Turning Loan Enquiries into Underwriting-Ready Applications

PixieDust is an AI-native operating system for SMB lenders and merchant service providers, founded by experts from PayPal, Funding Circle, JPMorgan, FICO and Dun & Bradstreet — people who ran the lending infrastructure they are now rebuilding. Their customers are lenders and merchant service providers serving small businesses with working capital, revenue-based finance and SBA-aligned products. PixieDust is trying to help them turn more qualified SMB leads into funded loans, without proportionally increasing their sales and operations headcount.

• Increasing borrower drop-off

  Borrowers arrive at a lender's site with questions that decide whether they will apply: what they qualify for, at what rate, over what term, with what documentation.

• Challenges in efficient scaling

  Support teams cannot handle the volume of borrower questions through live chat, while traditional bots fail on complex product and rate-related queries.

• Compliance & revenue risk

  Inaccurate or inconsistent answers can create disclosure/compliance issues while also causing lenders to lose qualified borrowers and potential revenue.

PixieDust conceptualised, and Ryewalk executed a fully AI-enabled solution, Business Loan Assistant, whereby the SMB Lending process can be accelerated for the end-customers and the lenders, but with increased compliance scrutiny and controls. The envisaged solution was to fully AI-agentify the customer application engagement, from initial conversation-based application enrichment to aiding qualifying the application for success. The solution deployed an inbound AI agent on each lender's site. Instead of dropping borrowers into a static form, it answers product, rate and eligibility questions from that lender's own approved knowledge base, qualifies the borrower through natural conversation, and then pre-fills the application from everything the conversation already established — so the borrower never re-enters what they've just told us. With every answer traceable to its source and nothing submitted without the borrower confirming it. Because PixieDust is an operating system rather than a single-tenant product, the assistant runs across five brands, each with distinct products, credit box, personality and approved material, from one platform rather than five builds.

Ryewalk built and operates the productised AI-assistant - entirely through Claude. With over 10+ Claude certified architects, Ryewalk was able to seamlessly transform the client's business vision & idea into an implementable, tangible solution - within timelines & budget. As part of the solution, Claude does the reasoning, with each turn separating structured understanding from response generation. The facts a conversation establishes, such as revenue, time in business, amount sought and purpose, are captured under a strict schema rather than parsed out of prose. That makes the conversation auditable. Model selection is tiered by task to keep per-conversation cost predictable.

Three Claude capabilities carry the design as follows:

1. 01

   Tool use makes fact extraction a schema contract rather than a parsing exercise.

2. 02

   Files API grounding with citations lets the agent read a lender's approved knowledge base in full and cite it, rather than assembling answers from retrieved fragments — (removing the retrieval-miss failure mode from a rate conversation)

3. 03

   Prompt caching keeps the system economical at volume.

Containment is architectural rather than a matter of instructing the model well:

• The model cannot act

  The response call carries no tools at all. Claude's only outputs are language for the borrower and the structured facts that pre-fill an application, which the borrower reviews before anything reaches the lender.

• Every change is traceable

  Each fact is stored with its provenance, so any borrower's conversation history can be reconstructed for review.

• Compliance owns the boundaries

  Lender behaviour is configuration rather than code, so the material the compliance team approves is the material the agent is held to. Adding a new lender brand is a configuration exercise, not a build.