How I Used Claude to Build an AI Loan-Intake System Inside HubSpot

I directed and approved. Claude did the build. Here's exactly how a commercial lender went from hand-sorting broker emails to an AI intake system running inside HubSpot.

The problem

A commercial lender's loan officers receive more than 1,000 inbound emails a month. Buried in them are broker loan requests, and each one has to be:

• spotted in a crowded inbox
• logged in HubSpot as a Lead
• checked against the lender's criteria
• answered, often with a decline

All of it was done by hand.

Step 1: Start with the data, not a guess

Before building anything, Claude analyzed the HubSpot portal: email volume per inbox across four loan officers, the Lead pipeline with its stages and existing properties, and the owner IDs and lending criteria. The design started from real numbers.

Step 2: Two custom agents

Claude wrote and configured two HubSpot custom agents:

1. Submission Check. Is this email a loan request? Is it a duplicate? If it's new, create the Lead, name it by property and city, and assign it to the officer who received it.
2. Loan Intake Agent. Extract the loan details and score them against the lending criteria: Meets, Needs review, or Does not meet.

Step 3: Test before going live

Claude ran free simulated tests on 10 real broker contacts. Nine of ten were correct on the first pass. The miss was an edge case that matched a new request to the wrong existing deal, so Claude rewrote the rules: a new property from the same broker counts as a new deal. Every fix was re-tested.

Step 4: Catch what breaks

Testing surfaced problems that would have been costly in production:

• CC'd contacts were triggering the agent
• auto-replies and calendar accepts slipped through
• forwarded deals still needed a way in

The fix: Claude rebuilt the trigger as an email-based workflow with tight sender, recipient and subject filters.

Step 5: The workflows

Claude built two workflows, step by step in the browser:

• Workflow 1: incoming broker email → Submission Check
• Workflow 2: new Lead → Loan Intake Agent → branch by result

Each result routes differently. Meets creates a follow-up task. Needs review creates a review task. Does not meet marks the Lead Disqualified and creates a task to review and send the decline. A human sends every reply.

Step 6: Cost controls

HubSpot AI agents consume credits, so Claude measured credits per run type during testing, projected the monthly cost from real email volume, set per-agent run limits and a hard monthly credit cap, and published the agents without running them until go-live. No surprise bills.

Step 7: A pilot dashboard

Claude built five HubSpot reports to judge the pilot:

• Leads created
• Agent results (Meets / Needs review / Does not meet)
• Disqualifications by reason
• Tasks by owner and type
• Declines waiting to be sent

It then seeded a test record to prove every report worked, flagged for deletion before launch.

The takeaway

Claude didn't replace the consultant. It replaced the clicking. I made the calls on scope, criteria, approvals and go-live. Claude did the research, build, testing and documentation. The project went from idea to launch-ready in days, not weeks.

Want an AI intake system like this in your HubSpot portal? Book a consultation with CRM Assist.