WEEK 03

Data Strategy and Readiness

For Low-Resource Environments
Data quality · Thin-file populations · Low-resource languages · Data pipelines
Reading time~60–90 min
FormatRead · Apply · Deliver
ContextAfrican fintech examples

In this session

The most common constraint on AI product quality in African markets is not model capability — frontier models are accessible to any team with an API key. The constraint is data. Specifically: the right data, in the right format, at the right quality, with the right permissions, about the right population. This session is about making an honest assessment of your data position before building anything. It will be uncomfortable in places. Organisations that have been collecting data for years often discover, when they examine it rigorously, that the data they have is not sufficient for the AI systems they want to build. That discovery is valuable — it is far better made before six months of engineering effort than after. We will cover how to audit data for AI readiness, the specific challenges of thin-file and low-resource language contexts, what to do when the data you need does not exist, and how to design data collection strategies that build AI capital over time.

What makes data 'AI-ready'

Data that is adequate for business operations is frequently inadequate for AI development. These are different standards, and conflating them is one of the most common causes of expensive false starts in AI projects. Business operations require data that is available (you can retrieve it when needed) and actionable (it tells you something useful for the decision at hand). AI development additionally requires data that is representative (it covers the full distribution of cases the model will encounter in production), labelled (for supervised learning, you need ground truth outcomes paired with inputs), sufficient in volume (enough examples for the model to learn robust patterns), permissioned (you have legal rights to use it for training), and stable (the underlying distribution it represents is not changing faster than your model can adapt). Each of these additional requirements deserves examination. Representativeness is the most frequently violated. A lender who has been operating for five years has historical data — but only on applicants who were approved (or in some cases, applicants who were applied). Declined applicants, whose true creditworthiness is unknown, are absent. This creates survivorship bias: any model trained on this data learns the credit characteristics of people your previous process found acceptable, not the true distribution of creditworthy people in the market. The model will reproduce and potentially amplify your historical approval patterns, including any biases embedded in them. Label quality is the second most common problem. Labels are the ground truth outcomes paired with training examples. In credit, labels are typically 'defaulted' vs 'repaid'. But what counts as a default? 90 days past due? Write-off? First missed payment? If the label definition changed over the five-year dataset, or was inconsistently applied, the model is learning an inconsistent target. Label quality problems are invisible to standard data quality checks — you need to understand the business processes that generated the labels. Volume thresholds vary significantly by task. A simple binary classifier on structured data can achieve reasonable performance with a few thousand labelled examples. A fine-tuned LLM for a specialist task requires tens of thousands. A new model for a new market or population requires a minimum volume of examples from that specific distribution — you cannot simply transfer a model trained on Kenyan borrowers to an Ivorian context without significant data from Ivory Coast. Permission is the requirement most frequently assumed rather than verified. Data collected for one purpose (processing a loan application) may not be permissioned for a secondary purpose (training a machine learning model). In markets with data protection regulation — Kenya's Data Protection Act, South Africa's POPIA, Nigeria's NDPR — this is not a theoretical risk. An AI system built on improperly permissioned data is both legally exposed and potentially required to be shut down, destroying the investment.

The thin-file challenge

A thin-file customer is someone with insufficient formal credit history for a traditional credit assessment. In African markets, the thin-file population is not a niche — it is the majority. Across Sub-Saharan Africa, fewer than 20% of adults have formal credit bureau records. This means that any AI-powered credit product that requires bureau data as input has structurally excluded most of its addressable market before it starts. The response to this challenge is alternative data — signals about creditworthiness that are correlated with repayment behaviour but do not require formal credit history. The most valuable alternative data sources in African market contexts are: Mobile money transaction data. M-Pesa, Airtel Money, MTN MoMo, and similar services provide rich longitudinal records of income, expenditure, savings behaviour, and network connections. Transaction regularity, income recurrence, the ratio of inflows to outflows, and patterns of financial stress (frequent small withdrawals, balance below threshold for extended periods) are all predictive of credit behaviour. The challenge is access: this data is owned by mobile money operators (MNOs), who are potential partners but also potential competitors, and who have regulatory obligations around data sharing. Utility and rent payment history. Regular payment of electricity, water, and rent bills is predictive of general financial discipline. Collection and standardisation of this data is less mature than mobile money data, but growing through partnerships with utilities and property managers. Airtime purchase patterns. The frequency, amount, and timing of airtime purchases is a surprisingly strong proxy for income level and cash flow patterns. Customers who buy airtime in large weekly amounts behave differently from those who buy small daily amounts — and these patterns are accessible to MNOs and, through partnership, to lenders. Behavioural and engagement data. In app-based lending contexts, the customer's interaction with the app itself — session frequency, feature exploration, time to complete applications — can be predictive, though this data source requires careful governance to avoid discriminatory patterns (e.g. penalising customers who use older phones or slower connections). Social network data. In markets where social lending and group savings (chamas, tontines, ROSCAs) are common, social network characteristics are predictive of creditworthiness. A customer who is a consistent contributor to a chama is demonstrating financial commitment that credit bureau data would not capture. The critical governance issue with alternative data is proxy discrimination. Many alternative data signals that are predictive of credit behaviour are correlated with protected characteristics — geography, phone model, transaction type. A model that uses these signals may produce discriminatory outcomes even without explicitly using a protected characteristic as input. This requires both technical mitigation (careful feature selection, bias testing across demographic groups) and legal clarity (what does non-discrimination mean in the absence of well-established credit scoring regulation in your market).

Low-resource language challenges

Africa has approximately 2,000 languages, of which only a handful — Swahili, Amharic, Hausa, Yoruba, Zulu, and a few others — have significant representation in the training corpora of frontier AI models. For the remaining languages, AI models trained primarily on English or French data are not just less capable — they may produce outputs that are actively harmful: grammatically incorrect, culturally inappropriate, or carrying unintended connotations. The practical implications for AI product development are significant. An AI voice agent deployed for collections calls in Tanzania needs to handle Swahili — including regional dialectical variation and code-switching with English or Arabic. An AI-powered customer service chatbot for a Nigerian lending platform needs to handle Nigerian Pidgin, which is distinct from both standard English and formal Yoruba or Igbo. A document extraction system processing loan applications from a mixed Francophone/Anglophone West African market needs to handle documents in multiple languages with inconsistent transliteration of names and addresses. There are four approaches to the low-resource language problem, which are not mutually exclusive: Use specialist vendors. Companies like ToumAI, Lelapa AI, and Masakhane are building language technology specifically for African languages. Their models are trained on curated African language data and, for the languages they cover, typically outperform frontier general models. The tradeoff is depth of coverage (they support a subset of African languages) and the usual build/buy questions around integration and dependency. Fine-tune on domain-specific data. A base model with reasonable (not excellent) performance in a target language can often be significantly improved for a specific task through fine-tuning on in-domain data. A customer service LLM fine-tuned on 10,000 examples of actual loan-related customer queries in Swahili will typically outperform a general Swahili model on that specific task. The constraint is data collection: you need those 10,000 examples, which requires either a labelling programme or the deployment of a less capable system to generate real interaction data. Translation as a bridge. For tasks where the linguistic nuance is lower (structured data extraction, classification), routing inputs through a translation layer (translate to English, process, translate back) can be acceptable. The quality of translation for many African languages has improved significantly with recent models. The risks are translation errors compounding with model errors, latency (two API calls instead of one), and loss of culturally specific nuance. Invest in data collection. The most durable approach is to invest in building the data infrastructure to collect labelled examples in target languages. This requires a labelling programme (recruiting native speaker labellers, building annotation tooling, quality control processes) and a longer time horizon. But it produces a proprietary data asset that competitors cannot easily replicate, and it creates the foundation for fine-tuning that will compound in value over time.

Building a data readiness scorecard

A data readiness scorecard is a structured assessment of your organisation's data position for AI development. It is most useful when done honestly — the temptation to be optimistic about data quality is real, but it produces downstream costs that dwarf the discomfort of the honest assessment. The scorecard has six dimensions, each rated on a simple three-level scale: Ready (this dimension is not a blocker), Workable (this dimension is a moderate challenge that can be addressed with effort), and Blocker (this dimension is a hard constraint that must be resolved before building). Volume. Do you have enough examples of the target behaviour to train or fine-tune a model? For structured classification tasks, a minimum of 5,000 labelled examples is a rough threshold for 'workable'; 20,000+ is 'ready'. For LLM fine-tuning, thresholds are higher. For a brand-new market with no historical data, this is typically a blocker. Label quality. Are your outcome labels accurate, consistently defined, and correctly paired with inputs? This requires a manual audit of a sample of records, not just a programmatic check. Common label quality issues: inconsistent default definitions over time, data entry errors in outcomes, look-ahead bias (labels derived from information that was not available at the time of the original decision). Representativeness. Does your historical data cover the full population you intend to serve, or is it biased toward a subpopulation? For a lender expanding from salaried to informal sector customers, training data from the salaried period is not representative of the new target population. Completeness. What percentage of records have complete data for the features you intend to use? Missing data rates above 20% on key features are typically a workable challenge; rates above 40% are a blocker without a missing-data strategy. Recency. How old is your training data, and does the underlying distribution change over time? In markets with high economic volatility — common in African markets — data more than 2–3 years old may represent a different environment than today. This affects how much weight to give older cohorts in training. Permission and compliance. Do you have verified legal rights to use the data for model training? Has a privacy impact assessment been completed? Is PII managed appropriately (anonymised or pseudonymised for training)? For any organisation operating in a regulated sector (banking, telecoms), this dimension is non-negotiable and should be assessed with legal counsel.

Data collection strategies for AI-native organisations

The most forward-looking approach to data strategy is to design AI data collection into your product from the start, rather than trying to retrofit it from historical records. An AI-native data strategy has three components. Outcome capture by design. Every decision your product makes should have an associated outcome that is automatically captured and linked back to the original decision record. For a lending product, this means: the application and all its features → the model's output at decision time → the loan officer's final decision → the repayment outcome at defined intervals → the default or write-off outcome if relevant. Most organisations capture some of these but not all, and the linkage between them is incomplete. Without complete linkage, you cannot compute model accuracy, detect drift, or train new model versions. Exploration budgets. A purely exploit strategy — using the best current model to make every decision — generates data only about the applicants your current model finds acceptable. To improve the model over time, especially for underserved populations, you need an exploration strategy: allocating a defined fraction of decisions to a randomised or exploratory policy, specifically to generate data about the model's blind spots. This is a business cost (you will accept some bad loans deliberately) but it is an investment in model improvement. In credit contexts, exploration should be done at the lowest possible loan amounts to limit expected loss. Labelling infrastructure. For tasks where ground truth labels are generated by human judgment (document classification, fraud determination, credit decisions on edge cases), you need tooling and processes to capture those human labels efficiently and in a format usable for model training. This is frequently an afterthought — labels exist in email threads, spreadsheets, or decisions recorded in ways that are difficult to link back to the original input data. Building a simple labelling interface early, even before you need it for model training, is a high-return investment.

Self-check questions

Before moving to the next session, make sure you can answer these without re-reading. If you cannot, go back to the relevant section.

Session deliverable

Data Readiness Scorecard

Complete an honest data readiness assessment for an AI initiative you are considering. Rate each dimension, explain your rating, and identify the top three gaps to close before building.

Show worked example ↓
Thin-file credit scoring model for informal sector (jua kali / market trader) customers in Kenya
Show worked example ↓
Workable — moderate gap, addressable with effort
Show worked example ↓
We have 85,000 labelled loan records but only ~4,200 are from informal sector customers (vs. salaried). For the target population, this is borderline-workable for a structured classifier but insufficient for fine-tuning. Need a 12-month data collection push targeting informal sector originations.
Show worked example ↓
Workable — some inconsistency, manageable with cleaning
Show worked example ↓
Default definition changed from 60DPD to 90DPD in mid-2022 — affects ~18,000 records. Cohorts before 2022 need relabelling against the new definition. Outcome tracking is complete for all records.
Show worked example ↓
Blocker — significant population bias
Show worked example ↓
Training data is 95% salaried or formal employment customers. Informal sector customers are structurally different (irregular income, mobile money primary rather than bank account). Model trained on current data will perform poorly on the target population and may produce discriminatory outcomes.
Show worked example ↓
Workable — requires review and some remediation
Show worked example ↓
Consent for data processing captured for all customers but does not explicitly mention secondary use for ML model training. Pre-2022 cohorts need a legal review under the Data Protection Act 2019. PII not stripped from training dataset — anonymisation workstream needed before any model development.
Show worked example ↓
1. Population gap: recruit 8,000 informal sector customers through agent network over next 6 months, with explicit ML training consent. Owner: Head of Data, target: Q3. 2. Label remediation: relabel 18,000 pre-2022 records against 90DPD definition. Owner: Data Engineering, 4-week workstream. 3. Permission remediation: legal review of consent language + PII anonymisation pipeline. Owner: Legal + Data Engineering, 6-week workstream.
← Week 02 Week 04 →