Case study: how we'd build an AI agent for quotes

This is a reference implementation (not a delivered project): how we'd build an AI agent that produces quotes for a typical make-to-order business — say a joinery making windows and doors. Goal: the owner describes a job in Claude Desktop and gets back a numbered PDF quote. The key principle from the previous article holds: a tool computes the price, not the model.

Architecture

Claude Desktop  ──MCP/HTTP──▶  MCP server (Cloudflare Worker)
                                  ├─ pricing rules (deterministic)
                                  ├─ data: D1  OR  Intrix CRM (via API)
                                  └─ PDF generator (+ R2 storage)

The agent only gathers parameters and calls tools; all logic and data stay on the server, under control.

Step 1 — the pricing model

First we turn the price list into rules. For a window, for example:

area_m2 = (width_mm * height_mm) / 1_000_000
price = area_m2 * price_m2[material]       // wood, alu-wood …
      + glazing[type]                      // single/double/triple
      + hardware
      + (install ? install_m2 * area_m2 : 0)

The pricing tables (price_m2, glazing, install_m2) are data you can edit without touching code. This is 80% of the project's value.

Step 2 — MCP tools

The server exposes a few tools over MCP. calculate_quote with a clear schema:

{
  "name": "calculate_quote",
  "inputSchema": {
    "type": "object",
    "properties": {
      "items": { "type": "array", "items": {
        "type": "object",
        "properties": {
          "type": { "enum": ["window", "door", "fire_door"] },
          "width_mm": { "type": "integer" },
          "height_mm": { "type": "integer" },
          "material": { "enum": ["wood", "alu-wood"] },
          "glazing": { "enum": ["single", "double", "triple"] },
          "quantity": { "type": "integer" },
          "install": { "type": "boolean" }
        }, "required": ["type", "width_mm", "height_mm", "quantity"]
      }}
    }, "required": ["items"]
  }
}

It returns priced line items and a total (with and without VAT). A second tool, create_quote, saves the quote, assigns a number and returns a PDF.

Step 3 — storage and PDF

A simple table in D1:

CREATE TABLE quotes (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  number TEXT UNIQUE, customer TEXT,
  items TEXT,          -- JSON
  total_net REAL, total_gross REAL,
  created_at TEXT DEFAULT (datetime('now'))
);

The PDF is built from a template with their logo, terms and validity; stored in R2 and returned as a link.

Step 4 — connecting Claude Desktop

No local server — Claude Desktop connects over HTTP with a bearer token (same in Claude Code):

claude mcp add --transport http quotes https://example.com/mcp \
  --header "Authorization: Bearer <TOKEN>"

Option: integration with Intrix (existing ERP)

If the company already uses Intrix (a CRM with the inquiry → quote → deal path), the agent doesn't need to build a separate customer directory. The MCP server connects to Intrix via its API:

• find_customer — reads customers from the Intrix CRM (no data duplication).
• create_quote — writes the quote back into Intrix as a quote linked to the customer, so it stays in the existing sales pipeline.

The pricing calculation stays in the MCP server (deterministic rules), since make-to-order prices are often not part of the ERP. You get a single source of truth — the sales process in Intrix is unchanged; AI only shortens the step from description to quote. The integration scope depends on Intrix's API/integration options, which we align with the vendor (Intera).

Example flow

Owner: Quote for Novak — 3 wooden windows 1200×1400 double glazing with installation, 1 fire door 900×2100.

1. Claude calls calculate_quote with three line items → gets prices and a total.
2. Claude calls create_quote → quote 2026-041, PDF link (and a record in Intrix, if connected).
3. The owner reviews the PDF, approves and sends. A minute in total.

Security and phases

• Deny-by-default bearer token; the agent gets only the quote tools.
• Customer data in the EU (GDPR).
• Phases: (1) capture the price list, (2) calculate_quote + create_quote + PDF, (3) Claude Desktop and onboarding, (4) CRM/email.

The result is a consistent, numbered, stored process — a draft a human approves. That's exactly our AI integration service.

More about AI and MCP ↗ · Why the model must not price ↗