# Build a Broker Copilot

> Use this recipe to scope the architecture, tool flow, and review checkpoints for an AI workflow that helps brokers move deals forward.

Source: https://www.lev.com/docs/build/broker-copilot

Last updated: May 2026

---
## Target Workflow

The minimal useful copilot usually answers three questions:

1. What is the current state of the deal?
2. What has already been sent or discussed?
3. What should the broker do next?

## Recommended Architecture

**Core building blocks**

- **Primary records:** Deals, deal financials, deal properties, placements, term sheets, and account/team context.
- **AI-native workflow:** Use MCP for interactive tooling or use a server-side orchestrator that can call the Lev API deterministically.
- **Decision layer:** Let the model summarize trade-offs, but keep writes and outbound actions behind explicit approval.
- **Review layer:** Show the retrieved records and the exact proposed action before the user confirms it.

## Suggested Tool Sequence

Pull the deal record alongside deal financials.

Fetch recent placements and any available term sheets for the deal.

Summarize the opportunity and identify the next recommended broker action.

If a write is required, show the exact change and request confirmation.

## Human Review Points

  Use the model for synthesis and recommendation. Keep the broker responsible for final approval whenever a placement, status change, or outbound communication could affect a live relationship.

- Review before creating or updating placements.
- Review before summarizing or comparing term sheets for external sharing.
- Review before any outbound lender communication is triggered by an agent workflow.