What regulators and financial institutions have proven about AI
Prateek Shrivastava | July 2026 | Washington DC
Artificial intelligence is already producing real, measurable results in financial services, and the strongest evidence for it is more substantial than the conference circuit usually gives it credit for. The appendix of this article contains the database of case studies referenced in the text.
The most credible evidence sits in two places: regulators using AI to improve supervisory work, and lenders and insurers using it to assess risk, extend credit, design cover and process claims. This piece brings together the strongest examples from both groups.
In reviewing the evidence, we found it useful to sort every case into one of two buckets. Measured deployments are live or substantially operational systems with published adoption, accuracy, service or portfolio metrics, the clearest proof that AI is already working. Measured pilots are real implementations with credible technical outputs, often the leading edge of where a measured deployment is headed next. Broader landscape reports, from the World Bank, the Bank for International Settlements (BIS), the Organisation for Economic Co-operation and Development (OECD) and others, matter for understanding where the market is moving, and sit alongside this exercise rather than inside it.
That distinction matters: The measured deployments below are proof that AI already works in this sector. The measured pilots are a good sign of what is about to work more widely.
Where regulators are getting practical value
The research shows that, at the moment, the strongest regulator-side examples are not futuristic “AI supervisors.” They are focused applications that improve the quality and speed of work regulators already do: handling complaints, monitoring markets, catching fraud and checking data.
Example 1. Brazil: finding anomalies in a very large credit registry, and being precise about how few cases that covers
Banco Central do Brasil has tested machine-learning techniques to identify potential anomalies in submissions to its public credit registry, the SCR. Working with the International Finance Corporation (IFC), it applied an Isolation Forest model to paycheck-linked lending data drawn from a 55-million-loan training set, against a registry that processes 1.1 billion operations a month (IFC case study, May 2025). The model flagged 278 loans as potential outliers across nine institutions. Of the 206 cases where the central bank received feedback from eight of those institutions, 179 were confirmed as genuine issues, an 87 percent precision rate. After segmenting high-income borrowers from the rest of the portfolio and retraining, precision rose to 97 percent.
The validated sample is 206 cases, not the full registry. This is a strong suptech example because the problem is specific, the model supports an existing control process and performance can be measured. It should be described as a measured pilot unless there is public evidence that the approach has been deployed broadly across the registry and has produced sustained improvements over time.
Example 2. The Philippines: making complaints handling accessible
The Bangko Sentral ng Pilipinas’ Online Buddy, or BOB, is one of the clearest emerging-market examples of AI-enabled consumer protection. It gives consumers a digital channel to submit complaints and uses natural-language processing to categorise and route cases.
An Innovations for Poverty Action evaluation recently analysed 1,372,534 messages sent to BOB over its first year of operation, from August 2020 to July 2021 (IPA, 2022). BOB correctly categorized 55 percent of complaints immediately, rising to 74 percent after follow-up questions. Only 20 percent of chatbot interactions reached the final message in the conversation flow. That drop-off rate is as informative as the accuracy figure: it tells us where the design still loses people, which is exactly the kind of workflow metric a regulator should be tracking and publishing on an ongoing basis. The evaluation itself is now several years old and, as far as we can find, has not been refreshed publicly since. That combination of regulator reporting and independent evaluation still makes BOB unusually useful: it provides evidence on adoption, system performance and where users disengage.
It also demonstrates an important design principle. The value was not created by inventing an entirely new supervisory function. It came from improving access to an existing complaints process and converting unstructured messages into usable supervisory data.
Example 3. India: detecting “mule accounts” before fraudulent funds disappear
The most current addition to the suptech list addresses a problem the other cases here do not directly: identifying “mule accounts”, which are used to receive and move the proceeds of cyber-enabled financial fraud. The Reserve Bank of India’s (RBI) innovation arm, the Reserve Bank Innovation Hub, built MuleHunter, a machine-learning model that screens transaction and account data against known behavioural patterns to identify these mule accounts. It was piloted with two public-sector banks in 2024 and referenced by the RBI governor in the December 2024 monetary policy statement. A right-to-information response reported in December 2025 indicated that 23 banks had adopted it. In May 2026, the Department of Financial Services publicly directed banks to accelerate adoption.
That adoption figure is worth noting for a reason beyond the number itself. It was not disclosed by the RBI or the vendor. It came out through an independent information request, a stronger form of provenance than most of the figures in this space. Please note: we have not found a verified accuracy percentage from a primary RBI source, and any figure circulating in secondary coverage should be treated as unconfirmed until the central bank publishes one directly.
Peru, Ghana and Indonesia: promising work still building toward a track record
Peru’s Superintendencia de Banca, Seguros y AFP, the Bank of Ghana and Indonesia’s Otoritas Jasa Keuangan (OJK) have all built AI-enabled market monitoring and complaints tools, using social-media scraping, sentiment analysis, chatbots, complaint case management and supervisory dashboards.
These initiatives are credible and worth watching. Some have already processed substantial volumes of public comments and produced usable monitoring interfaces. What they do not yet have publicly is the same depth of outcome evidence as Brazil, the Philippines or India, lasting changes in consumer outcomes, enforcement effectiveness or supervisory productivity measured over time. That is less a verdict on the technology than a sign of where each program sits on its own curve, earlier than Brazil, the Philippines and India, but on the same path. They belong in the database as measured pilots, the group most likely to produce the next mature deployment.
3 lessons from the regulator cases stand out
First, AI works best when it is attached to an existing supervisory process with a clear operational bottleneck. Complaints arrive as unstructured text. Registry submissions contain anomalies. Transaction networks conceal mule accounts. Public conversations reveal emerging conduct risks. These are well-defined problems.
Second, workflow metrics matter, and the honest ones include the uncomfortable numbers. A regulator should be able to measure classification accuracy, false positives, completion and drop-off rates, cases routed, and whether supervisory action improved, not just the headline volumes.
Third, independently sourced figures are worth more than self-reported ones. The RBI adoption number came from a right-to-information request. The BOB evaluation came from an independent research organisation, not the regulator’s own communications team. That distinction should shape how much weight any of us puts on a claim.
Where lenders and insurers have stronger evidence
On the financial-service-provider (FSP) side, lenders, card issuers, agri-finance firms and insurers, the evidence base is broader, particularly in credit underwriting and product design. But this is also where attribution becomes slippery. A lender may have a large portfolio, rapid growth and an AI model. That does not automatically mean the AI caused the portfolio outcomes.
MYbank: automated SME lending at meaningful scale, dated honestly
MYbank’s “3-1-0” model, a three-minute application, one-second approval and zero manual intervention, is one of the most prominent examples of automated small and medium-sized enterprise (SME) underwriting. The model uses transaction and ecosystem data to assess businesses that may not have conventional financial statements or collateral.
The most recent figure I can source is 53 million SMEs served cumulatively, disclosed by Ant Group and MYbank in April 2024 as part of MYbank’s 2023 annual report (Businesswire, April 2024). An earlier IFC reference from 2020 cited 45 million SMEs served and a default rate near 1.5 percent, roughly 60 percent below the industry average at the time. We have not found a more recent independent update to either figure, so we are dating both explicitly rather than presenting them as current performance.
Context worth adding: the largest figures come from MYbank and Ant Group’s own disclosures, real and substantial, though not independently audited. The clearest claim, and a strong one on its own terms, is that AI-enabled underwriting has been embedded successfully in one of the largest SME lending operations anywhere.
JUMO: large-scale African digital lending with independent customer-protection assessment
One instructive counterpoint on the independent-verification question is JUMO, the pan-African lending infrastructure platform operating across Ghana, Kenya, Tanzania, Zambia, Uganda, Côte d’Ivoire, South Africa and Cameroon. JUMO reports $7.9 billion disbursed to more than 31 million people since 2015, a company-disclosed figure with the same attribution caveat as MYbank’s.
What sets it apart is that JUMO became the first digital financial services provider assessed against the Cerise+SPTF customer-protection standard, an externally administered review, and scored 92.2 percent. That is a rare instance in this field of a third party, rather than the company itself, producing the headline number.
Konfío: the rare case with an actual comparison group
Most of the lending examples in this field compare a portfolio against itself, before and after, or against an industry average. Konfío, a Mexican fintech lender using AI-driven credit scoring for micro, small and medium-sized enterprises (MSMEs), is one of the few with something closer to a genuine evaluation design. An analysis by IDB Invest, the private-sector arm of the Inter-American Development Bank, compared businesses that received a Konfío loan with similar applicants that were rejected. Two years on, borrowers who received financing grew sales 19.4 percent more than the rejected comparison group, and 41.9 percent more among women-owned businesses (IDB Invest).
That comparison-group design is a materially higher evidentiary bar than a simple before-and-after story, and it is why we would put Konfío ahead of most of the credit-scoring examples circulating in this sector. It does not remove every question. IDB Invest is also an investor in the company, so the evaluation is not fully arm’s length. But it is the closest thing in this list to a counterfactual, and it should be the reference point when a lender says its underwriting model “improves outcomes.”
Eshandi in Zambia: promising gender evidence, but not causal proof
Eshandi uses machine-learning scoring in a mobile-money nano-loan product and has disbursed close to a million loans to women. Borrower-level analysis in a 2026 IFC report found that women score marginally higher on the model’s own credit score and are more likely to become repeat borrowers (IFC, “Cracking the Credit Code,” May 2026). The report is careful about what this means, noting that most of these models are gender-neutral by design and that inclusion gains of this kind are largely incidental rather than engineered.
That is the right level of caution. The evidence shows an association between the scoring model, repeat-loan behavior and gender. It does not by itself prove that AI increased women’s financial inclusion or caused better economic outcomes. The case is useful precisely because it raises the right next questions: who enters the funnel, who is approved, who repays, and whether the model reproduces or reduces gender gaps by accident or by design.
Vexi in Mexico: product use matters as much as approval
Vexi’s digital credit card combines alternative-data underwriting with direct customer research. The same 2026 IFC report surveyed more than 3,000 Vexi borrowers and found that women who operated personal or family businesses used the card for business expenses at higher rates than men in the same survey, and that their average credit limits pulled ahead of men’s after roughly four years of card tenure.
This makes Vexi more than a credit-scoring story. It is a product-design example showing how access to flexible digital credit can support household enterprises and women-owned businesses over time, not just at the point of approval. The same caution applies as with Eshandi: surveyed usage and credit-limit growth are evidence of inclusive product use, not proof that the underwriting model alone caused business growth.
Agri-finance and climate insurance: real scale, harder AI attribution
Two African lenders and insurers, Apollo Agriculture and Pula, belong in this same FSP category, and they show real scale while illustrating why AI claims need to be separated from product outcomes, and why a target and a result are not the same thing.
Apollo combines data-driven credit assessment with inputs, agronomic advice and insurance for smallholder farmers in Kenya and, more recently, Zambia. A 2024 British International Investment case study cites a target of reaching more than 2.3 million farmers by 2026. That is a forward goal, not current reach. The most recent reach figures we can find put actual coverage at roughly 400,000 farmers across both markets as of 2025, with the Zambia operation still under 30,000 farmers after its first 18 months. The yield and quality-of-life figures Apollo and its investors cite, including a reported 2.6 times average yield versus other Kenyan farmers, are drawn from company and investor surveys, not independent evaluation. The defensible claim is that Apollo has built a scaled, digitally enabled agri-finance model that is still growing toward its stated ambitions, not that AI has independently been proven to increase yields or climate resilience. Those outcomes are influenced by credit, seed, fertiliser, advice, weather, insurance and farmer behaviour together.
Pula uses satellite, weather and agronomic data in agricultural insurance design, pricing and claims processes. By the company’s own current disclosure, it has insured 20.1 million farmers across 22 countries, paid out $133.9 million in claims to 2.8 million farmers, and protected $2.6 billion in agricultural investment. In October 2024, Pula processed what it describes as Africa’s largest single agricultural insurance payout, more than $29 million to over 800,000 farmers in Zambia following a severe drought. That evidence matters, insurance only creates resilience when policies are distributed at scale and claims are paid quickly enough to matter, but the number of farmers insured and the amount paid demonstrate the scale and performance of the insurance operation. They do not isolate the incremental contribution of machine learning or satellite analytics. Pula is a strong climate-insurance case with an AI-enabled operating model, not a controlled test of AI’s impact on resilience.
A better test for AI case studies
For boards, regulators and FSPs, we would apply five tests before calling an AI initiative successful.
- Is the system actually in use? A functioning prototype is useful, but it is not production.
- Is the AI component clearly identified? “Digital” and “AI-enabled” are often used too loosely.
- Are the results about the model or merely about the institution? A lender’s portfolio growth does not prove its algorithm worked.
- Are the metrics independently verified? Company-reported numbers can be valuable, but they should be labelled as such, and a genuine comparison group, like the one behind Konfío’s evaluation, or an independently obtained figure, like the one behind India’s mule-account rollout, should carry more weight than a company’s own press release.
- Are outcomes measured for customers as well as the institution? Faster processing is useful. Better access, fairer decisions, lower losses and improved resilience are more important.
What this means for regulators and FSPs
The evidence is now strong enough to move beyond generic experimentation, but not strong enough to suspend judgement.
For regulators, complaint handling, data-quality checks, fraud and mule-account detection, and targeted market monitoring are practical starting points. They use existing data, solve visible bottlenecks and allow human supervisors to remain accountable.
For lenders and insurers, alternative-data underwriting and automated decisioning are already expanding reach, particularly for thin-file consumers, women, MSMEs and smallholder farmers. The strongest examples measure inclusion across the full funnel, from application and approval to limits, pricing, repeat use, repayment and customer outcomes. Konfío’s comparison-group evaluation is the standard worth building toward, and more of the sector is capable of it than currently attempts it.
For the investors and development institutions that finance both sides of this equation, the useful diligence question asks what decision has changed, what evidence shows it is better, and who benefits, rather than simply whether the company uses AI.
Regulators are already catching fraud and clearing complaint backlogs faster. Lenders are reaching borrowers who were invisible to conventional underwriting. Insurers are paying claims within days instead of months. The evidence for AI in financial services is stronger, and more specific, than the conference circuit usually credits it for. The task now is to keep building that evidence base with the same rigor these examples were held to, so the next wave of cases earns its place beside them. The winners in this sector may not be the institutions with the biggest AI budgets. They may be the ones who can show their work
