is not optional here.

Access Control: Who Can Use It

Restrict which Slack users can interact with the agent:

{
  "channels": {
    "slack": {
      "enabled": true,
      "appToken": "xapp-...",
      "botToken": "xoxb-...",
      "dmPolicy": "allowlist",
      "allowFrom": [
        "U01ABC123",
        "U02DEF456",
        "U03GHI789"
      ]
    }
  }
}

Find Slack user IDs by clicking on a user's profile > "More" > "Copy member ID".

For a more open setup where everyone in the workspace can use it (smaller teams where all members are trusted):

{
  "channels": {
    "slack": {
      "dm": {
        "policy": "pairing"
      }
    }
  }
}

With pairing mode, anyone can message the bot, but they must approve a one-time pairing code first. This is safer than "open" mode and still easy for small teams.

Channel Restrictions

Control which channels the agent responds in:

{
  "channels": {
    "slack": {
      "groups": {
        "C01GENERAL": { "requireMention": true },
        "C02PROJECTS": { "requireMention": false },
        "C03SENSITIVE": { "enabled": false }
      }
    }
  }
}

This means:

• In #general: only responds when @mentioned
• In #projects: responds to all messages (active assistant)
• In #sensitive: completely ignores all messages

Command Restrictions

Limit what the agent can execute on the server. For office use, start with the most restrictive setting and expand only as needed:

{
  "tools": {
    "exec": {
      "security": "deny"
    }
  }
}

This completely disables shell command execution. For most office use cases (answering questions, drafting content, summarizing), the agent doesn't need shell access at all.

If you need limited command access (checking service status, calling internal APIs):

{
  "tools": {
    "exec": {
      "security": "allowlist",
      "allowlist": [
        "date",
        "cal",
        "curl https://api.internal.*"
      ]
    }
  }
}

See the exec approvals docs for fine-grained control including per-command approval prompts.

Data Handling Rules

Add explicit data handling rules to the agent's instructions:

## Data Handling (add to AGENTS.md)

### What I Can Store
- Project names and statuses
- Process documentation
- Meeting action items
- General team preferences

### What I Must Never Store
- Passwords, API keys, or credentials
- Client personal data (PII)
- Financial account numbers
- Health information
- Salary or compensation details

### If Someone Shares Sensitive Data
- Do NOT write it to memory
- Process the request in the current session only
- Remind the user to share sensitive data through secure channels

Step 6: Build an Office Knowledge Base

The agent's memory system becomes a living knowledge base for the office. Seed it with information the team frequently asks about.

Creating the Initial Knowledge Base

Write key information into MEMORY.md:

# MEMORY.md - Office Knowledge

## Company Info
- Founded: 2020
- Headquarters: San Francisco, CA
- Team size: 24

## Processes
### Expense Reports
- Submit via [Expensify/Concur]
- Deadline: 5th of each month
- Approval chain: Manager > Finance
- Receipts required for amounts over $25

### Time Off Requests
- Submit through [HR platform]
- Minimum 2 weeks notice for planned PTO
- Manager approval required
- Emergency PTO: notify manager directly

### Client Onboarding
- Step 1: NDA signing (via DocuSign)
- Step 2: Kickoff call scheduling
- Step 3: Project setup in [tool]
- Step 4: Slack channel creation
- Owner: Account Manager

## Tools and Access
### WiFi
- Network: [CompanyName-Secure]
- Password: Ask IT at #it-help

### VPN
- Software: [Cisco/OpenVPN]
- Setup guide: [internal link]
- Contact for issues: #it-help

## Frequently Asked Questions
### "How do I get access to [tool]?"
Submit a request in #it-help with your manager CC'd.

### "What's the PTO policy?"
[Summary of PTO policy]

### "Where are the brand guidelines?"
[Link to brand guidelines document]

Growing the Knowledge Base Over Time

The agent learns from conversations. When someone asks a question and the answer isn't in memory, tell the agent to "remember this for next time." Over weeks, the knowledge base grows organically from real team questions.

Set up a weekly review:

openclaw cron add \
  --name "Knowledge Base Review" \
  --cron "0 10 * * 5" \
  --session isolated \
  --message "Review this week's conversations. Identify any questions that were asked multiple times. If the answers aren't in MEMORY.md, add them. Also flag any outdated information in memory that needs updating." \
  --delivery announce

Step 7: Train the Team

Deploying the technology is half the work. The other half is getting people to actually use it.

Launch Communication

Send a clear message to the team when rolling out:

Hi team,

We've set up an AI assistant in Slack called @Office Assistant.

What it can do:
- Answer questions about our processes (PTO, expenses, onboarding)
- Draft emails, reports, and summaries
- Set reminders and follow-ups
- Look up information across our tools

How to use it:
- DM it directly for personal requests
- @mention it in channels for team requests
- It only responds when mentioned in public channels

What it does NOT do:
- Access confidential HR or financial data
- Send messages on your behalf without confirmation
- Make decisions - it drafts, you approve

Try asking it: "What's our expense report process?" or
"Draft a client update email for [project]"

Questions? Reach out to [admin name].

Setting Expectations

Usage Patterns That Work

• "Draft this for me": Ask for a first draft of any standard communication
• "Summarize this": Paste meeting notes or long threads for a clean summary
• "How do we...?": Process questions get instant answers from the knowledge base
• "Remind me to...": Set follow-ups without leaving Slack
• "What happened with...?": Get project context from stored memory

Step 8: Monitor and Improve

Track Usage

Check what the agent is being asked and how it performs:

# View recent sessions
openclaw sessions list --limit 20

# Check specific session history
openclaw sessions history --session-key <key>

# View gateway health
openclaw status --all

Monitor Costs

API costs scale with usage. Track monthly spending:

# Check token usage
openclaw status

Typical office agent costs:

These estimates assume Claude Sonnet as the primary model. Opus costs roughly 5x more per token.

Gather Feedback

After the first week, ask the team:

1. What tasks are you using it for most?
2. What did it get wrong?
3. What do you wish it could do?
4. Is anything confusing about how to interact with it?

Use this feedback to update MEMORY.md, refine the AGENTS.md instructions, and add new automations.

The Complete Office Configuration

Here's a production-ready configuration for a Slack-based office agent. Tokens are stored as environment variables (never hardcoded in config files that might be version-controlled):

{
  // Keep tokens in env vars, not inline
  "env": {
    "SLACK_APP_TOKEN": "xapp-your-app-token",
    "SLACK_BOT_TOKEN": "xoxb-your-bot-token",
    "BRAVE_API_KEY": "YOUR_BRAVE_SEARCH_KEY"
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-20250514"
      },
      "heartbeat": { "every": "30m" },
      "timeoutSeconds": 1800
    }
  },
  "channels": {
    "slack": {
      "enabled": true,
      "dm": {
        "policy": "allowlist",
        "allowFrom": ["U01ABC123", "U02DEF456"]
      },
      "groupPolicy": "allowlist",
      "channels": {
        "C01GENERAL": { "allow": true, "requireMention": true },
        "C02PROJECTS": { "allow": true, "requireMention": true }
      },
      "groupAllowFrom": ["U01ABC123", "U02DEF456"],
      "replyToMode": "all"
    }
  },
  "tools": {
    "exec": {
      "security": "allowlist",
      "allowlist": ["date", "cal"]
    },
    "web": {
      "search": {
        "enabled": true
      }
    }
  }
}

After saving the config, restrict file permissions:

chmod 600 ~/.openclaw/openclaw.json

This prevents other users on the machine from reading the config file (which contains API tokens).

## Scaling Beyond One Team

Once the first team is comfortable:

1. **Add more channels**: Invite the bot to additional Slack channels with appropriate mention settings
2. **Create team-specific skills**: Build skills for each department (marketing has different needs than engineering)
3. **Set up per-channel instructions**: Different channels can have different system prompts so the agent adjusts its behavior
4. **Consider multiple agents**: For large organizations, separate agents for different departments prevent context mixing and improve security

```json5
{
  "channels": {
    "slack": {
      "groups": {
        "C01MARKETING": {
          "requireMention": false,
          "systemPrompt": "You are the marketing team assistant. Focus on content, campaigns, and brand voice."
        },
        "C02ENGINEERING": {
          "requireMention": true,
          "systemPrompt": "You are the engineering team assistant. Focus on technical accuracy and code review."
        }
      }
    }
  }
}

What Not to Do

Measuring Success

After 30 days, measure the impact:

Next Steps

• OpenClaw Setup Guide: Personal agent setup for individual use
• How to Build an AI Agent Team: Multi-agent orchestration with coordinator and specialists
• Best AI Tools for Solopreneurs: Complementary tools that work alongside your agent
• Why Every Business Needs an AI Strategy: The business case for AI adoption

Official resources:

• OpenClaw Documentation (complete reference)
• Slack Channel Setup (detailed Slack configuration)
• Google Chat Setup (Google Workspace integration)
• Gateway Configuration (all config options)
• Exec Approvals (command safety controls)
• OpenClaw Discord (community support)

Start with one team, one channel, and two or three specific tasks. Build trust through consistent, accurate results. Then expand. The best office AI deployments don't happen in a big bang launch. They grow organically as the team discovers new ways to use an assistant that never sleeps, never forgets, and never minds being asked the same question twice.