AI Is Becoming More Accessible to Smaller Financial Institutions. Who Is Paying for the Safeguards? 

A community lender usually knows its borrowers well, but what it often lacks is the staff to process their applications quickly, and AI software can help with that. A bank or credit union with a small technology team can now buy it from a specialist instead of building it from scratch. 

And there is big demand for this, with an MSME finance gap at $5.7 trillion, with 70% of micro, small, and medium enterprises in emerging markets lacking adequate financing. 

But a vendor can supply a model. It can’t take over the duty to explain what the model decides. When software influences a declined loan or a policy that isn’t renewed, the institution using it still has to tell the customer why, and answer to the regulator if the decision was misguided or even wrong. A large financial institution will have specialists dedicated to reviewing models and push back on vendors. A community lender, though, might have one compliance officer who also covers all kinds of other tasks. 

Buying AI gets an institution the capability and leaves the accountability with it, too. And because the work of oversight doesn’t shrink with the size of the firm, whether AI widens access at a small lender depends on whether it can afford that work. 

Smaller institutions are buying. AI use at small banks, for instance, rose from 22% to 52% in two years, and at small insurers from 27% to 46%, as per Deloitte EMEA research. 

Jifiti’s recent partnership with Peach Finance shows what that looks like in lending. Announced in May, it gives U.S. community and regional banks one product that runs from loan origination through servicing. A bank with a small technology team can launch a lending program it couldn’t have built on its own. Importantly, the partnership helps banks keep “full control over the customer experience, compliance, and lifecycle operations.” Control over compliance also means responsibility for it. 

Buyers don’t always see inside what they’ve bought. Many vendors of off-the-shelf software have started embedding AI in their products, often without their customers having full visibility or understanding of it. 

Decisions still have an owner

Insurance shows what’s at stake. Insurers have long priced by segment, such as drivers of a certain age in certain regions, because the cost of a single policy only shows up in claims, sometimes years after the sale is signed. Soteris, a Y Combinator-backed company that sells machine-learning software to property and casualty insurers, scores each policy on its own and flags the ones likely to lose money. It says customers of its original product improved loss ratios by five to 15 points within a year.

There’s an open question, though: what happens to the customer behind a flagged policy? The insurer might reprice it or decline to renew it. Whether the savings reach other policyholders as lower premiums is a separate matter, and the answer to that depends on how the insurer uses the information. A vendor can supply a risk score, but the insurer has to explain the decision in terms the customer understands. 

Lending works the same way. In the U.S., a lender that denies consumer credit must give the applicant specific reasons why, so a bank that licenses a scoring model needs the model to produce them. The bank signs the notice, not the vendor. 

Before signing, a lender should ask the vendor to walk through a real or representative declined application. What can the lender see about the outcome, and what records would exist if a customer challenged it or a regulator asked?  A demonstration tells you more than an assurance that the system is transparent and working correctly. 

Liability has a steep price 

Insurance may not step in either. Some standard policies could carry “absolute AI exclusion” clauses, so cover for a model’s mistakes can’t be assumed. Founder Shield, a brokerage that arranges insurance for startups and small businesses, is working on the gap from the other side. It builds scenarios for AI malfunctions and hallucinations into its professional services policies, so the risk is addressed explicitly in the wording instead of left to silence (and assumptions). That gives a buyer something concrete to hold up against what its own policy says. 

A bank or insurer that deploys a vendor’s model should ask its broker, in writing, whether a claim arising from that model would be covered. The answer may turn on the type of loss and on any exclusions, and it’s better to learn that at renewal than after a claim.

The fixed-cost problem 

Oversight doesn’t peter out. Someone has to review the vendor’s contract and keep records of how decisions are made, and those jobs don’t get easier or lighter because a lender has fewer customers. A large bank can staff a team for them, but a community lender often can’t. 

The evidence supports the shape of this concern but doesn’t yet size it for AI. In Deloitte’s survey, more than half of the respondents named transparency and explainability as a hurdle, which is linked to the growing use of vendor tools, and 39% cited internal skills as an additional challenge. 

More accessible software from vendors puts tools within reach of the smaller players that couldn’t afford them a few years ago. Whether that widens access depends on who picks up part of the oversight bill. Vendors can supply decision records and documentation as standard, and insurers can write cover that rewards sensible controls instead of excluding AI across the board. Until someone other than the smallest institution carries part of that cost, cheaper software alone won’t widen access. But more proactive vendors, and small institutions looking for the right traits, can help.