How to Automate B2B Lead Triage with AI

Direct answer

AI lead triage reads every inbound enquiry, extracts structured facts (company size, need, urgency), scores it against your ideal customer profile, and routes it to the right owner or reply path within minutes. The reliable pattern is a deterministic pipeline with a language model inside one step, plus a human review queue for borderline cases.

Why response time is the lever

Speed to first contact is the most studied variable in inbound sales. A Harvard Business Review study of 2,241 US companies (Oldroyd, McElheran and Elkington, 2011, *The Short Life of Online Sales Leads*) found:

Manual triage rarely sustains a one-hour response outside office hours. That gap is the case for automating the first step. We make no promise about your revenue impact: the study shows a correlation in 2011 data, so measure your own baseline first.

Reference architecture

Form / chat / email
        |
        v
  [1] Capture  ->  validated payload stored in the database (Supabase)
        |
        v
  [2] Enrich   ->  normalise company, domain, role, free-text need
        |
        v
  [3] Classify ->  language model returns JSON: intent, size, urgency
        |
        v
  [4] Score    ->  deterministic rules against your ICP (no model)
        |
   +----+-----------------+
   |                      |
 high score          borderline / low
   |                      |
 [5a] notify owner,   [5b] human review queue,
 propose booking      nurture reply drafted

Two design choices matter more than model quality:

• Keep scoring deterministic. The model extracts facts; plain code applies your thresholds. You can then explain, audit and tune every routing decision.
• Persist before you process. Store the raw submission first, so a failed model call never loses a lead.

Step by step

1. Capture and persist

Validate the form server-side and write the lead to a database row with a status column (new, triaged, needs_review). Chat front ends such as Voiceflow can post the same payload to the same endpoint, so every channel feeds one pipeline.

2. Classify with a strict schema

Ask the model for a fixed JSON shape: intent, company_size_band, urgency, and a short reason. Reject and retry any response that fails schema validation. Never let free-text model output drive routing directly.

3. Score with rules you control

Combine classified fields with firmographic rules, for example: company size band in your target range, a stated need that matches an offer, and a business email domain. Each rule adds points; thresholds decide the route.

4. Route and reply

• High score: notify the account owner immediately and send a reply that proposes a booking slot.
• Borderline: hold in a review queue with the model's reason attached, and draft a reply for a person to approve.
• Low fit: send a polite resource-based reply; log it for later analysis.

5. Close the loop

Record the final outcome of every lead (booked, qualified, lost). Comparing outcomes with the original score is how you tune thresholds and catch drift.

Guardrails to ship on day one

• A human approves any message that makes a commitment on price or scope.
• Log every model input, output and routing decision with a timestamp.
• Minimise personal data sent to the model and document retention in your privacy policy.
• Alert on classification failures and on queue age, not only on errors.

What to measure

If you want this designed around your own lead sources, an AI automation audit maps the pipeline to your stack before anything is built.