Where Proximity Still Matters

In U.S. bank small-business lending, the hardest test for AI is not a borrower with a clean history and standardized records. It is a start-up, an unusual firm, or a business whose creditworthiness is partly visible only through context. This comparison asks which problems are solved by better information processing and which still require meetings, local relationships, flexible judgment, or a guarantee. The answer helps distinguish a productive hybrid model from a remote process that has simply removed the people who used to fill its information gaps.

The evidence that AI can improve distant lending is significant, but it does not make every borrower equally legible. A bank can use AI to extract figures from documents and find patterns across transactions. It cannot assume that a thin record is equivalent to a rich one, or that a score captures the same value as a trusted explanation. The difference becomes clearest when the bank must decide whether to finance a firm whose future is not well represented by its past.

Start-ups expose the thin-file boundary

Start-ups are a useful boundary case because they often lack the operating history that makes standardized underwriting comfortable. The FDIC report commonly describes a start-up as a business less than two years old, but the survey allowed banks to use their own definitions. That qualification matters: the category is not a single legally fixed population. It is a group of borrowers whose age or limited history can make conventional evidence less complete.

Even so, banks do lend to them. Sixty-six percent of banks say they typically make loans to start-ups. The share is 66% among small banks and 54% among large banks. This difference is not simply a judgment about the willingness of small banks to take risks. The FDIC finds that banks relying more heavily on hard information, emphasizing credit-bureau information, or often requiring personal credit scores are less likely to lend to start-ups. A firm that has not yet accumulated a strong business credit history may be screened out by a process designed around portable records.

The pattern mirrors the broader contrast between relationship and transactional lending. A small bank can use a conversation to learn how an owner understands the market, what experience the management team brings, and why a business plan is plausible despite limited historical revenue. Those judgments are not automatically accurate, and a meeting can reveal weaknesses as readily as strengths. But it gives the bank a way to gather information that a credit file does not contain.

The FDIC numbers show that meetings are associated with start-up lending most clearly among small banks. Sixty-nine percent of small banks that meet applicants lend to start-ups, compared with 45% of small banks that do not. Among large banks, the corresponding figures are 52% and 53%, a small difference. Large banks therefore appear to use a different compensating mechanism when the borrower’s records are thin.

That mechanism is often a government guarantee. Among large banks, 59% of those that make Small Business Administration loans lend to start-ups, compared with 25% of large banks that do not make such loans. Among banks that lend to start-ups, 55% of large banks say that at least half of their start-up loans use a government guarantee, compared with 22% of small banks.

A guarantee and a meeting address uncertainty in different ways. A meeting can add information to the bank’s assessment. An SBA guarantee can reduce the lender’s exposure to a portion of the loss if the loan defaults. The guarantee does not make screening unnecessary: lenders retain exposure and still have a reason to distinguish viable businesses from poor risks. But it changes the economics of a decision when the bank’s information is incomplete.