AI Lead Scoring Engine — Automatically Score and Prioritise Every Inbound Lead So Sales Teams Focus Only on the Ones That Will Actually Clos
Built by AetherGlide ·
About this build
A Claude and n8n powered workflow that takes every new inbound lead, scores them across 8 dimensions using company data, behaviour signals and fit criteria, assigns a priority tier and routes high priority leads to sales within 90 seconds of them submitting a form — so no good lead ever sits cold in a CRM waiting to be
What was built
Built this after watching a sales team of 4 people spend their mornings manually reviewing every new lead from the previous day before they could start any actual selling. By the time they had triaged the overnight leads and decided who to call first it was already 11am. Half the leads they prioritised turned out to be early stage researchers with no buying intent. Half the ones they deprioritised turned out to have been ready to buy and had already talked to a competitor. The workflow runs automatically the moment a new lead submits a form anywhere in the marketing stack. Step 1 — Lead ingestion and enrichment When a form is submitted in HubSpot the webhook triggers n8n. The lead's email domain gets passed to Clearbit for company data enrichment — employee count, industry, funding stage, technology stack, annual revenue estimate and location all come back within 2 seconds. This enrichment data combined with the form submission data forms the input for scoring. Step 2 — Lead scoring via Claude Claude scores the lead across 8 dimensions each weighted based on historical conversion data from the company's own closed-won deals. Company fit — does the company size, industry and geography match the ideal customer profile. Scored 1 to 10. Budget signal — does the company's revenue estimate and funding stage suggest they have budget for this type of solution. Scored 1 to 10. Technology fit — does their current tech stack suggest compatibility with or need for this product. Scored 1 to 10. Role fit — is the person who submitted the form in a role that typically drives purchase decisions for this category. Scored 1 to 10. Intent signal — what did the person actually ask for or write in the form. A demo request scores higher than a content download. A specific question about pricing scores highest of all. Scored 1 to 10. Timing signal — are there any signals suggesting active evaluation right now such as recent funding, headcount growth, a new hire in a relevant role, or a job posting for a position this tool would support. Scored 1 to 10. Competitive risk — is there any signal the prospect is already evaluating or using a direct competitor. Scored 1 to 10 where 10 means high competitive risk requiring immediate response. Engagement history — has this person or company engaged with marketing content before, visited the pricing page, attended a webinar. Pulled from HubSpot. Scored 1 to 10. Claude outputs a composite score out of 100, a tier classification of Hot, Warm or Cold, a one paragraph rationale explaining the score in plain English, and 2 to 3 specific talking points the sales rep should use when they reach out based on what the scoring found. Step 3 — CRM update The score, tier and rationale get written back to the lead record in HubSpot automatically. No manual data entry. The lead owner gets assigned based on the tier — senior reps handle Hot leads, junior reps handle Warm, Cold leads go into a nurture sequence automatically without any human touching them. Step 4 — Sales notification for Hot leads Hot tier leads trigger an immediate Slack notification to the assigned rep containing the lead's name and company, the composite score, the tier, the one paragraph rationale and the specific talking points. The rep gets this within 90 seconds of the lead submitting the form. Response time for Hot leads went from an average of 4.2 hours to under 12 minutes. Step 5 — Score calibration via Airtable Every scored lead and their eventual outcome — closed won, closed lost, disqualified, still in pipeline — gets logged to Airtable. After 60 days of data this creates a calibration dataset. The scoring weights get reviewed monthly against actual conversion rates by tier to confirm the model is predicting correctly and adjusted if specific dimensions are underperforming. What broke during development: The role fit scoring was initially too binary — it could only identify exact job title matches to the ideal buyer persona. Fixed by expanding the Claude prompt to reason about organisational influence rather than just title matching. A Head of Operations at a 15 person startup often has more purchase authority than a Procurement Manager at a 2,000 person enterprise even though the title sounds less senior. The timing signal dimension was the hardest to get right. Early versions were flagging too many false positives — companies that had recently raised funding but were in industries completely unrelated to the product. Fixed by adding an explicit instruction to only count timing signals as positive when they combine with strong company fit scores. The Slack notification length was too long initially. Reps were not reading past the third paragraph. Cut the rationale to one paragraph maximum and limited talking points to exactly 3. Shorter notifications get read. Results after 12 weeks: Hot tier conversion rate to qualified opportunity — 67 percent. Warm tier — 23 percent. Cold tier — 4 percent. This confirmed the scoring was working as intended. Average response time to Hot leads dropped from 4.2 hours to 11 minutes. Pipeline generated from Hot leads in the first 60 days was 3.4 times higher than the same period the previous quarter when leads were being worked manually without scoring. The most valuable unexpected output was the calibration data. After 8 weeks the Airtable data showed that technology fit was being underweighted relative to its actual predictive value for closed-won deals. Adjusted the weight and saw an immediate improvement in Hot tier precision. Will share the full n8n workflow JSON, Claude scoring prompt and HubSpot webhook configuration in the comments for anyone who wants to deploy this.