Chapter 8
Play 3: Open the Model to Every AI
Let people choose the interface
Play 3: Open the Model to Every AI
Goal: people can ask about how the enterprise runs — and eventually act on it — from the assistant or chat tool they already use.
How to run it
- Expose the model through MCP. Use a remote MCP server so assistants such as Claude, ChatGPT, Copilot and coding assistants can search the model, follow relationships and read history.
- Sign in as the person. Every connection should go through your normal sign-in, with an explicit consent step, and carry the person's own permissions. No shared accounts, no service users with broad access.
- Read first. Start read-only. Let people get used to asking the model questions before any AI can change it.
- Bring the model into chat. Use Slack or Teams for the moments where the model needs to reach people: the update that needs an owner, the approval that's waiting, the readout before the meeting.
- Link every answer to its source. Answers should carry links to the records they came from, so anyone can check the AI's work.
- Turn on writes deliberately. When you're ready, allow AI to propose and then make changes — status updates, field values, relationships — with drafts shown first and destructive changes confirmed.
A first set of questions to open up
| Who | Asks | From |
|---|---|---|
| CFO | Where are we over budget, and which strategic bets does it touch? | A general assistant |
| Portfolio lead | Which initiatives have stalled, and what depends on them? | A general assistant |
| Architect | Which applications reach end of life next year, and what runs on them? | A coding assistant |
| Risk lead | Which control waivers expire soon, and on which systems? | A general assistant |
| Team lead | (Answers a status request from an agent) | Slack or Teams |
Outputs
- The model reachable from the assistants your people use.
- Chat-tool messages for updates and approvals.
- A written policy for when AI may write to the model.
Signs it's working
- Leaders ask the model before they ask an analyst.
- The same question gets the same answer whichever assistant it's asked in.
Pitfalls
- Opening writes too early. Trust is built on reads.
- Letting each assistant build its own context. If an assistant answers from documents instead of the model, you've recreated the problem.
- Forgetting the people who never open an assistant. Chat-tool messages reach them where they already are.