Every demo request scored by Claude and given an owner in about 3 seconds.

A lead routing and deal risk system I built on HubSpot, n8n and Claude for a made-up accounts payable software company. Watch it run, then score a test lead.

to score 1,000 leads: $0.44
test leads routed correctly, three runs in a row: 20 of 20
confidence needed to route straight to a rep: 60%
scheduled RevOps jobs on the same portal: 4

Play the 1:20 walkthrough (sound on)

A real run: n8n scores the lead, then the record updates in HubSpot. The walkthrough has captions and an AI-generated voiceover.

Try it

Pick a test lead

Tier A is the best fit and goes to a sales rep first. Tier D is nurtured or disqualified.

Or score your own lead

Fill it in the way a buyer would. Use a real company or make one up.

Same scoring code as the live form, minus the website lookup. Lead details go to Claude. I keep the company, email domain, job title, result and any ref tag from your link. Nothing is written to HubSpot.

What else it runs

The same contacts and deals also go through these jobs.

Job	Runs	What it does
Speed to lead	Every 5 minutes	Checks new leads for a first call, email or meeting. A missed deadline puts a task on the sales manager.
Deal risk	Mondays	Rules score every open deal. Claude reads recent calls and notes, can move the score one level and writes the owner a next step.
Forecast snapshot	Mondays, 7:20am	Saves the open pipeline each week to show what was added, won, lost or slipped. Claude writes the CRO brief from figures the code computed.
Portal audit	Mondays, 7:40am	31 read-only checks on data quality, pipeline hygiene and governance. Claude ranks the fixes.
Reporting	Always on	Four HubSpot dashboards: pipeline and forecast, funnel by source, speed to lead, and data quality.
Change control	Every deploy	Fields, pipeline and forms deploy from a spec locked to this portal. Every AI decision is logged with its model and prompt version.

See the command center and health check

See the full system map

Decisions I made

Claude reads the lead. Code assigns it.: The model judges fit. Owner and call deadline come from plain rules, so sales leadership can change routing without touching a prompt.
When unsure, send it to a person.: If Claude fails, or a lead comes back under 60% confidence or contradicts itself, a manager gets it instead of a guessed route.
Logic in n8n, records in HubSpot.: One scoring workflow serves the live form, this page and the test set. HubSpot stays the system of record.

Testing before a change ships

20 test leads with known answers, including spam, a competitor and three prompt injections. A prompt change ships only if it clears all three bars.

Bar	Needed	v5 got
Spam, competitor and injection cases handled	All	All
Right tier, owner and deadline	19 of 20	20 of 20
Leads that pass every check	17 of 20	20 of 20

v5 is live. It scored 20 of 20 on three runs in a row on Oct 17, at about $0.009 per full test.

Contact

Happy to walk you through the portal and the n8n workflows. All the data is synthetic.

davidengoy@gmail.com/LinkedIn