Lead Qualification & Follow-up Engine

Python · Claude Messages API · CRM stub

The problem

In car sales, speed and prioritisation win deals — a lead contacted in ~5 minutes converts far better than one contacted an hour later. But a sales team's inbox mixes a cash-ready buyer in with someone "just looking", so reps burn time sorting instead of calling the hot ones first.

The approach

For each inbound lead, one Claude call:

Extracts BANT signals — budget, timeline, financing, trade-in, vehicle of interest
Assigns a tier — hot / warm / cold against an explicit rubric
Recommends the next action for the team
Drafts a personalised follow-up that names the actual vehicle

It logs lead-shaped data to a CRM and prints a prioritised call queue — hottest first — while flagging fleet orders and trade-in valuations for a salesperson.

A deliberate choice: tiers, not a 1–100 score

An open numeric score wobbles run-to-run (the same lead scores 72, then 68, then 75). So the prompt uses stable, well-defined tiers, and the sort priority is derived in code — the model never invents a number, and the queue is reproducible by construction.

Architecture

inbound lead ↓ orchestrator (BANT rubric + brand voice + few-shot examples) ↓ Claude → schema-enforced JSON ↓ { name, vehicle, budget, timeline, financing, trade_in, ↓ tier, rationale, next_action, draft_followup, needs_human } ├──→ CRM stub (lead-shaped columns) ├──→ prioritised queue (priority derived from tier, in code) └──→ salesperson flag (fleet / trade-in)

Sample run — the prioritised queue

PRIORITISED QUEUE (call hot leads first)
  [HOT ] Jenny Lao          Mitsubishi Xpander
  [HOT ] FastCargo Inc.     fleet (6–8 vans)   → salesperson
  [WARM] —                  Toyota Hilux 4x4
  [COLD] —                  general enquiry

"I want to buy the Mitsubishi Xpander. I have cash ready and I'd like to test drive this Saturday. Please call me. - Jenny Lao"

tier: HOT · budget: strong · timeline: immediateAUTO-DRAFT
Follow-up: "Hi Jenny, thanks for your interest in the Xpander — we'd love to set up that test drive this Saturday. Expect a call shortly, or reach us on (02) 555-0143..."

Measuring & improving it (evals)

The demo ships with a labelled eval harness. It caught an over-escalation bug: the original rule flagged "high-value leads" for a salesperson, so it escalated a cash-ready buyer and a general financing question that should just get a drafted reply. An explicit rule fixed it:

Version                          needs_human   tier      exact-match
Before (vague "high-value")         75% (6/8)   100%      6/8
After (explicit: fleet + trade-in) 100% (8/8)   100%      8/8

Tiering accuracy held at 100% — the fix removed false positives without touching qualification.

Cost & model choice

Defaults to Claude Haiku 4.5 — qualification is a structured judgement over short text, cheap at sales-team volume. Opus 4.8 is one env var away for nuanced, high-value leads.

← Back to work