African financial services operate on data that no machine learning textbook describes: transaction histories built on USSD sessions, airtime recharge patterns that proxy income, and mobile money ledgers that are often the only financial footprint for millions of users. Engineering teams building AI systems for African FinTech work with this data directly. The challenges are specific to the continent, and the solutions that work here have started influencing how AI is applied to financial services globally.
This post covers four areas where African FinTech engineering teams are actively building ML systems: fraud detection, alternative credit scoring, AML transaction monitoring, and conversational AI on constrained channels. For each, the engineering decisions differ from what Western FinTech teams typically build, and the reasons matter for anyone deploying these systems.
Fraud Detection on Mobile Money Networks
Fraud detection in mobile money operates on a fundamentally different transaction graph than card-based payment fraud. There are no BIN ranges, no merchant category codes, and no static cardholder billing addresses. What exists is phone numbers, agent IDs, transaction amounts, session timing, and the sequence of USSD interactions that preceded the transaction.
Feature engineering for mobile money fraud typically relies on velocity patterns (how many transactions from a given number in a given time window), agent network anomalies (unusual float drawdown at specific agent points), SIM swap recency (a recently swapped SIM attempting a high-value transfer is a strong fraud signal), and device fingerprinting where applicable.
Latency requirements are tighter than in card fraud. USSD sessions time out after 180 seconds. Fraud scoring must complete within the session window or the user experience breaks entirely. This means inference latency for real-time fraud models on mobile money networks is measured in tens of milliseconds, not hundreds. Teams that migrated fraud models designed for card transactions directly to mobile money found that model accuracy held but inference latency caused session timeouts at scale. The architectural response: lightweight gradient-boosted models at the transaction layer, with heavier neural models running asynchronous risk scoring for pattern detection over longer time windows.
For the broader context on combating fraud in FinTech, the patterns for mobile money extend to any mobile-first payment context where transaction data is thin and session latency is constrained.
Alternative Credit Scoring for Thin-File Borrowers
The credit bureau problem in African markets is not that credit data does not exist. It is that credit data for the population most in need of credit has never been captured by traditional credit infrastructure. A borrower who has used M-Pesa for five years has a detailed transaction history, but none of it appears in a credit bureau report.
Alternative credit scoring in African FinTech means building models on data that Western credit scoring ignores: mobile money transaction frequency and regularity, airtime recharge patterns as a proxy for cash flow, utility payment behavior where captured, and in some implementations, social graph analysis where appropriate consent is obtained.
The engineering challenges are distinct from model design. Consent and data minimisation requirements under Kenya's Data Protection Act, Nigeria's NDPR, and South Africa's POPIA create feature selection constraints. Features that might improve model performance may not be permissible under the applicable data protection framework. Compliance must be integrated into the feature pipeline, not checked at the end. Companies like Tala, Branch, and M-Shwari have operated these models in production across African markets and have published on the challenge of maintaining model performance and fairness simultaneously in high-volatility economic environments.
AML Transaction Monitoring Architecture
Anti-money laundering transaction monitoring in Africa is complicated by a structural feature of mobile money: cash is the input and output for much of the network. Cash-in at an agent point and cash-out at another agent point is a legitimate pattern for millions of users who have no other way to move money. It is also a common pattern for layering in money laundering. Distinguishing between the two requires context that rule-based AML systems do not capture well.
ML-based AML systems for mobile money networks typically run on a graph structure: the transaction graph of relationships between phone numbers, agent IDs, and counter-parties. Graph neural networks have shown strong results for detecting structured layering behaviour in this context, but the engineering cost is high. Graph construction, feature generation, and inference at mobile money transaction volumes requires significant infrastructure investment.
Regulatory requirements add a compliance layer to the engineering architecture. Nigeria's CBN, South Africa's FSCA, and Kenya's CBK all have AML/KYC reporting obligations with specific timelines for suspicious transaction reports. The monitoring system needs to output not just risk scores but audit trails that satisfy regulatory examination. Engineering teams that built monitoring systems without explicit compliance output found they had to retrofit audit logging, which is significantly more expensive than building it in from the start. For a broader view of how AI governance intersects with FinTech compliance requirements, see the AI governance in FinTech guide.
Conversational AI on Constrained Channels
USSD and WhatsApp dominate customer communication for mobile money users across Africa. Both channels impose significant constraints on conversational AI. USSD has a 182-character session limit per message and a linear menu structure that does not support free-form natural language. WhatsApp requires a Business API integration and explicit opt-in. Neither supports the open-ended dialogue that most conversational AI systems are designed for.
Effective conversational AI on USSD is a decision-tree problem with ML-assisted intent classification at each node. The NLP component is narrow: classifying ambiguous user inputs within a constrained menu context. Engineering teams that attempted to port general-purpose chatbot architectures to USSD found that session timeout and character limits made the approach unworkable. Purpose-built intent classifiers trained on USSD session data perform significantly better.
WhatsApp-based financial services have more headroom for natural language but still operate under significant constraints: message delivery is not guaranteed, sessions do not persist, and latency in low-connectivity environments means multi-turn dialogue design must assume gaps between turns and handle session resumption explicitly.
Build vs. Buy for AI in African FinTech
The build vs. buy decision for AI systems in African FinTech is shaped by data availability more than model sophistication. Fraud detection and credit scoring models built on proprietary transaction data cannot be purchased off the shelf because the models are the data. A fraud detection vendor's model is trained on their transaction data, which may not match your transaction data in a different market with different fraud patterns.
The general pattern across African FinTech: buy for non-differentiating AI components such as document OCR for KYC and face matching for identity verification, and build for fraud detection, credit scoring, and AML monitoring where proprietary transaction data is the competitive moat. Companies that have outsourced fraud detection to generic providers typically find that model performance degrades as African transaction patterns diverge from the training distribution.
Engineering teams building these systems from scratch benefit significantly from working with engineers who have already solved the infrastructure and compliance integration problems on live African mobile money rails. For the broader context on engineering decisions in this space, see the FinTech Engineering Playbook. For teams looking to work with FinTech software development organisations that have direct experience building fraud, credit scoring, and AML systems on African payment rails, Scrums.com has engineers across Nairobi, Lagos, and Cape Town with production experience on these systems.
Frequently Asked Questions
What AI/ML techniques do African FinTech companies use for fraud detection?
Fraud detection on mobile money networks typically uses gradient-boosted models for real-time transaction scoring within USSD session latency constraints, with graph-based models running asynchronously for pattern detection. Feature engineering focuses on velocity patterns, agent network anomalies, SIM swap recency, and transaction sequence characteristics specific to mobile money flows.
How does alternative credit scoring work in African markets?
Alternative credit scoring uses non-traditional data to assess creditworthiness: mobile money transaction history, airtime recharge frequency, utility payment behavior, and social graph data where consent permits. Models are trained on these features to predict repayment for users with no formal credit history. Feature selection must comply with applicable data protection frameworks in each market.
What are the AML compliance requirements for FinTech in major African markets?
AML compliance varies by market. Nigeria's CBN, South Africa's FSCA, and Kenya's CBK each have reporting obligations with specific timelines for suspicious transactions. Engineering teams need to build audit trail output into monitoring systems from the start, not retrofit it later, as retroactive compliance logging is significantly more expensive and harder to validate under regulatory examination.
How is conversational AI constrained by USSD in African FinTech?
USSD has a 182-character limit per message and a linear menu structure that does not support free-form natural language. Conversational AI on USSD is narrow-intent classification within a constrained decision tree. Teams should build purpose-built intent classifiers trained on USSD session data rather than adapting general-purpose chatbot architectures.
Should African FinTech companies build or buy their AI systems?
Build for fraud detection, credit scoring, and AML monitoring where proprietary transaction data is the competitive moat. Buy for non-differentiating components like document OCR and face matching. Generic AI vendors' fraud models are trained on data that may not match your market's fraud patterns, and credit scoring models require your own transaction history to perform accurately on thin-file users.
