Work / Roofing lead intake (demo build)
AI lead intake that files roofing leads in HubSpot and Asana
Web forms, emails and voicemails read by AI, sorted by urgency and filed in HubSpot with a drafted reply and an Asana task. Tested on 40 realistic leads.
- 40 of 40
- test messages sorted into the right category
- 8 of 8
- emergencies flagged as emergencies
- 2 of 2
- repeat callers matched to their open deal
The problem
A roofing company’s leads arrive in three places: the website form, the inbox and the voicemail. Many are incomplete. Some are spam or vendor pitches. A few are emergencies, such as water coming through a ceiling, and the company that calls back first usually gets the job. Someone has to read everything, type it into the CRM and decide what comes first.
What I built
A small service on Cloudflare that takes each new lead and does that first pass.
- Reads the message. A Claude model pulls out the category, name, phone, email, city, job type, urgency and insurance status, and writes a short summary and a draft reply.
- Drops the noise. Spam, vendor pitches and job seekers never reach the CRM.
- Files the lead in HubSpot. It finds or creates the contact, opens a deal in the first pipeline stage and attaches a note with the original message and the draft reply.
- Catches repeat callers. A second message from someone with an open deal attaches to that deal, and an emergency follow-up escalates it.
- Creates the task. An Asana task lands in Emergency dispatch, Estimate visits or Customer service, due the same day for emergencies.
Draft replies wait for a person to approve them. Nothing is sent to customers automatically.
How it was tested
There was no real roofer’s data to start from, so I wrote 40 realistic messages for a fictional company in Parker, Colorado, with an answer key. The set includes spam, a Spanish-language voicemail, repeat callers, a warranty claim, a complaint, an out-of-area job and leads missing a name or phone number. Every person and address in it is invented.
A scoring script sends all 40 to the live service and compares the results with the answer key, field by field. On the final live run:
| Field | Correct |
|---|---|
| Category | 40 of 40 |
| Name, phone, email | 34 of 34 each |
| Job type | 34 of 34 |
| City | 33 of 34 |
| Insurance claim | 20 of 21 |
| Urgency | 26 of 33 |
Every urgency miss was one level off, such as a hail claim with no deadline rated normal instead of high. The run created 29 deals and 32 tasks.
What it shows
Most of the value came from wiring together tools the company would already pay for. The AI model does the one step plain automation cannot: reading a messy message and deciding what it is. The test set made the build measurable, and the first live run surfaced two problems before any customer could have seen them: a missing HubSpot permission and a missing field in the Asana request.
The next step is a pilot with a real home service company. See AI automation for how this applies to your business.