AI bank lending may be changing more than how quickly a credit decision is made. Researchers at the Federal Reserve Bank of San Francisco found that banks with higher levels of AI-related hiring devoted a smaller share of their lending to small and medium-sized businesses than banks with less AI activity.
Across the full sample, the high-AI group had roughly 12 per cent of lending in the small-business category compared with about 21 per cent for the low-AI group. The gap appeared across different bank sizes, although the levels varied substantially between small, medium and large institutions.
The researchers are explicit that this is an association, not proof that AI caused banks to retreat from small-business lending. The result is still worth attention because smaller firms often depend on information that does not fit neatly into a standard data field.
How the AI bank lending comparison was built
The San Francisco Fed team needed a way to estimate how intensively each bank was adopting artificial intelligence. Rather than relying on a survey, it used the share of a bank’s job advertisements that asked for AI-related skills. Data from Lightcast provided the hiring information, which the researchers combined with regulatory lending data.
That measure is useful because it captures visible investment in AI capability. It is also imperfect. A bank can use AI without advertising many AI jobs, while a bank recruiting aggressively may still be years away from embedding those systems deeply in credit decisions. The measure should therefore be read as an indicator of adoption intensity, not a direct switch marked AI on or off.
The underlying Federal Reserve Economic Letter is also catalogued through Fed in Print. Both sources frame the findings cautiously and avoid treating the observed relationship as a causal experiment.
Why small firms depend on information that is hard to automate
A large listed company leaves a long trail of audited accounts, market data, analyst coverage and formal disclosures. A small business may be understood through a much more local mixture: the owner’s history, the quality of management, a relationship with the bank, seasonal trading, a key contract or knowledge of the local market.
Economists often describe that distinction as hard and soft information. Hard information is structured, comparable and easier to feed into a model. Soft information depends more on context and judgement. Relationship banking has traditionally helped smaller borrowers because a human lender can gradually accumulate that context.
That does not mean AI is incapable of using rich information. It does mean that an organisation optimising around scalable, standardised data may naturally become better at serving borrowers who already produce plenty of it.
The raw lending gaps are large enough to notice
The paper reports that high-AI small banks devoted about 13.3 per cent of their lending to small businesses, compared with 21.4 per cent among low-AI small banks. For medium banks the comparison was about 3.8 per cent against 7.8 per cent. Among large banks it was roughly 2.2 per cent against 3.3 per cent.
Those are descriptive comparisons, not a controlled before-and-after test. Banks that invest heavily in AI may also pursue different customers, operate in different markets or have different growth strategies. A technology-intensive bank could have been moving towards larger, more data-rich borrowers even without AI.
That is why the researchers avoid saying that an algorithm refused loans that a human would have approved. The paper instead raises a broader possibility: greater reliance on scalable analytical systems may coincide with a business model that gives relatively less weight to relationship-based SME lending.
Efficiency can improve while access changes
The study also finds that higher AI adoption is associated with stronger return on assets, alongside higher levels of problem loans and a lower small-business lending share. That combination is a reminder that a more efficient bank and a more inclusive credit market are not automatically the same thing.
An AI system can help a lender process documents, detect anomalies, price risk, triage applications and monitor portfolios. Those gains can reduce costs. Whether the savings reach a small borrower depends on what the institution chooses to optimise and what information its models can use confidently.
LiveAIWire has previously examined how AI can expand the information used in credit decisions. More information can improve prediction, but it also changes which behaviours become visible and which borrowers fit the model most comfortably.
Small banks face a different AI problem
Community and regional banks often compete partly through local knowledge. Their advantage may come from understanding customers that a national scoring system finds awkward. At the same time, they have fewer resources to build AI teams, modernise data systems and satisfy the governance requirements that come with automated decision tools.
That creates a strategic dilemma. If a small bank adopts expensive technology, it has to use it efficiently enough to justify the investment. If it does not adopt, it risks becoming slower and more expensive than larger competitors. Either route can put pressure on the human relationship model that historically supported some small businesses.
The answer is not necessarily to keep lending manual. A more plausible goal is to use automation for the parts of underwriting that genuinely benefit from scale while preserving an escalation path for cases where local knowledge or unusual circumstances matter.
What a small-business borrower may notice
For a business owner, the shift could appear in mundane ways: more requests for structured data, less tolerance for incomplete bookkeeping, faster automated screening and fewer opportunities to explain an unusual number in person. None of those changes is inherently unfair. They simply reward businesses whose financial information can be interpreted quickly.
That increases the value of accurate accounts, reliable cash-flow records and clear documentation. It can also widen the gap between firms that operate through modern digital systems and firms whose financial story sits mainly in the owner’s head.
LiveAIWire’s earlier look at AI in investment management reached a related conclusion: better analytical tools do not remove the need to understand how the data was produced and what the model cannot see.
AI can change the market without making the final decision
It is tempting to imagine AI lending as a robot approving or rejecting an application. In practice, technology can reshape credit long before it reaches the final yes or no. It can influence which customers are targeted, which applications are prioritised, how much information is requested, what counts as an exception and how much human time a borderline case receives.
That makes the San Francisco Fed result useful even without a causal verdict. It points towards a structural question for regulators and banks: as lending becomes more data driven, are small firms losing access because their risk is worse, or because the information that explains them is harder to standardise?
The study cannot answer that fully. It does show that AI adoption and the composition of bank lending are moving together strongly enough to deserve continued scrutiny.
The warning is about incentives, not machines
AI may ultimately help banks lend to more small firms by lowering underwriting costs and finding patterns humans miss. The same technology could also encourage institutions to concentrate on borrowers that are easiest to assess at scale. Both futures are technically possible.
The difference will come from incentives, data design and whether humans remain able to review cases that do not fit the standard pattern. For borrowers, that makes transparency around automated credit processes increasingly important.
LiveAIWire has also explored the limits of AI in personal financial planning. The shared lesson is straightforward: financial automation can be genuinely useful, but convenience should not be confused with neutrality. The system still reflects what the institution chooses to measure and reward.
The strongest conclusion is a question for the next study
The Fed researchers are careful not to turn correlation into a verdict. The next useful step would be to observe more directly where AI enters the lending process and whether the same bank changes its small-business behaviour as adoption deepens. That could help separate technology effects from differences in strategy between banks.
Until then, the current evidence is best treated as a warning signal. AI-heavy banks look different in ways that matter to small firms, and those differences are large enough to investigate rather than explain away.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.