AI Workflow

How to Give AI Your Business Context: CRM, Slack and Docs

By Gabriel Ceicoschi

April 6, 2026

AI Workflow AI for Non-Engineers AI Tools for Business AI Automation Claude Code AI Productivity Developer Productivity AI Agents AI Workflow Automation
Aligned openings in three pale planes reveal an orange block, representing connected business context.

The real problem

Here's where I see people get stuck:

  1. Open ChatGPT or Claude
  2. Try to explain the situation from scratch
  3. Get surface-level advice because the AI is missing most of the story
  4. Give up and go back to doing it manually

The AI isn't broken. It just can't see where you work.

Your real work is in:

  • Word docs and Google Docs with your team's comments
  • Slack threads where decisions got made
  • Your CRM with client history and past conversations
  • Templates you've refined over years
  • Email chains with context buried 15 replies deep
  • Spreadsheets with your calculations and notes

AI lives in a chat window. These worlds don't talk.

That's the gap.

Every trip between tools costs you focus. Gloria Mark, a professor of informatics at UC Irvine, told CNN that people now focus on one screen for an average of 47 seconds, and that it takes about 25 minutes to return to the original project after switching away.


The fix: let AI read your tools

Instead of copying everything into a chat, flip it around.

Give AI permission to read your actual apps.

This exists right now. It's called Model Context Protocol (MCP). Anthropic introduced it as an open standard and donated it to the Agentic AI Foundation under the Linux Foundation in December 2025. Claude, ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code all support it.

In Claude, an MCP connection is called a connector. In ChatGPT, it's called an app.

Without connectors With connectors
You copy and paste context into the chat AI reads the tool directly
Context limited to what you pasted AI looks up what it needs when you ask
Stale as soon as the data changes Reads the current version each time
Generic responses Responses grounded in your records

Broken workflow:

  1. A client asks a question in Slack
  2. You dig through the CRM for their account
  3. You open the proposal doc to check what you promised
  4. You paste it all into a chat
  5. You ask for a response
  6. You edit it because the AI missed half the context anyway

With context:

  1. "Claude, draft a response to the client's question in #client-acme"
  2. Claude reads the Slack thread
  3. Claude checks the CRM
  4. Claude opens the proposal
  5. Claude drafts something that fits
  6. You review, tweak and send

AI does the legwork. You keep the judgment.


The plan: 4 steps to context-aware AI

You don't need to be technical. Here's the framework I use.

Step 1: Map where your context lives

Write down where the information you need actually exists:

  • Client context: CRM (HubSpot, Pipedrive, a Notion database), email history
  • Team decisions: Slack, Teams, project management tools
  • Work in progress: Word docs, Google Docs, shared drives
  • Templates and standards: Your proposal templates, report formats, checklists
  • Knowledge base: Your Notion wiki, internal docs, SOPs

You're not changing your tools. You're making them visible to AI.

Step 2: Choose what the AI can access

You control permissions. Start narrow:

  • Read first: In Claude, open a connector under Customize > Connectors and set any tool that writes or deletes to Blocked
  • One tool at a time: Turn on only the connectors a task needs. You can switch each one on or off per chat from the + menu
  • Non-sensitive data: Start with internal processes, not client financials

Connectors that use your own sign-in usually see what your account already sees. HubSpot, for example, respects your existing HubSpot permissions. If you want the AI to see less than you do, connect a separate account with narrower access.

Expand access later. Start with what makes you comfortable.

Step 3: Connect the AI to those tools

This is where MCP comes in. Three approaches:

Option A: Add a connector from the directory (easiest)

In Claude on the web or the desktop app:

  1. Open Customize > Connectors
  2. Find the tool, such as Slack, Notion, Google Drive, HubSpot or Microsoft 365
  3. Click Connect and sign in to that tool

On Team and Enterprise plans, an Owner enables the connector for the organization first. Some tools also need their own admin to approve. Slack asks for workspace admin approval, and Microsoft 365 needs a business tenant and a one-time consent from a Microsoft Entra Global Administrator.

No coding required. ChatGPT works the same way through its apps, though which apps and actions you get depends on your plan.

Option B: Add a custom connector by URL

If your tool has a remote MCP server that isn't in the directory, go to Customize > Connectors, click Add custom connector and paste the server URL. Claude's docs cover the details. The Free plan allows one custom connector.

Option C: Build a connector (for unique workflows)

  • Hire someone or work with us to build a connector for your specific CRM or industry software
  • Plan for upkeep. APIs and the MCP spec change, so someone needs to own it after launch

If the data you need sits in your ERP instead of a tool with a ready-made connector, the setup is bigger. We describe how we built one safely in what it takes to build an AI system on your own sales data.

Step 4: Test and expand

Start with one workflow:

"When a client asks a question in Slack, draft a response using:

  • The CRM record for context on their account
  • The proposal doc to check what we agreed to
  • Our internal wiki for our standard process"

Once that works, expand:

  • Generate reports from CRM data and templates
  • Draft proposals using past proposals as examples
  • Summarize Teams meeting chats into action items

Real examples: context integration for consultants

These are illustrative workflows. Your results depend on how clean your data and templates are.

Proposal writing (Financial Advisor)

Without context:

You copy client info from the CRM, paste parts of last year's proposal, adjust the numbers and rewrite the personal sections.

With context:

Prompt: "Draft a financial planning proposal for this client using:

  • Their profile in HubSpot
  • Last year's proposal for a similar client as a template
  • Our standard fee structure from the wiki"

Claude pulls all three sources and writes a first draft. You check the numbers and add the personal touches.

Client follow-up (Legal Consultant)

Without context:

A client emails about invoice terms. You search your inbox for the original contract, scan for payment terms and draft a response.

With context:

Prompt: "Draft a response to this email using the payment terms from Contract_ABC_2025.docx."

Claude reads the contract, finds the clause and drafts a response. You confirm the clause before you send.

Report generation (Energy Auditor)

Without context:

You run the site audit, collect measurements, fill the report template by hand from Excel and write the same compliance sections every time.

With context:

Prompt: "Generate an energy audit report using:

  • Site data from Audit_2025_04_Site_B.xlsx
  • Site notes from the Google Drive folder Site_B
  • Standard compliance text from AuditTemplate_v3.docx"

Claude builds the draft. You verify the measurements and add your expert commentary.

Team coordination (Agency Owner)

Without context:

A designer asks in Slack about a client's brand guidelines. You search Notion, find the doc and paste the relevant section.

With context:

You prompt: "Answer the designer's question in #project-acme using the brand guidelines from Notion."

Claude reads the Slack message, searches Notion and drafts a reply that quotes the guideline.


What this actually feels like

Once AI can read your context:

You stop being a search engine for your own information.

Instead of:

  • "Let me find that doc"
  • "Give me a sec to check the CRM"
  • "I'll look through Slack and get back to you"

You say:

  • "Claude, pull the details from the CRM and draft a response"
  • "Summarize the decision from last week's Slack thread"
  • "Generate the report using the template and the client data"

You stay in the decision-making layer. AI handles retrieval and drafting.

You're not rushing through grunt work. You're reviewing drafts that already have the right context.


Common mistakes

Mistake 1: Giving too much access too fast

Start narrow. One tool, one workflow, write tools blocked. Expand once you trust it.

Mistake 2: Not labeling your sources

If your templates are named Final_FINAL_v3.docx, the AI won't know which one to use. Clean up naming conventions first.

Mistake 3: Treating AI output as final

The AI gives you a draft. You add judgment, tone and human touches. Don't skip the review step.

Mistake 4: Skipping the "why"

When you prompt, tell it why you need the information. "Draft a response to calm an upset client" gets better output than "Draft a response."

Mistake 5: Connecting servers you don't trust

A connector can read data and, unless you block it, take actions in that tool. Only connect servers from organizations you trust, and review what each one asks for when you sign in.


For software teams: technical implementation

If you're a developer or technical team lead, here's how I'd set up context integration with MCP today.

Setup: GitHub + Slack + Figma

GitHub, Slack and Figma now run official remote MCP servers. You don't install npm packages anymore. The old @modelcontextprotocol/server-github and @modelcontextprotocol/server-slack reference packages are archived and deprecated on npm, so don't use them in new setups.

Servers:

Claude Code:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ -H "Authorization: Bearer YOUR_GITHUB_PAT"
claude mcp add --transport http slack https://mcp.slack.com/mcp
claude mcp add --transport http figma https://mcp.figma.com/mcp

Slack and Figma open a browser sign-in the first time you use them. Figma also offers a plugin: claude plugin install figma@claude-plugins-official.

Claude web and desktop:

Add Slack from Customize > Connectors. You can try other remote servers as custom connectors by URL, but each provider decides which clients it accepts. Figma, for example, only allows clients listed in its MCP catalog, so check Figma's docs before you rely on it outside Claude Code.

If you need a server that runs on your own machine, use a desktop extension under Settings > Extensions in Claude Desktop. For GitHub, the local option is the ghcr.io/github/github-mcp-server Docker image with a GITHUB_PERSONAL_ACCESS_TOKEN.

Usage patterns

PR review with full context:

Prompt: "Review PR #234 in repo/project. Check:
- The Figma design spec linked in the description
- Related Slack discussion in #engineering
- Similar PRs from the last month

Focus on: design fidelity, edge cases, test coverage."

Claude fetches the PR diff, loads the Figma design, searches Slack, finds related PRs and writes review comments with context from all of them.

Feature spec from Slack + Figma:

Prompt: "Write a technical spec for the feature discussed in #product-ideas (thread from April 3).
Use the Figma mockup: figma.com/file/abc123
Reference our standard spec template from GitHub: docs/templates/feature-spec.md"

Claude pulls all three sources and drafts the spec.

Incident response:

Prompt: "Summarize the production incident from:
- Slack #incidents (last 2 hours)
- GitHub issues with label 'incident'
- Related PRs merged in the last 48 hours

Format: timeline, root cause hypothesis, action items."

Claude correlates events across tools and builds an incident summary.


What comes next

Once AI can read your context:

You stop doing retrieval work. You start doing judgment work.

The AI becomes an assistant that knows what's going on. It doesn't replace your expertise. It handles the grunt work so you can focus on the decisions only you can make.

This isn't the future. It's available right now.

The question is: will you connect your AI to where you actually work, or keep copying and pasting context into a chat window?

Frequently Asked Questions

How do I connect AI to my CRM?

Check whether your CRM has a connector. HubSpot has an official Claude connector that respects your HubSpot permissions. In Claude, open Connectors in the Customize menu, find your CRM, click Connect and sign in. Then test it by asking Claude to pull one client record. If your CRM isn't listed, look for a remote MCP server URL you can add as a custom connector.

Can ChatGPT read my Slack messages?

Yes, if you connect it. ChatGPT now supports MCP through apps and developer mode, and Slack runs an official MCP server. Which apps and actions you get depends on your ChatGPT plan, and your Slack workspace admin may need to approve the connection first. Claude connects to Slack the same way from its Connectors page.

What is Model Context Protocol (MCP)?

MCP is an open standard that lets AI assistants connect to external tools and data. Anthropic introduced it and, in December 2025, donated it to the Agentic AI Foundation under the Linux Foundation. Claude, ChatGPT, Gemini, Microsoft Copilot and others support it. Think of it as giving your AI assistant a badge to read your files instead of you copying everything manually.

Is it safe to give AI access to my business data?

You control what it can reach. Connectors that use your sign-in usually see only what your account can see. In Claude you can block individual tools, such as anything that writes or deletes, and switch connectors on or off per chat. Only connect servers from organizations you trust, and start with non-sensitive data.

Do I need to know how to code to set this up?

No. For tools in the Claude connectors directory, such as Slack, Notion, Google Drive and HubSpot, you click Connect and sign in. On Team and Enterprise plans an Owner enables the connector first. You only need technical help to build a connector for a tool that doesn't have one.

Where would AI help your team?

Tell us about your team and the work you want to change. We'll help you pick the right next step.