AI Automation • Zapier • CRM Intelligence

Top Air Ductboard Fabrication — Lead Qualification & AI Routing System

An AI-powered inbound lead triage system that classifies new HubSpot contacts into five operational paths, updates the CRM, logs an audit trail, sends context-aware emails, and escalates complex fabrication work for human review.

Project: IsraelPlatform: ZapierHubSpotOpenAI GPT-4o-miniGoogle SheetsMandrillGmail

Project Overview

Business ProblemNew fabrication inquiries required manual sorting, qualification, follow-up and escalation.

Automation GoalClassify every new lead consistently and route it into the correct operational workflow.

Human-in-the-loopComplex, custom or high-risk jobs are surfaced for fabrication/sales review instead of being guessed by AI.

What the System Automates

Polls HubSpot for contacts whose lead status is exactly New.
Uses GPT-4o-mini to classify each lead into one of five predefined buckets.
Routes each bucket through a dedicated CRM + logging + email sequence.
Uses GPT-4.1-nano for plan/file takeoff analysis when a plans URL is available.
Logs structured AI outputs and lead data to Google Sheets for auditability.
Uses Mandrill templates for automated customer communication and Gmail to detect replies after 24 hours.

Architecture

HubSpot Contacts

→

Zapier Poll

→

New Status Filter

→

GPT-4o-mini

→

5-Path Router

HubSpotTrigger/source of truth for lead identity, fabrication fields, lifecycle and automation status.

ZapierOrchestrates filters, AI steps, conditional paths, delays and application connectors.

OpenAIGPT-4o-mini performs lead classification; GPT-4.1-nano performs takeoff analysis when applicable.

Google SheetsMaintains a structured audit row containing contact data, AI outputs and path-specific analysis.

MandrillSends templated customer communication for missing information, plans, manual review, quotes and reminders.

GmailChecks for a lead reply after a 24-hour delay before triggering a reminder.

AI Classification Engine

Decision	Operational meaning	CRM outcome
DECLINED / SPAM	Fake, invalid, gibberish or off-topic inquiry	UNQUALIFIED; audit only; stop
MANUAL_REVIEW_REQUIRED	Custom, conflicting, unclear or high-risk scope	ATTEMPTED_TO_CONTACT; human review
NEED_MORE_INFO	Standard request but critical fields are missing	IN_PROGRESS; request details
TAKEOFF_PENDING	Plans/drawings or broader scope needed	IN_PROGRESS; takeoff or plans request
READY_FOR_QUOTE_REVIEW	Complete and quotable request	OPEN_DEAL + opportunity

End-to-End Workflow

1. Trigger & Gate

Zapier polls HubSpot and continues only when hs_lead_status = New. Other contacts stop immediately.

2. AI Classification

The classifier evaluates identity, fabrication, dimensions, quantity, material, plans and notes using a fixed decision order.

3. Route

Exactly one of five paths executes for each lead, preventing overlapping customer actions.

4. Execute

Each path updates HubSpot, writes to Sheets and sends the appropriate email—or stops safely when no email is required.

5. Follow-up

For applicable paths, the workflow waits 24 hours, searches Gmail for a lead reply, then sends a reminder only when a message is found.

6. Human Review

Fabrication/sales teams use risk flags, missing information and summaries to resolve complex or incomplete jobs.

Path Details

Path	Typical reason	Automated actions
A — Manual Review	Custom fittings, unclear geometry, conflicting specs, high-risk scope	HubSpot → Sheets → acknowledgment email → 24h reply check → reminder if reply found
B — Need More Info	Missing dimensions, quantity, material/gauge or openings	HubSpot → Sheets → missing-info email → 24h reply check → reminder
C — Takeoff Pending	Plans required or scope cannot be determined from form	With file: GPT-4.1-nano analysis + Sheets + HubSpot. Without file: request plans + follow-up.
D — Ready for Quote Review	Complete, clear and quotable fabrication request	Promote to opportunity → Sheets → quote/payment-link email → 24h follow-up
E — Declined / Spam	Fake, invalid or off-topic inquiry	HubSpot UNQUALIFIED → audit log → stop; no customer email

Takeoff Analysis Layer

When a TAKEOFF_PENDING lead includes a website/file URL, GPT-4.1-nano analyzes the file and returns structured fields such as quote status, price-match status, item summary, quote total or reason it cannot be determined, assumptions, missing variables, margin risk, and a professional customer message.

Safety principle: the documented AI rules explicitly prohibit guessing dimensions, quantities, scope or pricing.

Business Impact

Faster ResponseNew inquiries can move from intake to a relevant next step within minutes rather than waiting for manual sorting.

Sales FocusHuman attention is concentrated on manual-review, takeoff and quote-ready cases rather than every new inquiry.

Consistent RoutingFixed decision strings and path gates create deterministic downstream behavior.

AuditabilityEvery processed lead gets a structured Sheets record with classification and operational outputs.

Spam ControlSpam is marked UNQUALIFIED and exits without outbound customer email.

Follow-up AutomationApplicable paths perform a 24-hour Gmail reply check before sending reminders.

Operational Reliability

Exact classification strings are preserved because path filters depend on them.
Only contacts with status New are processed, preventing repeated handling of old records.
AI is instructed to return facts only and to choose a conservative classification when uncertain.
Production prompt changes are intended to be tested against five known fixtures before replacement.
Day-2 operations include daily Zapier history review, HubSpot queue review, Sheets risk-flag review, weekly spam false-positive checks and template deliverability checks.

Workflow Screenshots

Demo Recording

Technology Stack

ZapierHubSpotOpenAIGoogle SheetsMandrillGmailAI by ZapierConditional Paths24h Delay

Implementation Notes

The system uses standard Zapier application connectors rather than custom webhooks or Code-by-Zapier steps. The documented architecture is intentionally modular: a shared trigger/filter/classifier/router spine feeds five isolated operational paths.

The source KT documents approximate volume distribution of roughly 10–15% spam, 15–20% manual review, 20–25% missing information, 25–30% takeoff pending and 20–25% quote review; these are documented expectations, not measured performance claims.