Case study · BlueVine · 2020 to 2021
$4.5B in PPP loans, shipped against a rulebook that changed overnight
BlueVine funded more than 155,000 small businesses through the Paycheck Protection Program. I was the lead product manager on the origination platform, from the application a borrower started to the forgiveness flow that closed the loan out. The federal specification kept moving while the platform was live, once landing hours before the application window opened.
Program totals are BlueVine's own published figures. The conversion lift is the one I measured on the PPP application funnel.
The situation
Two weeks of runway and a fixed pot of money
I joined BlueVine on March 1, 2020. Two weeks later the country shut down. The CARES Act was signed on March 27, Treasury published the terms on March 31, and applications opened on April 3.
What existed was a small business term loan and line of credit built for BlueVine's own underwriting. PPP was a different animal. The loan was federally guaranteed, the eligibility rules came from the SBA rather than from us, and the appropriation was first come, first served. Every hour of engineering delay was a borrower who would not get funded before the money ran out.
It was all hands on deck across the company. I was the lead product manager, working directly with the head of engineering, the head of design, the head of operations and the head of servicing, plus our contact at the SBA and the partner banks originating the loans.
Problem one
A government form wearing a website
What I inherited was BlueVine's existing small business loan application mapped onto the SBA's form requirements. Every field the form could eventually need was asked up front, in the order the form listed them, before the borrower had any signal that they qualified or how much they would get.
I rebuilt it as a progressive flow that asks for information only at the point the decision actually needs it. Two changes did most of the work. Questions that could not change an outcome came out entirely. Open-ended questions became yes or no questions bounded by the underwriting limit, because when a threshold decides the answer you do not need the number, you need to know which side of the line the borrower is on. That is faster for a business owner working from a kitchen table and it is a cleaner input for the model behind it.
Inherited
- 1Full business profile, every field the SBA form might need
- 2Open-ended payroll and ownership questions
- 3Document upload
- 4Submit, then wait to learn whether you qualified
One long form, ordered by the paperwork. The borrower spent the most effort before getting any signal back.
Rebuilt
- 1Entity type and the few facts that decide eligibility
- 2Bounded yes or no questions set at the underwriting limits
- 3Loan amount shown, then documents requested for that amount
- 4Submission assembled into the SBA form behind the scenes
Ordered by the decision instead of the document. Onboarding conversion rose 30%.
Problem two
Automate everything but the moving part
At PPP volumes, anything that needed a person was a queue, and a queue was borrowers missing the appropriation. We automated the whole path except eligibility. Eligibility stayed with a human on purpose, because it was the part of the product the federal government rewrote most often, and a rule we had hardcoded on Tuesday could be wrong by Thursday.
That split held up. It also set the design constraint for everything else. Terms, disclosures and the generated loan documents had to be things we could change without a release, because the rate and the rules were inputs from outside the building.
Problem three
Fraud signal without a slower queue
PPP was the largest fraud surface in recent small business lending, and speed was the whole point of the program. Those two pull against each other, so the question was which signals we could act on without adding a step for legitimate borrowers.
A cross partner view
Originating through several bank partners gave us a picture no single lender had. Application stuffing was visible across the set, not inside any one of them.
Clustering on shared details
Multiple companies filing from one address. One phone number attached to a run of unrelated applications. Patterns that are invisible in a single file and obvious across a population.
Known customers first
We prioritized existing BlueVine members whose KYC we already held. The fastest identity check is the one you finished months ago.
BlueVine's fraud rate came in well below the lenders it was compared against. I still think we left signal on the table. Some of what we could see across partners was treated as a flag to log rather than as grounds to decline, because the program did not require a lender to act on it. If I ran that queue again I would set the rule before the volume arrives. A signal we trust enough to record is a signal we act on, whether or not the regulator is asking.
The hard part
The specification changed while the platform was live
Most product teams treat the spec as the thing they control and the schedule as the thing they negotiate. In PPP it was the other way round. The spec arrived from Treasury and the SBA, it arrived late, and the date it applied from was not negotiable. Two examples from the first week set the pattern for the year.
terms published
Apr 2, 2020
terms replaced
Apr 3, 2020
window opens
The interest rate doubled 12 hours before launch
What changedTreasury's initial documentation set the loan rate at 0.5%. Lenders spent the next two days building to it. Late on April 2 the SBA's first interim final rule raised it to 1.0%, because commercial banks had said they could not clear and service the loans at half a point.
What it costThis was not a constant to swap. Amortization schedules, every sample payment term shown to a borrower, and the backend generators that compiled promissory notes and loan disclosures all had to be rebuilt and re-checked.
How long we hadUnder 12 hours. It landed the night before applications opened.
timeline split
Apr 3, 2020
businesses apply
Apr 10, 2020
contractors apply
One funnel, suddenly two start dates
What changedEarly guidance read as though every applicant could file on day one. The April 2 rule split the timeline. Small businesses and sole proprietorships could apply on April 3. Independent contractors and the self-employed had to wait until April 10.
What it costWe had built a single intake funnel. It now had to read the applicant's entity type, let an S corp or a sole proprietorship through, and hold a 1099 contractor back without losing them, which meant classifying entity structure correctly at the top of the flow rather than at underwriting.
How long we hadHours, deployed before Friday morning traffic arrived.
In a regulated launch the specification is an input that moves, not a contract that holds.
The durable lesson was architectural. Rates, eligibility gates and generated documents belong in configuration, owned by the people who read the rules, not in code behind a release cycle. We learned that in the first 48 hours and it paid for itself every time the program was amended after.
Closing the loop
Forgiveness is where the loan actually ends
A PPP loan was only half a product. The borrower still had to get it forgiven, and the SBA published three ways to ask. The full 3508, the shorter 3508EZ, and the 3508S for the smallest loans. Getting the choice wrong meant a rejected filing and a loan that stayed a loan.
The manual path was exactly what it sounds like. A borrower reads three forms, picks one, fills out a PDF, and an operations person reviews it by hand. We replaced all three with a single progressive flow that asks about the borrower's situation and works out which form they qualify for. The business owner never chooses a form number. A compliance workflow that had been staffed turned into product that ran itself.
What carries over
What I would bring to a lending team now
PPP compressed a decade of lending problems into eighteen months. These are the five that still shape how I build.
- 01The full lifecycle, not the application. Origination, identity, decisioning, servicing and the closeout are one product. The forgiveness flow taught me more about the loan than the application did.
- 02A funnel measured on conversion, not completeness. Every question in a lending flow has a cost, and most of them can be bounded by the limit that actually decides the outcome.
- 03Fraud signals are product decisions. What you do with a pattern, and how fast, belongs in the flow design, not only in a risk queue downstream of it.
- 04Terms and rules as configuration. Anything set by a regulator, a partner bank or a pricing committee should be changeable without a deploy.
- 05Partner banks and regulators are stakeholders you design with. I worked our SBA contact and our originating banks the same way I work engineering, which is how rule changes reached us early enough to matter.