What Banks Actually Automate—and What They Keep Human
In U.S. bank small-business lending, “technology adoption” is too broad a description to answer whether AI can replace proximity. This chapter maps where financial technology enters the lending workflow, distinguishes process support from AI underwriting, and shows why a bank can become more digital without becoming fully remote or relationship-free.
The distinction matters because a loan is not a single action. It is a chain that begins when a business looks for financing, passes through application and underwriting, and continues through closing, servicing, monitoring, and portfolio management. A tool that makes compliance faster changes a different constraint from a model that evaluates an applicant. Treating both as evidence that AI has replaced local knowledge would confuse the technology with the job it performs.
FinTech touches a workflow, not one magic decision
The FDIC’s Small Business Lending Survey asked banks about financial technology at ten stages of their small-business loan process. The survey counted technology whether a bank developed it internally or obtained it from an outside firm. It also treated a bank as a “FinTech bank” if it used technology in at least one stage. That is a useful measurement choice for describing modernization, but it deliberately includes far more than artificial intelligence.
The distribution of use shows why the distinction is necessary. In 2022, 31% of banks were using FinTech somewhere in the process, and another 22% were discussing or developing it. But the most common use was regulatory compliance, reported by 21% of banks. Banks also used it for closing and loan performance tracking or servicing, each at 12%, and for portfolio analytics at 10%. These are important activities, but they are generally downstream of the question “Should this borrower receive this loan?”
Use was less common at the front end. Nine percent of banks used FinTech for application submission, 8% for application processing, 8% for underwriting and decision-making, and 8% for post-decision quality control such as anti-fraud work. Only 4% used it to identify potential borrowers, and 5% used it for loan funding. The figures overlap because a bank could select several stages. They should therefore be read as a map of functions, not as a sequence in which each percentage represents a separate set of banks.
The breadth of adoption was limited as well. Sixty-nine percent of banks used FinTech in no stage of the small-business loan process. Fifteen percent used it in one or two stages, 8% in three or four, and 8% in five or more. Even among adopters, using a tool to process a class of financial statements is not the same as letting a model make the final decision. The FDIC report explicitly cautions that a stage count cannot show how deeply a technology is integrated. It cannot distinguish between automating one narrow task and changing the bank’s entire underwriting model.
This is the first boundary around the word “digital.” A bank may use software to check regulatory documents, organize a borrower’s records, close a loan, and monitor a portfolio while preserving a human review of creditworthiness. In that configuration, technology reduces administrative cost or improves visibility after approval. It may make the relationship more efficient, but it has not necessarily replaced the relationship as the source of judgment.
The same logic applies to communication. Almost all banks permit customers to perform application steps at a branch, and remote communication is widely available for consultation, submission, and supporting documents. Online portals are less common: 25% of banks accept a formal application online, 23% accept supporting information, 19% offer online product consultation, and 13% allow signing online. Forty percent allow some application steps to be completed online, but only 6% allow all four surveyed steps to be completed online.
That last figure is more revealing than the existence of an online portal. A borrower can upload a form without completing a credit relationship online. The bank may still need to ask questions, verify the records, negotiate terms, explain a decision, or monitor the loan. Digital access changes the route by which evidence reaches the bank. It does not establish that the evidence is sufficient, that the borrower trusts the institution, or that the bank has learned enough about the business to take risk responsibly.
Scale changes the shape of adoption
Large banks are more likely to use FinTech, but the difference is not as simple as large banks being digital and small banks being traditional. Forty-two percent of large banks used FinTech in at least one lending step, compared with 30% of small banks. Another 31% of large banks were discussing or developing it, compared with 21% of small banks. Thus 73% of large banks and 51% of small banks were at least considering some FinTech use.
The breadth of use differs more sharply than the presence of use. Among banks that use technology, large banks are much more likely to apply it across five or more steps, while small banks are more likely to use it in only one or two. This is consistent with a scale effect: the fixed cost of integrating a system, training staff, and maintaining a data infrastructure is easier to justify when a bank processes a high volume of similar activity.
But scale does not dictate the purpose of the tool. When small banks use FinTech, they tend to use it in secondary processes such as compliance and performance tracking rather than at the earliest stages of borrower evaluation. Large FinTech banks are more likely to use it for application submission, processing, and underwriting. Both groups use technology at similar rates for some post-approval activities, including compliance, servicing, and portfolio analytics. The implication is not that large banks have discovered a universal replacement for relationships. It is that their operating scale makes deeper integration more economically feasible.
The underwriting evidence makes the contrast clearer. The FDIC report defines automated underwriting as using a credit-scoring model to produce a lending decision from numeric inputs without human discretion. Only 9% of small banks use a credit-scoring model for some small-business lending, compared with 46% of large banks. Even among large banks, automatic approval is concentrated in small loans: 28% can automatically approve a loan of $25,000, compared with 11% for a $250,000 loan and 1% for a $1 million loan.
This pattern gives the word “automation” a practical boundary. It is most feasible where the bank can define the applicant, the data, and the acceptable risk range in advance. As loan size, complexity, or exception risk increases, the cost of a bad decision rises and more information may be needed. A model may narrow the routine cases that require staff attention without making human review unnecessary for unusual cases.
There is also a difference between a scoring model and artificial intelligence. A credit score may be one input in a highly standardized process, while the Census study defines AI through technologies such as machine learning, machine vision, natural language processing, and voice recognition. Some AI can help create or interpret hard information; some FinTech may do neither. The categories overlap in the marketplace, but they answer different research questions.
A bank’s customer-facing style does not follow automatically from its underwriting style. The FDIC finds that FinTech banks are less likely than non-FinTech banks to have a decision-maker meet applicants—78% compared with 92%—and that this difference remains after controlling for bank size. That suggests a relationship between technology use and less reliance on meetings in underwriting. Yet most FinTech banks still meet applicants, and the report finds that technology-using banks are generally as likely as other banks to engage in many high-touch practices, including building relationships, accepting walk-ins, taking referrals, and participating in the community.
This coexistence is not a statistical curiosity. It reflects two different jobs. A bank may use technology to lower the cost of gathering or checking information while asking staff to preserve the trust and continuity that keep a borrower engaged. A mobile app can make a deposit or document transfer easier, while a banker still helps the owner choose a product. A servicing system can flag a change in cash flow, while a relationship manager decides whether the change calls for a conversation rather than an immediate response. The tool and the relationship occupy different interfaces in the institution.
The bank’s incentives also help explain why adoption is gradual. Seventy-two percent of banks classify cybersecurity risk as a large concern when considering FinTech, where “large” combines large and very large concern responses. Sixty-nine percent say the cost of implementation—including integration with legacy systems and investment in human capital—is a large concern. Data sharing with third parties is a large concern for 51%, reputational risk for 46%, and compliance or regulatory risk for 45%.
These concerns have transmission paths. A security failure can expose borrower information or interrupt a critical process. Integration costs can consume the expected savings before the tool works reliably with existing systems. A bank that sends data to a third party may lose control over how information is handled or explained. A model that produces an opaque recommendation can create reputational and compliance problems even when its average prediction is useful. The technology is not an isolated purchase; it changes dependencies across the institution.
Banks also report softer barriers. Thirty-nine percent see a lack of perceived benefit to their current business model as a large concern, and 29% say the same about a lack of awareness of relevant technology. A bank may understand that an online lender has a convenient interface yet conclude that the available tools do not fit its borrowers, product mix, or way of underwriting. The absence of adoption is therefore not proof that a bank is incapable of modernizing. It may reflect cost, risk, limited relevance, or a judgment that the proposed technology does not improve the constraint that matters most.
The human boundary is a design choice with consequences
The FDIC data suggest that technology is entering bank lending in layers. Administrative and post-approval tasks are relatively accessible because they can be standardized without asking the system to understand the entire borrower. Customer communication can move partly online, but full online origination remains uncommon. Underwriting can be scored or partially automated, particularly for smaller loans, while complex or unusual cases continue to require judgment.
That layered pattern is compatible with relationship lending. It can even strengthen it if staff spend less time transcribing documents and more time understanding a business. But complementarity is not guaranteed. If a bank treats a model’s output as a substitute for asking questions, the loss of soft information may matter. If it uses an online channel to reach a borrower but cannot explain a decision or detect a bad input, the apparent reach may not produce durable access to credit.
The choice also differs by bank size. A large bank may use technology to make a standardized small loan faster and cheaper, but it may still reject or escalate a case that falls outside the model’s boundaries. A small bank may use fewer integrated systems while relying on staff who can assemble a fuller picture through a meeting. Neither configuration is automatically superior. The relevant test is whether the bank’s way of collecting, processing, and acting on information fits the risk and needs of the borrower.
This is why broad FinTech statistics cannot answer the proximity question on their own. They tell us that banks are adopting tools, where those tools are concentrated, and what prevents wider use. They do not tell us whether AI has reduced the cost of replacing local knowledge with portable evidence. That narrower question requires looking at what AI processes, how distance is measured, and whether outcomes improve for borrowers who are harder for a nearby relationship lender to know.
The next step is to isolate that mechanism. AI matters for proximity only if it can do more than move a document across a network. It must make messy information usable, integrate evidence that would otherwise remain scattered, or identify patterns that conventional review misses. Only then could it change the informational reason a bank has historically needed to stay close.