Lead qualification

Inbound leads arrive from forms and imports at any hour. Millstream normalises each one, drops duplicates, enriches the company, has an agent score the fit, and puts the promising leads in a sales inbox before anything reaches the CRM.

The pipeline

flowchart LR
  W(["webhook: form"]) --> N[normalise] --> D{{dedupe}} --> E[enrich company]
  E --> Q["qualify (agent)"]
  Q -->|fit at least 60| R[/sales review/] --> C[CRM upsert]
  Q -->|other| U[nurture list]

pipeline: lead-qualification
description: Qualify inbound leads before they reach the CRM

vars: { minFit: 60 }

triggers:
  inbound: { webhook: { rate: 60/min } }

streams:
  leads.normalized: { retention: 180d, redact: [/email, /phone] }

steps:
  normalize:
    action: ./actions/normalize-lead
    in: trigger.inbound
    out: leads.normalized
  dedupe:
    action: dedupe
    in: leads.normalized
    out: leads.unique
    config: { key: "item.emailNorm", window: 90d }
  enrich:
    action: lab/enrich-company@1
    in: leads.unique
    out: leads.enriched
    rate: 30/min
  qualify:
    action: lab/qualify-lead@1
    in: leads.enriched
    concurrency: 4
    out:
      - { when: "item.fit >= vars.minFit", out: leads.promising }
      - { when: "true", out: leads.nurture }
  sales-review:
    action: review
    in: leads.promising
    out: leads.approved
    config:
      inbox: sales
      timeout: 4h
      onTimeout: { escalate: sales-leads }
      editable: [/tier, /owner]
      reasons: [notOurMarket, competitor, student, existingCustomer]
  crm-upsert:
    action: lab/crm-upsert@1
    in: leads.approved
  nurture:
    action: lab/nurture-export@1
    in: leads.nurture

budget: { dayUsd: 20 }

The qualifying agent's manifest sets its model and limits:

name: lab/qualify-lead
version: 1.0.0
kind: agent
effects: none
runtime:
  agent:
    harness: api
    model: claude-haiku-5-5
    network: [api.anthropic.com]
    maxTurns: 2
    maxCostUsd: 0.05

What Millstream does for it

Intake

• Each form post is POST /hooks/lead-qualification/inbound with the trigger's token, rate limited to 60 a minute.
• A sender that retries with the same Idempotency-Key gets back the run it already opened.
• The same address submitted twice within 90 days is filtered by dedupe with reason duplicate.

The agent

• qualify answers a schema with fit, tier and its reasons; an answer that does not match is retried.
• Routes in out send each lead to the first stream whose condition holds, and the threshold stays in the pipeline file.
• dayUsd: 20 caps the agent spend per day, and the per-delivery maxCostUsd caps a single lead.

Sales

• Promising leads wait in the sales inbox; one undecided after four hours escalates to sales-leads.
• A rejection needs a reason code, so notOurMarket and student become counts instead of free text.
• The CRM action is declared effects: external and receives an idempotencyKey per delivery, stable across retries.

How it learns

The insights projector reads the review decisions. When rejections are predictable from a field known before the agent runs, such as a country or a company size, the improve pipeline drafts a topology proposal: a filter placed before qualify, which saves the agent call. Its backtest must show it would have wrongly dropped at most 1% of past approvals. Recurring edits to /tier attributed to qualify become a lessons proposal for that step. Each proposal is drafted in development, validated against evidence exported from production, and accepted by a person.

Figures to watch

Next

Human review · How agents run · Self-improvement