Lead Scoring Fails When Intent Lives in Three Places

Most “lead scoring” is just a spreadsheet cosplay that collapses the moment the first reply comes in from a real person, because the data that should drive the next step is trapped in three places: the form tool, the inbox, and somebody’s memory. The result isn’t bad copy or weak demand gen. It’s operational latency. This playbook builds a working lead-intel loop that turns inbound form fills into an enriched, prioritized, and routed record in under two minutes, with a reason attached to every decision. Tools: Webflow (capture), n8n (orchestration), Perplexity (enrichment + intent extraction), HubSpot CRM (routing + SLA) Workflow Analysis: Lead Enrichment + Routing That Doesn’t Lie 1) Capture the right raw signal in Webflow Don’t just collect name/email. Add one high-friction field that reveals intent: “What are you trying to replace?” plus a “timeframe” dropdown. You’re buying disambiguation up front. Yes, it reduces volume. Good. 2) Orchestrate in n8n with a strict event contract Trigger on Webflow form submission. Immediately write a “Lead Intake” event to a dedicated n8n workflow log table (even a simple Postgres node works). If the workflow fails, you want proof, not vibes. 3) Enrich with Perplexity, but constrain it like a system Send the company domain + free-text “replace” answer. Ask Perplexity for: company category, estimated size band, likely tech stack, and a one-sentence “why they’re shopping now.” Force JSON output. If it can’t answer, it must return nulls. 4) Create/update HubSpot with deterministic routing n8n upserts the contact and company, attaches the enrichment JSON, then applies routing rules: timeframe <= 30 days goes to SDR queue A; “replacing competitor X” triggers a battlecard task; unknown domain goes to “manual verify” with a 1-hour SLA. What you ship isn’t “automation.” It’s a lead pipeline that can be audited, tuned, and argued with—using data instead of stories.

Maya runs growth at a 25-person B2B SaaS. Mondays are supposed to be “pipeline hygiene.” Instead, she opens Webflow submissions, copies emails into HubSpot, pings an SDR in Slack, and hopes the “replace” field wasn’t just “none.” By lunch, three leads have replied. One is hot. One is a student. One is a competitor scouting. All of them look identical in the CRM because the only structured data is “name” and “company.” So she ships the loop. A form fill hits Webflow at 10:07. n8n triggers instantly and writes a Lead Intake event to Postgres: timestamp, raw payload, workflow version. Not glamorous. But now there’s a receipt. Then n8n calls Perplexity with the domain and the free-text: “Replacing Intercom. Need it live this month.” It comes back as JSON: SaaS, 51–200, likely using Segment, “shopping now because support volume spiked after pricing change.” Nulls where it can’t know. HubSpot gets an upsert. Contact, company, enrichment blob. Routing runs. Timeframe <= 30 days? SDR Queue A. “Replacing Intercom”? Create a battlecard task, attach a snippet for objections. Unknown domain? Manual verify, 1-hour SLA. It mostly works. Then friction hits. Perplexity starts returning “51–200 employees” as a string one day and as an object the next. n8n happily maps it into HubSpot’s number field. HubSpot rejects the update. No one notices because the workflow kept running and only the enrichment step failed. The lead gets routed without the competitor signal. SDR responds with a generic “Thanks for reaching out.” The prospect replies: “Did you even read the form?” Maya checks the log table. Sees the failed node. Fixes the schema with strict validation and a fallback: if JSON doesn’t match, store raw text and route to manual review. But the bigger question lingers. How much “intent” can you really extract from one annoyed sentence typed into a form?

Here’s the uncomfortable part: the loop is only “two minutes to enriched and routed” as long as you pretend language is a stable API. It isn’t. The moment you lean on a model to turn a messy sentence into a field that triggers automation, you’ve built a system whose core input is ambiguity disguised as structure. That doesn’t mean don’t do it. It means treat it like credit scoring, not copywriting. If we’re being honest, the real scaling bottleneck isn’t n8n or HubSpot. It’s exception handling. You’ll spend more time on the 8% of leads that break the contract than the 92% that glide through. Competitor names show up as nicknames, “timeframe” becomes “yesterday,” domains are personal emails, and half your “replace” answers are emotional: “Your competitor’s support is trash.” Great signal, impossible schema. So the system you actually want is less “enrichment” and more “triage with accountability.” We’ve had success treating the model output as a hypothesis that must earn the right to route. A few patterns help: score confidence, store evidence, and gate actions. If confidence on competitor extraction is low, don’t create a battlecard task; create a one-click review task that includes the original sentence, the model guess, and a button that writes the corrected value back to HubSpot. Now you’re not betting pipeline on a guess—you’re turning review into training data. Also: stop trying to predict intent perfectly. Optimize for catching the costly mistakes. A competitor scouting and a student inquiry are the ones that waste cycles and leak info. If your loop only gets great at flagging those two, it pays for itself. Build “negative routing” before you build fancy prioritization. And yes, this means you’re signing up to run a small ops product internally: schema versioning, alerting on reject rates, sampling audits, and a weekly “why did this route?” review. Not glamorous. But it’s the difference between automation and a system you can trust under stress.