How to Build an AI Lead Intake Engine for B2B Service Firms
Slow first response kills deals before the pitch. This is the blueprint for an AI lead intake engine that qualifies every inquiry and books the call in under five minutes. Full build on the blog.
This is the blueprint for a system that captures every inbound inquiry, qualifies it with AI, and books a discovery call before a human reads the message. The goal is a five-minute response SLA on every channel, with the closer receiving a prepared brief instead of a raw forwarded message.
The problem
Inbound leads die in the inbox. The pattern repeats across agencies, professional services firms, and manufacturing companies in Southeast Asia. A prospect finds your website, fills out a contact form, or sends a WhatsApp message. That inquiry lands in one inbox monitored by one person. The founder, the sales lead, or an office manager.
That inbox is the bottleneck. During business hours, first response averages two to four hours. After hours, it stretches to the next morning. During a founder trip, it compounds. You return from a two-week trip to stalled proposals and leads who already signed elsewhere.
The cost compounds silently. Every hour of delay reduces the probability of conversion. Buyers interpret slow responses as disinterest or capacity limits. They move on to the next vendor. The revenue you lost was never tracked because the lead never entered the pipeline properly. It sat in an inbox, aged, and went cold.
The root cause is structural. You built the business around personal responsiveness. That worked when you had five clients and one inbox. It breaks when you have prospects across three channels and a team that depends on you for every response.
Who this is for
This blueprint is for B2B service founders in Southeast Asia running businesses between $500k and $5M in annual revenue. Agencies, professional services firms, manufacturing companies with a consultative sales process.
You fit this build if you receive 5 to 50 inbound inquiries per week across email, web forms, and WhatsApp. Your sales team is one to five people. First response time is inconsistent and depends on who is available. You have a CRM in place, even a basic one.
You do not fit this build if you run a high-volume transactional business with thousands of low-value leads per day. You need a different architecture. You also do not fit if you have no CRM and no intention of installing one. The intake engine requires a data backbone to store, route, and track lead records. Finally, if your sales cycle is under 24 hours with no qualification step, speed alone solves your problem. You do not need the AI layer.
System architecture
The AI Lead Intake Engine is a pipeline with four stages: capture, qualify, route, and book. Each stage is a discrete component with clear inputs and outputs.
Capture. Every inbound channel feeds into one entry point. Web form submissions trigger a CRM webhook. Email inquiries forward to a dedicated intake address parsed by an email API. WhatsApp messages route through the WhatsApp Business API. The system normalizes all channels into a single lead record with these fields: name, company, email, phone, channel, message body, and timestamp.
Qualify. The lead record enters an AI qualification step. The model reads the inquiry text and returns a structured assessment. It classifies intent into three buckets: informational (researching), evaluation (comparing options), or decision-ready (ready to talk). It scores fit against your ICP criteria on a 1 to 10 scale. It extracts key data points: company size, industry, stated budget, timeline. The qualification step runs in under 10 seconds.
Route. Based on the qualification output, the system applies deterministic routing rules. Decision-ready leads with strong ICP fit go to the closer immediately. Evaluation-stage leads enter a nurture sequence. Low-fit leads receive a polite redirect. Routing rules are code that you write and control. The model reads and classifies. You decide what happens next.
Book. For qualified leads, the system sends a calendar link and books a discovery call. The closer receives a notification with the full lead brief: inquiry text, qualification score, extracted data points, and suggested talking points based on intent classification. The brief arrives in Slack or your team chat.
The full flow: inquiry enters, system qualifies, system routes, system books, closer receives brief. Total elapsed time from inquiry to booked call: under five minutes. The human touches the process only at the discovery call itself.
Build steps
1. Install the data backbone
Set up a CRM that accepts webhooks and stores lead records. HubSpot, Pipedrive, or a Supabase table all work. Define your lead record schema: name, company, email, phone, channel, message body, intent score, ICP fit score, status, created timestamp. Every inbound inquiry creates exactly one record. This is the foundation. Every subsequent component reads from and writes to this record.
2. Configure channel ingestion
Connect every inbound channel to the CRM. Set up your web form to POST on submission. Forward intake emails to an email parsing service that extracts sender, subject, and body, then sends the parsed data to the CRM webhook. Connect WhatsApp Business API to forward inbound messages to the same endpoint. Test each channel end to end. Confirm that a lead record appears in the CRM within seconds of submission on each channel.
3. Deploy the orchestrator
Install n8n on a VPS. Create a workflow that triggers on every new lead record in the CRM. The workflow executes four steps in sequence: call the AI qualification step, apply routing rules, execute the routing action, and trigger the calendar booking if qualified. Self-hosted n8n runs on your existing VPS with no additional monthly cost.
4. Build the AI qualification prompt
Write a system prompt that accepts the inquiry text and returns a structured JSON object. The JSON contains: intent classification (informational, evaluation, or decision-ready), ICP fit score (integer 1 to 10), extracted data points (company size, industry, budget, timeline if stated), and a one-paragraph brief for the closer. Constrain the model to only extract data explicitly present in the inquiry. Test the prompt against 20 real past inquiries. Compare the model output to your own manual qualification. Iterate until accuracy is consistent across the test set.
5. Set deterministic routing rules
Write routing rules in the orchestrator based on the qualification JSON. If intent is decision-ready and ICP fit is 7 or above, route to the closer and send the calendar link in the first response. If intent is evaluation, enter the three-touch nurture sequence. If intent is informational, send a knowledge base link. If ICP fit is below 4, send a polite redirect. These rules are explicit code. The model does not make routing decisions. You do.
6. Integrate the calendar
Connect Cal.com to the orchestrator workflow. When a lead qualifies for a discovery call, the system includes a scheduling link in the first response email or WhatsApp message. Configure the closer calendar with buffers between calls and reasonable blackout windows. Cal.com handles timezone conversion, conflict prevention, and confirmations. The system books the call without human intervention.
7. Build the first response template
Write a response template in your brand voice. The template has three parts: acknowledgment of the inquiry, confirmation of next steps, and the calendar link for qualified leads. The AI model fills in specifics from the inquiry context, referencing the product asked about and acknowledging the stated need. Review the first batch of responses manually before allowing the system to send autonomously. This is the quality gate.
8. Build the closer brief
When a call is booked, the system compiles a brief for the closer. The brief includes: the original inquiry text, the qualification score and intent classification, the extracted data points, the channel, and suggested talking points derived from the intent classification. The brief arrives via Slack or your team chat. The closer walks into the call with context, not with a raw forwarded email.
9. Install the nurture sequence
For evaluation-stage leads, build a three-touch sequence. Touch one sends immediately: a personalized response acknowledging their evaluation and offering a resource. Touch two follows on day three with a relevant case study or comparison. Touch three arrives on day seven with a direct call to action to book a call. The model drafts each touch using the lead context. The sequence caps at three touches. If the lead responds at any point, the sequence pauses and the closer takes over.
10. Set up monitoring and review
Build a dashboard tracking three metrics weekly. First response time: average across all channels, target under five minutes. Booking rate: track the ratio of qualified leads who book a call and watch for drops week over week. Qualification accuracy: spot-check leads per week and compare the model assessment to the closer post-call judgment. Review every Monday. The dashboard tells you where the system leaks before it costs you deals.
Tools and costs
CRM (HubSpot Free or Pipedrive Essential). Stores lead records, accepts webhooks, tracks pipeline stages. HubSpot Free covers up to 1,000 contacts at $0 per month. Pipedrive Essential is approximately $14 per month per seat. Pick one. The CRM is the data backbone.
n8n (self-hosted on your VPS). The orchestrator. Runs every workflow, handles routing logic, triggers the model and calendar integrations. Cost: $0 per month beyond your existing VPS. Open source, self-hosted, no per-execution fee.
AI Model (Gemini Flash via API). Runs the qualification prompt on each inbound inquiry. Estimated cost: $0.001 to $0.01 per call depending on input length. At 50 inquiries per week, monthly cost is under $2. This is an estimate based on current published API pricing. Verify pricing before deployment.
Cal.com (self-hosted or free tier). Handles scheduling, sends confirmations, syncs with the closer calendar in real time. Cost: $0 per month for the self-hosted version. The free cloud tier also works for a single closer.
WhatsApp Business API. Required only if WhatsApp is an inbound channel. Per-conversation pricing varies by country, approximately $0.05 per conversation in most Southeast Asian markets.
Total estimated monthly cost: $2 to $10 for a firm receiving 50 inquiries per week. The system runs on infrastructure you likely already have. The AI cost is negligible at this volume.
Failure modes
The model invents data. The qualification prompt may produce budget numbers or company details that the lead never stated. This is the most common failure. Fix: constrain the prompt to extract only data explicitly present in the inquiry text. Add a validation step that compares extracted fields against the source text. Flag any extraction with no matching source text for manual review.
Over-messaging kills the relationship. The nurture sequence sends too many touches or the tone drifts from helpful to aggressive. Fix: cap the sequence at three touches over ten days. Monitor reply rates weekly. If a lead responds at any point, pause the sequence immediately and hand off to the closer. Never let an automated sequence argue with a prospect.
The webhook fails silently. A channel goes down and leads stop creating records. Nobody notices for days because the failure is invisible. Fix: set up a daily reconcile job that counts inbound messages per channel. If any channel drops to zero for 24 hours, send an alert to Slack. Test the full webhook path weekly with a synthetic lead.
The calendar double-books. Two leads book the same time slot due to a sync delay between the calendar API and the closer calendar. Fix: use Cal.com or Calendly with real-time two-way calendar sync. Set a buffer between calls. Let the scheduling tool handle conflict resolution rather than building custom logic.
The first response feels robotic. The model drafts a generic reply that signals automation to the buyer. Fix: write a response template with specific brand voice markers. The model fills in details but follows the template structure strictly. Run the first batch of responses through manual review. If a response feels generic, revise the template before going autonomous.
What good looks like
The installed system has measurable properties. Every inbound inquiry receives a first response within five minutes, on every channel, during and after business hours. The closer walks into every discovery call with a prepared brief that includes the inquiry, the qualification score, and suggested talking points. The nurture sequence runs on schedule. The dashboard shows first response time, booking rate, and qualification accuracy every week.
The cost per intake is under $0.10. The system runs while the founder travels, sleeps, or closes deals in person. Revenue tracks the pipeline, not the founder calendar.
The intake engine does not replace the closer. It removes the latency, the manual triage, and the single point of failure. The closer spends time closing, not reading inboxes.