Use cases · Kilwa Research

AI that paid off in markets like ours.

Eight deployments with measured outcomes, from Asia, Latin America and Africa, each told the same way: the result first, then what explains it. Then two opportunities in South Africa and Nigeria, modelled by Kilwa with every input sourced and the arithmetic shown.

One framework, every case. Outcome before narrative, so a committee can stop reading at the point it has what it needs.

01OutcomeThe result, first
02Key findingWhat explains it
03ProblemWhat was broken
04SolutionWhat was built
05Success metricsAs the source states them
06Why it travelsThe read for African markets

The evidence base

Chosen for what they measured, not for what they promised.

Every case reports an outcome in the source's own numbers. Half rest on peer-reviewed studies; the rest on company filings, regulator data or the publishers behind the strategy playbooks. Where a company reports its own result, the page says so.

documented deployments, each with measured outcomes

countries and regions, from China and India to Rwanda and the flood basins of a hundred countries

sectors: credit, banking, insurance, agriculture, health, logistics, climate

3 of 8

cases with a peer-reviewed study behind the headline figure

01MYbank 02WeBank 03DBS Bank 04Ping An 05ICRISAT 06Google 07Zipline 08Google Research, Flood Hub South Africa · Kilwa estimate Nigeria · Kilwa estimate

MYbank (Ant Group)

China
Small-business credit
2019–2021

45 million small businesses financed without collateral, at a bad-loan rate of about 1.5%.

Key finding. When transaction data does the underwriting, the cost of a small loan collapses and the risk does not rise with it.

Problem

Small and micro businesses had no collateral and thin credit files, and their loans were too small for a bank to underwrite by hand. Eight in ten of MYbank's borrowers had never had a business loan from a bank.

Solution

The 310 model: under three minutes to apply on a phone, under one second to approve, zero human intervention, with risk models built on payment, e-commerce and operating data rather than collateral.

Success metrics

small and micro business clients served by end-2021, up nearly 30% in a year: 45m; small and micro business clients served by end-2021, up nearly 30% in a year
non-performing loan ratio on the small-business book in 2021 (1.52% in 2020): 1.53%; non-performing loan ratio on the small-business book in 2021 (1.52% in 2020)
the industry average for comparison, per the regulator, in 2019: 3.22%; the industry average for comparison, per the regulator, in 2019
average loan size in 2019, about US$4,300; 80% of users were first-time bank borrowers: RMB 31k; average loan size in 2019, about US$4,300; 80% of users were first-time bank borrowers

Why it travels to African markets

Nigeria's unmet MSME credit demand was put at ₦13 trillion by IFC. Mobile-money and payment-switch data are the African equivalent of the transaction trail MYbank underwrites on, and the model is the template for the Nigeria estimate below.

WeBank

China
Digital banking
2024–2026

A bank that runs an account for RMB 1.9 a year in IT cost, at 99.999% availability.

Key finding. Serving the mass market profitably is an infrastructure-cost problem before it is a credit problem.

Problem

At conventional per-account IT costs, low-balance customers and small businesses are loss-making to serve, so they are not served. Fraud and credit decisions took days.

Solution

A distributed core system on commodity infrastructure, more than 100 AI applications across credit, fraud and service, and digital employees handling routine work, with new models deployed in as little as a day.

Success metrics

IT operation and maintenance cost per account per year: RMB 1.9; IT operation and maintenance cost per account per year
product availability, with 1.4 billion transactions a day at peak: 99.999%; product availability, with 1.4 billion transactions a day at peak
individual customers and 7.6 million small and medium enterprises served: 440m; individual customers and 7.6 million small and medium enterprises served
anti-fraud processing time per case; credit assessment ten times more efficient: 3 days → 5 min; anti-fraud processing time per case; credit assessment ten times more efficient

Why it travels to African markets

Mobile-money operators and tier-two banks across Africa carry tens of millions of low-balance accounts. The unit economics WeBank reports are the benchmark for whether those accounts can be served at a profit.

Company filing or releaseWeBank, press release via PR Newswire, July 2026: Best AI-Driven Bank of the Year, with the figures quoted

DBS Bank

Singapore
Banking, enterprise AI
2023–2024

More than S$750 million of measured economic value from AI in a single year, double the year before.

Key finding. The value came from hundreds of small, measured use cases run on one platform, not from a flagship project.

Problem

A decade of data investment with no line on the income statement that management could verify, and a customer base of thirteen million to serve individually.

Solution

An industrialised machine-learning platform with more than 1,500 models across more than 370 use cases: personalised nudges, scam and fraud detection, and productivity tools, each with its economic value measured.

Success metrics

economic value from data analytics and AI/ML in 2024, up from S$370m in 2023: S$750m+; economic value from data analytics and AI/ML in 2024, up from S$370m in 2023
models in production across more than 370 use cases: 1,500+; models in production across more than 370 use cases
personalised nudges sent to more than 13 million customers: 1.2bn; personalised nudges sent to more than 13 million customers
engaged Singapore customers saved, invested and were insured more than non-users: 2× · 5× · ~3×; engaged Singapore customers saved, invested and were insured more than non-users

Why it travels to African markets

The pan-African banks already hold the data estate. What DBS shows is the operating discipline: a platform, a portfolio of small use cases, and a value number per use case that a board can audit.

Company filing or releaseDBS Group Holdings, Annual Report 2024: Innovating impactful solutions for customers
Company filing or releaseDBS Group Holdings, Annual Report 2023 (economic value of S$370 million in 2023)

Ping An

China
Insurance claims
2024–2025

Claims closed in an average of 7.4 minutes, and more than RMB 10 billion a year saved from fraud.

Key finding. Automating the routine claim frees the money and the adjusters for the claims that need judgement.

Problem

High-volume motor and personal-injury claims were slow to settle, expensive to handle and leaked to fraud, while service volumes outgrew what human representatives could absorb.

Solution

Smart Quick Claim: one-sentence reporting, one-click upload and one-minute review, with image recognition for loss assessment, AI service representatives for first contact, and fraud models across the claims book.

Success metrics

average time to close a claim with Smart Quick Claim, first nine months of 2024: 7.4 min; average time to close a claim with Smart Quick Claim, first nine months of 2024
of customer service volume handled by AI representatives: about 1.34 billion interactions in nine months: 80%; of customer service volume handled by AI representatives: about 1.34 billion interactions in nine months
of personal-injury claims settled automatically in 2025, in as little as 51 seconds, nearly one million cases: 70%; of personal-injury claims settled automatically in 2025, in as little as 51 seconds, nearly one million cases
claims savings from smart fraud detection in 2025, above RMB 10 billion for the third year: RMB 10.5bn; claims savings from smart fraud detection in 2025, above RMB 10 billion for the third year

Why it travels to African markets

South Africa has the continent's deepest short-term insurance market and Kenya and Nigeria are digitising motor cover. The claims economics Ping An reports are the anchor for the South Africa estimate below.

Company filing or releasePing An, results announcement for the first nine months of 2024
Company filing or releasePing An Group, Technology (2025 figures)

ICRISAT with Microsoft

India
Smallholder agriculture
2016–2017

Sowing-date advice by text message lifted smallholder yields by 10% to 30%.

Key finding. The cheapest AI intervention is often a date, delivered in time, on the phone the farmer already owns.

Problem

Rain-fed smallholders sowed on habit and hearsay. A mis-timed sowing costs a season's yield, and no extension service can reach every village in the window that matters.

Solution

Machine learning on thirty years of climate and sowing data, combined with weather forecasts, sends each farmer the optimal sowing window by SMS, with a village dashboard for soil health, fertiliser and a seven-day forecast.

Success metrics

average yield per hectare for groundnut in the 2016 pilot with 175 farmers in Andhra Pradesh: +30%; average yield per hectare for groundnut in the 2016 pilot with 175 farmers in Andhra Pradesh
farmers in the 2017 season across Andhra Pradesh and Karnataka: 3,000+; farmers in the 2017 season across Andhra Pradesh and Karnataka
yield gains across groundnut, ragi, maize, rice and cotton in 2017: +10–30%; yield gains across groundnut, ragi, maize, rice and cotton in 2017
of climate and sowing data behind the model: 30 yrs; of climate and sowing data behind the model

Why it travels to African markets

Most of Africa's farming is rain-fed and most farmers hold a basic phone. The intervention needs no new hardware, which is why it is the most portable case on this page.

Publisher case studyMicrosoft News, January 2017: Microsoft and ICRISAT's intelligent cloud pilot for agriculture in Andhra Pradesh
Publisher case studyMicrosoft Stories India: Farmers in India are using AI to increase crop yields

Google with Thai and Indian health partners

Thailand and India
Health screening
2018–2024

Screening for diabetic eye disease at specialist-level accuracy in clinics that have no specialist.

Key finding. A validated model is the start; the deployment work in the clinic is what decides the outcome.

Problem

Diabetic retinopathy causes preventable blindness, and there are far too few ophthalmologists to screen every diabetic patient, so most are never screened until damage is done.

Solution

A deep-learning model that grades retinal photographs, tested prospectively inside Thailand's national screening programme, then licensed to local partners who run it in public-sector clinics.

Success metrics

patients screened in the prospective Thai study, December 2018 to March 2020: 7,940; patients screened in the prospective Thai study, December 2018 to March 2020
accuracy for vision-threatening disease, with 91.4% sensitivity and 95.4% specificity: 94.7%; accuracy for vision-threatening disease, with 91.4% sensitivity and 95.4% specificity
screenings supported by the model in clinics worldwide by October 2024: 600,000+; screenings supported by the model in clinics worldwide by October 2024
no-cost screenings targeted over ten years with Forus Health, AuroLab and Perceptra: 6m; no-cost screenings targeted over ten years with Forus Health, AuroLab and Perceptra

Why it travels to African markets

Specialist scarcity is the African condition. The lesson Kilwa takes into its sovereign AI work is the second one: the model was the easy part, and the clinic workflow, connectivity and image quality decided whether it helped.

Zipline with Rwanda's Ministry of Health

Rwanda
Medical logistics
2016–2020 study period

Maternal deaths from haemorrhage fell 51% at hospitals served by autonomous drones.

Key finding. Autonomy in the last mile changed clinical outcomes, not just delivery times, and cut the inventory hospitals had to hold.

Problem

Blood products expire. Rural hospitals either stocked more than they could use or ran out in an emergency, and road delivery took hours.

Solution

Centralised blood inventory at two drone ports, with autonomous aircraft delivering to order in 15 to 60 minutes, so hospitals could hold less and get exactly what a patient needed.

Success metrics

in-hospital deaths of mothers with postpartum haemorrhage at drone-served facilities: −51%; in-hospital deaths of mothers with postpartum haemorrhage at drone-served facilities
trauma-patient mortality: −30%; trauma-patient mortality
blood-product wastage, with red-cell inventory down 63%: −40%; blood-product wastage, with red-cell inventory down 63%
public hospitals served from two ports by June 2020: 29; public hospitals served from two ports by June 2020

Why it travels to African markets

This is an African case, peer-reviewed, and the model has since been carried to Ghana, Nigeria, Kenya and Côte d'Ivoire. It is the clearest evidence on this page that an autonomous system can move a national health statistic.

Google Research, Flood Hub

Global, including most of Africa
Climate risk and public warning
2024

Reliable river-flood warnings up to five days ahead for basins that have no gauges at all.

Key finding. A model trained on the world's gauged rivers can give an ungauged African basin the forecast quality of a European one.

Problem

About 1.5 billion people, 19% of the world, are exposed to severe flood risk, and most African rivers are ungauged, so warnings arrive late or not at all.

Solution

A global AI hydrology model that learns from gauged basins and forecasts ungauged ones, published in Nature and delivered as Flood Hub forecasts and alerts through the products people already use.

Success metrics

lead time at which reliability matched or beat the current global system's same-day nowcasts: 5 days; lead time at which reliability matched or beat the current global system's same-day nowcasts
days of reliable lead time added to available global nowcasts, on average: 0 → 5; days of reliable lead time added to available global nowcasts, on average
countries covered in March 2024, with African and Asian forecasts raised to European levels: 80+; countries covered in March 2024, with African and Asian forecasts raised to European levels
people covered in more than 100 countries by the end of 2024: 700m; people covered in more than 100 countries by the end of 2024

Why it travels to African markets

Nigeria's 2022 and 2024 floods were forecastable events that arrived as surprises. Early warning is sovereign infrastructure, and it now exists at near-zero marginal cost for governments that choose to wire it into their response.

Peer-reviewed studyNearing et al., Nature, March 2024: global prediction of extreme floods in ungauged watersheds
Publisher case studyGoogle Research blog, March 2024: using AI to expand global access to reliable flood forecasts

Two opportunities, modelled by Kilwa

South Africa and Nigeria, with the arithmetic shown.

The same framework, applied forward. Each opportunity is anchored to the cases above, every input carries its source or is labelled a Kilwa assumption, and three scenarios bracket the answer, including the one where it loses money.

Kilwa estimate

South Africa · Short-term (non-life) insurance

Straight-through claims for South Africa's non-life insurers

Highlighted: South Africa. All 54 markets are covered on the platform.

Kilwa estimate: R2.7 billion to R8.0 billion a year of industry-wide pre-tax value, R4.9 billion in the base case, from AI claims handling and fraud detection.

Key finding. The value is in the ordinary claim, not the exotic one: settle the routine motor and property claim in minutes and put the adjusters on the claims that leak.

Problem

The primary non-life insurers earned about R139 billion of net premium in 2025 (annualised) and paid out about R71 billion in claims. The Insurance Crime Bureau's consensus is that 5% to 10% of claims paid are fraudulent, and the industry's management expenses ran at 31.6% of net earned premium.

Solution

Ping An-style straight-through processing for simple claims (photo-based loss assessment, automated settlement, exception escalation) and fraud models across the claims book, with the human adjusters concentrated on complex and suspicious cases.

What the cases above establish

Ping An settled claims in an average of 7.4 minutes and saved RMB 10.5 billion from fraud detection in 2025 (case 4).
Stanford's 51 deployments found escalation models, where AI handles 80%+ and humans review exceptions, delivered 71% median productivity gains (Latest “Best” AI Strategies, source 1).
BCG's leader's guide cites efficiency gains of up to 50% when a critical function is reshaped end to end (source 5).

Success metrics, targets

target share of simple motor and property claims settled straight-through within three years: 50%; target share of simple motor and property claims settled straight-through within three years
target settlement time for straight-through claims, from days today: Minutes; target settlement time for straight-through claims, from days today
base-case reduction in claims cost from fraud and overpayment detection: 3%; base-case reduction in claims cost from fraud and overpayment detection
base-case improvement in the industry combined ratio: 3.5 pts; base-case improvement in the industry combined ratio

The arithmetic, Kilwa estimate

Low · Base · High

Annual value = (net claims × leakage recovered) + (claims-handling cost × handling saving), where claims-handling cost = management expenses × the share attributed to claims handling.

Inputs to the South Africa estimate with sources
Input	Low	Base	High	Source
Net claims paid, 2025, annualised from the December-2025 quarter (R17.9bn × 4)	71.5 R bn	71.5 R bn	71.5 R bn	Prudential Authority, Selected insurance sector data, December 2025, non-life primary insurers
Management expenses, 2025, annualised (R11.0bn × 4)	43.9 R bn	43.9 R bn	43.9 R bn	Prudential Authority, December 2025
Share of management expenses attributable to claims handling (Kilwa assumption)	25 %	25 %	25 %	Kilwa assumption; loss-adjustment cost is typically a tenth to a sixth of claims in non-life books
Leakage recovered by fraud and overpayment detection, as a share of net claims	1.5 %	3 %	5 %	Below the SAICB consensus that 5–10% of claims paid are fraudulent; Ping An's 2025 savings anchor the high case
Claims-handling cost saved by straight-through processing	15 %	25 %	40 %	Stanford escalation-model gains and BCG's up-to-50% bound the range; the low case assumes partial adoption

Outputs of the South Africa estimate
Output	Low	Base	High
Leakage recovered	R1.1bn	R2.1bn	R3.6bn
Handling cost saved	R1.6bn	R2.7bn	R4.4bn
Annual pre-tax value, industry-wide	R2.7bn	R4.9bn	R8.0bn
Combined-ratio improvement (value ÷ net earned premium of R139bn)	1.9 pts	3.5 pts	5.7 pts
For one insurer with a 15% market share(pro rata)	R0.4bn	R0.7bn	R1.2bn

How to read it. The base case is worth about 3.5 points on an industry combined ratio that stood at 90.6% in 2025, which is the difference between a good year and an excellent one. The arithmetic does not include the cost of building it or the premium effect of faster settlement, both of which Kilwa would size in an engagement.

What would prove it wrong

Fraud detection recovers under 1.5% of claims after two years, which would mean the leakage is smaller than the Bureau's consensus or the models are not finding it.
Straight-through settlement stays below a quarter of simple claims because of policy wording, repair-network or regulatory constraints.
Management expenses attributable to claims handling prove to be well under a quarter of the total, which shrinks the handling saving proportionately.

Sources

Official statisticsSouth African Reserve Bank, Prudential Authority: Selected South African insurance sector data, December 2025
Official statisticsInsurance Crime Bureau (SAICB): fraud estimated at 5–10% of claims paid
Press reportKPMG, The South African Insurance Industry Survey 2023 (non-life gross written premium of R140.1bn in 2022)

Kilwa estimate

Nigeria · MSME lending

A 310-style lender for Nigeria's small businesses

Highlighted: Nigeria. All 54 markets are covered on the platform.

Kilwa estimate: a ₦650 billion AI-underwritten MSME book earns about ₦52 billion a year pre-tax in the base case, finances 325,000 businesses, and loses money if credit costs are not held near 4%.

Key finding. The opportunity is real and the margin is thin: at Nigerian funding costs the model only works if AI underwriting keeps losses closer to MYbank's 1.5% than to the banking system's 8% to 10%.

Problem

Nigeria has 39.7 million MSMEs producing 46% of GDP and 88% of employment, and IFC put their unmet credit demand at ₦13 trillion. Bank lending prices at 20% to 35%, the industry's non-performing loan ratio reached 8.1% in 2025 and 9.9% in early 2026, and manual underwriting cannot reach a ₦2 million loan profitably.

Solution

A MYbank-style digital lender or bank partnership: cash-flow underwriting on payment-switch, mobile-money and e-commerce data, approval in seconds, no collateral, exception review by humans, funded through deposits and on-lending lines.

What the cases above establish

MYbank served 45 million small businesses at a 1.53% non-performing loan ratio, against a 3.22% industry average (case 1).
WeBank cut anti-fraud processing from three days to five minutes and runs an account for RMB 1.9 a year (case 2).
Stanford's finding that model choice is a commodity in 42% of cases: the moat is the data and the orchestration, which in Nigeria means the payment rails (Latest “Best” AI Strategies, source 1).

Success metrics, targets

target annual credit-loss rate on the AI-underwritten book: ≤4%; target annual credit-loss rate on the AI-underwritten book
target time to a decision, from days today: Seconds; target time to a decision, from days today
base-case number of MSMEs financed at an average loan of ₦2 million: 325k; base-case number of MSMEs financed at an average loan of ₦2 million
base-case pre-tax return on the book: 8%; base-case pre-tax return on the book

The arithmetic, Kilwa estimate

Low · Base · High

Pre-tax profit = loan book × (gross yield − cost of funds − credit-loss rate − operating-cost ratio). Loan book = share of IFC's unmet demand captured over three years. MSMEs financed = book ÷ average loan.

Inputs to the Nigeria estimate with sources
Input	Low	Base	High	Source
Unmet MSME credit demand (IFC, 2022 survey; nominal, so conservative today)	13,000 ₦ bn	13,000 ₦ bn	13,000 ₦ bn	IFC, Market Bite Nigeria, 2022
Share captured over three years	2 %	5 %	10 %	Kilwa assumption
Gross yield on MSME loans	24 %	28 %	32 %	Between the average prime rate (19.5%, January 2026) and the maximum lending rate (32.7–35.2%, 2026), per CBN data
Cost of funds	15 %	12 %	10 %	Average savings rate 8.25% (October 2025) with the policy rate at 27%; blended deposit and wholesale funding
Credit-loss rate	7 %	4 %	2.5 %	Between MYbank's 1.53% and the Nigerian industry's 8.1–9.9% NPL ratio
Operating-cost ratio (digital origination and servicing)	6 %	4 %	3 %	Kilwa assumption; conventional MSME lending runs well above this
Average loan size	2 ₦ m	2 ₦ m	2 ₦ m	Kilwa assumption; MYbank's 2019 average was about US$4,300

Outputs of the Nigeria estimate
Output	Low	Base	High
Loan book	₦260bn	₦650bn	₦1,300bn
Pre-tax margin on the book (yield − funds − losses − opex)	−4.0%	8.0%	16.5%
Pre-tax profit per year	−₦10bn	₦52bn	₦215bn
MSMEs financed	130,000	325,000	650,000
Break-even credit-loss rate in the base case(yield 28 − funds 12 − opex 4)	·	12%	·

How to read it. Read the low case first: with losses at the banking system's level and funding at 15%, the book loses ₦10 billion a year. The case for AI underwriting in Nigeria is not that the demand exists, which IFC has established, but that the loss rate can be held near 4% while funding stays near 12%. That is a testable proposition, and it is the one Kilwa would underwrite before anything else.

What would prove it wrong

Credit losses on cash-flow-underwritten loans run above 7% after twelve months of seasoning.
Blended funding costs stay above 15% as the policy rate holds near 27%.
The Central Bank's rules on digital lending, data access or pricing cap the model's yield or its data inputs.

Sources

Development institutionIFC, Market Bite Nigeria: innovation offers key to the broader MSME finance market, 2022
Official statisticsNational Bureau of Statistics and SMEDAN, National Survey of MSMEs 2021
Official statisticsCentral Bank of Nigeria, monetary policy decisions (policy rate 27% at the November 2025 meeting)
Press reportBusinessDay, 2026: industry NPL ratio near 10% after the forbearance exit
Press reportThisDay, 2026: average maximum lending rate at 32.68% and 35.17%; average prime rate 19.54%

Notes on sources and estimates

Each case is Kilwa's paraphrase of the cited source, with figures quoted as the source states them and the period they refer to. Company names and product names identify the works; Kilwa is not affiliated with any organisation named here.

Sources are labelled by kind. Company filings and releases and publisher case studies report an organisation's own results; peer-reviewed studies are independent. Read the metrics with that distinction in mind.

Kilwa estimates are illustrative models of a market opportunity. Inputs are sourced where a source exists and labelled as Kilwa assumptions where it does not; the formula is stated so the reader can rerun it. They are not forecasts of any company's results, not investment advice, and not a recommendation of any security or transaction.

Where a regulator publishes quarterly flows, Kilwa annualises the latest quarter and says so. Rounded figures are shown; the underlying arithmetic uses the unrounded inputs.

The strategy behind these cases is synthesised in Latest “Best” AI Strategies

Size your own opportunity

Bring a market. We will bring the arithmetic.

Kilwa scopes and models opportunities like these for the markets you are actually in: inputs sourced, assumptions labelled, scenarios bracketed, and the conditions that would prove the case wrong stated before the money moves.

Scope an engagement Market entry and growth