AI Lead Qualification & Routing for a B2B SaaS Platform

Challenge

Inbound demo requests were triaged by hand in a shared inbox. Follow-up took over a day on average and sales had no consistent definition of a qualified lead.

Business problem

Demand was growing faster than the team could triage it. Because qualification happened in a shared inbox, response time depended on who was online, and no two people applied the same qualification standard. Marketing could not report reliably on lead quality, and sales discounted marketing-sourced pipeline because the definition of a qualified lead was inconsistent.

Solution

A scoring and routing layer combining firmographic fit with an LLM classifier reading the free-text request, writing directly into the CRM with SLA-backed assignment.

Architecture

• Form and chat submissions post to a webhook endpoint rather than directly to the inbox
• An enrichment step appends firmographic and technographic data to the record
• An LLM classifier extracts intent, use case and buying-stage signals from the free-text request
• A scoring service combines fit and intent into a single score with explainable contributions
• Routing rules write owner, score rationale and SLA timer directly to the CRM record
• Slack notifications and escalation fire from SLA timers, with all steps logged for audit

Technology stack

Implementation

• Audited two years of closed-won and closed-lost records to define fit criteria
• Built and validated the classifier against 500 human-labelled requests
• Ran routing in shadow mode for two weeks before switching it live
• Handed over documentation and a prompt-update runbook

Results

• First-response time reduced from over 24 hours to under 15 minutes
• Manual triage removed from the daily workload of two people
• Consistent MQL definition adopted by marketing and sales

Business impact

Sales capacity was redirected from triage to conversations, and pipeline reporting became consistent enough to plan against.

Lessons learned

• Shadow mode was the single most valuable phase: two weeks of comparing model decisions to human ones removed all internal resistance to switching over.
• Explaining the score mattered more than the score itself — sales trusted the system once it could show why a lead ranked highly.
• Deterministic rules handled most routing well; the model only earned its place on the free-text interpretation step.