Proximitty Wants to Run the Boring Middle of Commercial Lending

Walk into the lending department of a mid-sized bank and you will not find the drama that fintech pitch decks promise. You will find people waiting. Waiting on a borrower to send last year's tax return. Waiting on a rent roll that arrived as a photo of a printout. Waiting to key numbers from a debt schedule into a spreadsheet so a credit memo can finally get written. This is where commercial loans actually live - not in the yes-or-no of approval, but in the grinding logistics around it. Proximitty was built for that room.

The company, part of Y Combinator's Winter 2026 batch, describes itself as "the AI operating system for commercial loans." In practice that means software agents that request and chase borrower documents, read them, spread the financials, generate credit memos, watch covenants for breaches, and handle servicing across the life of a loan. The target customers are FDIC-insured banks, credit unions, and non-bank lenders working on C&I, commercial real estate, and SBA loans - the institutions that have the most paperwork and the least appetite for getting compliance wrong.

There is a reason this segment stayed manual while consumer finance got slick apps. A commercial loan is not one clean decision; it is a relationship that has to be tended for years. A borrower promises to keep a certain debt ratio, to deliver quarterly statements, to insure a building. Someone at the bank has to check that those promises hold, loan after loan, quarter after quarter. Multiply that by a portfolio of hundreds or thousands and the work becomes a treadmill. Proximitty's argument is that the treadmill is exactly the kind of bounded, rule-heavy labor that AI agents can now carry.

01The 15-to-2 story

The number Proximitty likes to lead with is a headcount. One fintech using its agents cut a loan servicing team from 15 people down to 2. On its own, that sounds like a story about layoffs. The more interesting part is what happened next: the company redeployed the freed-up staff and closed $20 million in new originations in a single quarter. The point the founders keep returning to is that automation here is not subtraction. It is redirection - taking people off document-chasing and putting them on the work that actually grows a loan book.

02What the agents actually do

Proximitty breaks the work into four connected jobs. First, ingestion: agents request, chase, and pull in borrower documents - financial statements, rent rolls, tax returns, debt schedules - and reconcile discrepancies in real time. The unglamorous flex here is format tolerance. The system is built to read blurry scans and handwritten notes, the kind of paperwork most software politely refuses to touch. Second, credit analysis: automated financial spreading and credit memo generation, using the institution's own business logic, with outputs designed to be audit-ready.

Third is portfolio monitoring and servicing - a unified data layer that tracks covenants, closing requirements, and borrower obligations, escalating breaches and flagging trouble early. Fourth, and most forward-looking, is what Proximitty calls its Servicing Studio: you record a screen or hand the system a standard operating procedure, and it turns that into a production agent that can execute actions across any interface, no engineering ticket required.

03The founders knew the room

You cannot automate commercial lending from the outside. The messy documents and unforgiving rules are exactly why this corner of banking resisted software for so long. Proximitty's two founders spent years inside it. CEO Wye Yew Ho advised banks and fintechs on risk strategy at McKinsey, then led FinCrime and Growth at Taptap Send as it scaled from $75 million to $200 million in ARR. He studied at the London School of Economics, and - a detail that says something about how he thinks - he is the highest-ranked Malaysian in the history of the Southeast Asian Maths Competition.

CTO Zi Zhang comes from the security side. He led security infrastructure at Bloomberg, work that involved protecting more than 300,000 terminals, and was Head of Engineering at ACI.dev, where he built what he describes as the first unified MCP, a project that reached 4,000 GitHub stars in a month. The pairing is deliberate: one founder who has felt the pain of a covenant breach caught too late, one who has shipped systems where a security mistake is not an option.

04Why banks would trust an agent

A bank will not run an AI it cannot audit. That constraint shapes the whole product. Proximitty leans on full audit trails, reproducible decisions, and a policy that customer data is never used to train models. On the compliance ledger it lists GLBA and GDPR alignment and a SOC 2 in its observation period. None of this is the flashy part of AI, but in a regulated industry it is the part that closes deals. The differentiator is less "our model is smarter" and more "our model shows its work."

That posture also explains the integration story. Rather than asking a bank to rip out its core, Proximitty plugs into the stack banks already run - Fiserv, FIS, Jack Henry, Finastra, Temenos, CSI, and Oracle FLEXCUBE, plus lending-specific platforms like nCino, Abrigo, and Moody's. The company frames onboarding as a three-week arc: integrate in week one, test in a sandbox in week two, deploy in week three. For a buyer who has lived through 18-month core migrations, a three-week timeline is itself part of the pitch.

05The business, and who buys it

Proximitty sells software to institutions, not consumers - the classic enterprise motion, with the added friction of a regulated buyer. The people who sign are chief credit officers, heads of lending, operations leaders, and the model-risk and compliance teams who have to bless anything that touches a lending decision. Its published economics point at operating expense: fewer people on servicing, faster underwriting, and analysts freed to originate. The framing is worth stealing for any enterprise founder. Proximitty does not lead with "features." It leads with capacity - the idea that one analyst can suddenly cover the ground of eight. Capacity is a number a banker can defend in a budget meeting.

Early customers are the kind of lenders that feel the squeeze most: community and regional banks, credit unions, and non-bank lenders that want to grow their books but cannot justify hiring proportionally to do it. For them the promise is not glamour. It is being able to say yes to more loans without adding a floor of analysts, and catching the loans that go quiet before they go bad.

06Where it sits in the market

Commercial lending software is not empty. Incumbents like nCino, Abrigo, Moody's, and Numerated already sell into banks, and most institutions still lean on in-house analyst teams and a patchwork of point tools. Proximitty's wager is that those options solve slices of the problem, while the day-to-day reality is end-to-end: a document arrives, a discrepancy surfaces, a memo is due, a covenant slips. By stringing the whole lifecycle together with agents that run on a bank's own logic, the company is trying to own the workflow rather than a step of it.

It fits a broader pattern in the YC W26 class - the return of the back office. The most durable enterprise AI companies are not chasing consumer attention. They are chasing the operations team's time, in industries where the paperwork is heavy and the tolerance for error is low. Commercial lending is close to a pure example: high stakes, high volume, and a workflow that has always been more manual than anyone would like.

The company is early. It is a small San Francisco team, backed by Y Combinator, CRV, Cohen Circle, and Uphonest Capital, with angels who lead teams at Stripe and OpenAI, and it presented at FinovateSpring 2026. The metrics it publishes are its own, drawn from early deployments, and the SOC 2 is still in observation. What Proximitty has, that is harder to manufacture than a funding round, is a pair of founders who know exactly which room they are automating - and why the people in it have been waiting.