Most AI projects fail before a line of code is written. They fail in the moment someone decides to 'add AI' to a product without being able to complete the sentence: AI will help our users do X better than they can do it today, and here is how we will know that it worked.
This session is about making that decision rigorously. It is the most important session in the programme, because every downstream choice — what data you need, which model to select, how to design the UX, how to measure reliability — flows from the clarity of your product thesis. A weak strategy produces expensive, slow, and ultimately abandoned AI systems. A strong strategy makes everything downstream easier.
We will cover how to identify genuine AI opportunities in the African market context, how to evaluate them against your constraints, and how to write a product thesis tight enough to make real decisions from.
Why most AI features fail before launch
The most common failure pattern in AI product development is not technical. It is strategic. A team observes that competitors are 'doing AI', or that a board member has asked about the company's AI strategy, or that a new LLM capability looks impressive in a demo. They begin building. Six months and considerable engineering effort later, they have a system that works technically but does not produce business value — and they cannot explain why.
The root cause is almost always the same: the problem was not defined before the solution was chosen. When AI is the starting point rather than the conclusion, teams optimise for the wrong thing. They optimise for model performance rather than user outcome. They optimise for technical novelty rather than problem severity. They optimise for what the technology can do rather than what their users actually struggle with.
There are three specific pathologies worth naming:
The demo trap. An LLM produces a compelling output in a controlled demo. The team extrapolates from the demo to production. What they fail to account for is that demos are carefully staged — the inputs are clean, the edge cases are absent, and the evaluator is charitable. Production is the opposite: inputs are messy, edge cases are constant, and users have zero tolerance for inconsistency. A feature that looks 90% ready in demo is typically 50% ready for production, and the gap between 50% and production-ready is where most projects die.
The feature inflation trap. AI is added to a product not because there is a user problem that AI uniquely solves, but because the product needs to look more sophisticated. The result is a product with an AI layer that users either ignore or actively distrust. This is particularly common in B2B products where the buying decision is made by someone other than the end user. The buyer wants to see 'AI-powered' in the brochure; the user wants to complete their task without friction.
The data assumption trap. A team decides to build an AI feature and only then goes looking for the data to train or fine-tune it. They discover that the data does not exist, is too noisy to use, or is owned by a third party. This trap is particularly dangerous in African market contexts, where formal data infrastructure is younger, alternative data sources are less standardised, and low-resource languages have sparse training corpora. We will return to this in detail in Session 3.
The antidote to all three is a disciplined product strategy process that precedes any technical decision.
The AI opportunity taxonomy
Not all AI use cases are equal. Before evaluating any specific opportunity, it helps to have a taxonomy of the types of AI value that exist, because different types require different data, different models, different UX, and different success criteria.
Automation opportunities replace a manual, repetitive human task with a system that performs it faster, at lower cost, or at higher scale. The value is efficiency. Examples: automated document classification, loan application pre-screening, invoice reconciliation. The key constraint is that the task must be sufficiently well-defined that the system can be evaluated against a clear standard. Automation AI almost always requires labelled examples of correct outputs.
Augmentation opportunities make a human expert more effective at a task they already perform, rather than replacing the task entirely. The value is quality or throughput improvement. Examples: a credit analyst using AI to surface relevant signals from a borrower's transaction history, a customer service agent using AI to draft responses that they then review and send. Augmentation is often lower-risk than automation because the human remains in the loop, but it also tends to produce smaller efficiency gains and requires careful UX design so that the AI recommendation does not override human judgment inappropriately.
Discovery opportunities surface insights that a human could not find manually, typically because the data volume or dimensionality is too large. The value is finding patterns that create competitive advantage. Examples: identifying creditworthy thin-file customers from mobile money transaction patterns, detecting fraud signals that no rule-based system would catch. Discovery AI tends to have the highest upside but the highest data requirements and the longest validation cycles.
Interaction opportunities create new ways for users to engage with a product — most commonly through conversational AI, voice interfaces, or personalized content. The value is user experience and accessibility. Examples: a USSD-based loan advisor that guides users through eligibility in natural language, a voice agent that handles collections calls in Swahili or Zulu. In African markets, interaction AI has particular strategic importance because it can bridge literacy gaps, language barriers, and the dominance of feature phones and USSD channels.
Prediction opportunities use historical patterns to forecast future outcomes with enough accuracy to improve decisions. The value is better decision-making under uncertainty. Examples: credit default prediction, customer churn prediction, market demand forecasting. This is the most mature category of applied ML and has well-established evaluation frameworks, but it is also the category most sensitive to data quality and distribution shift.
Most real AI products combine more than one type, but it is useful to identify the primary type for a given opportunity, because it determines what the minimum viable data requirement is and what success looks like.
Evaluating AI opportunities in African markets
African markets introduce a specific set of constraints and opportunities that change the calculus of AI product strategy in important ways. A framework that works for building AI products in the US or Europe may produce entirely wrong answers when applied to Nigeria, Kenya, Tanzania, or Ghana.
Connectivity and latency constraints. A large fraction of users in Sub-Saharan Africa access digital services on 2G or 3G networks with intermittent connectivity. This has direct implications for AI architecture: large model inference must either happen server-side with results cached, or lightweight models must run on-device. Streaming LLM responses — which depend on a persistent connection — work poorly on intermittent networks. Any AI strategy that assumes reliable broadband is building for the wrong user.
Channel diversity. In most African markets, USSD and SMS remain primary channels for financial services interactions. WhatsApp is dominant for messaging. Voice calls are more common than app-based interactions for certain user segments. This means that the UX surface area for AI is fundamentally different from Western markets. An AI feature designed exclusively for a smartphone app is inaccessible to a significant fraction of the target population. The best AI products in this context are multi-channel by design, not as an afterthought.
Language and literacy. Africa has over 2,000 languages, and the majority of formal digital products are built in English, French, or Portuguese — colonial languages that are second or third languages for most users. This creates both a problem and an opportunity. The problem is that AI systems trained primarily on English-language data perform significantly worse in African language contexts. The opportunity is that teams who invest in low-resource language capability — through fine-tuning, through data collection partnerships, or through working with specialist vendors — can create defensible moats that larger international players cannot easily replicate. We will explore this in depth in Session 3.
Regulatory environment. AI regulation in Africa is developing but uneven. Kenya, South Africa, Nigeria, and Egypt are the furthest advanced in articulating AI governance frameworks, but most markets still rely on sector-specific regulation (banking, telecoms, data protection) to govern AI applications. This creates both more freedom and more uncertainty than in heavily regulated markets like the EU. Teams operating across multiple African markets need to track regulatory developments actively and build for compliance headroom rather than minimum compliance.
The opportunity side. These constraints are real, but they co-exist with structural opportunities that are, in several respects, larger than those in more mature markets. Credit penetration is low, meaning the total addressable market for AI-powered credit decisioning is enormous. Formal employment data is sparse, creating high demand for alternative data-based underwriting. Mobile money penetration is high in East and West Africa, providing rich transaction data. Agent networks provide physical touchpoints that can be augmented with AI. The path from no formal credit history to AI-powered credit access is shorter in Nairobi than in New York, precisely because there is less legacy infrastructure to displace.
The product strategy framework
A good AI product strategy can be captured in five questions. The discipline is answering them in order, not skipping ahead.
Question 1: What is the specific problem? Not 'we want to improve customer experience' or 'we want to use AI for credit'. A specific problem has an identifiable user, a concrete failure mode, and a measurable cost. 'Our loan officers spend 40% of their time manually reviewing mobile money statements that could be automatically summarised, causing an average application processing time of 5 days when the market standard is 48 hours' is a specific problem. 'We want to be more efficient' is not.
Question 2: Why is AI the right solution? This question has a right answer (AI is uniquely suited to this problem because…) and a wrong answer (we want to use AI). AI is right when the problem involves processing unstructured data at scale (text, voice, images), when the decision space is too large for rules, when the system needs to learn from new patterns over time, or when the value is in generating novel outputs rather than executing defined logic. AI is wrong when the problem can be solved with a well-designed database query, a straightforward rule engine, or better process design. Using AI for a problem that doesn't require it adds cost, latency, and failure modes without adding value.
Question 3: What data exists to make this work? Every AI system requires input data to function and historical data to evaluate. Before committing to an AI approach, you need to know: what data will the system use at inference time (and is it reliably available, in the right format, at the right latency)? What historical data exists to train, fine-tune, or evaluate the system? Who owns that data, and what permissions exist to use it? In African market contexts, this question frequently surfaces blockers that require the strategy to pivot before any model selection happens.
Question 4: What does success look like, specifically? An AI product metric has three components: the thing being measured (accuracy, throughput, cost, user satisfaction), the threshold that constitutes success (95% accuracy, 10× throughput, 70% cost reduction, NPS above 40), and the time horizon over which it will be evaluated (after 1,000 inferences, after 90 days in production, after the first full market rollout). Without all three, 'success' is a feeling, not a standard. Two people on the same team can sincerely believe the product is succeeding and failing simultaneously, because they have different implicit thresholds in mind.
Question 5: What is the most likely reason this fails? Every AI product has a primary failure hypothesis — the one thing that, if it goes wrong, makes the whole thing unviable. Naming it explicitly forces the team to either mitigate it early or accept the risk consciously. Common failure hypotheses: the data does not exist at the quality required; user trust is too low for AI recommendations to be acted on; the model performs well on historical data but degrades on new data; the regulatory environment changes before we can recoup the investment; the unit economics only work at scale we cannot reach in the first 18 months.
This five-question framework is the foundation of the session deliverable. It is also the framework you should use whenever someone in your organisation proposes a new AI initiative — not as a gate, but as a forcing function for clarity.
Prioritisation and sequencing
If you run the five-question framework across multiple AI opportunities, you will typically surface more viable ideas than you can pursue simultaneously. Prioritisation is then required.
The most useful prioritisation matrix for AI opportunities in emerging markets has two axes: strategic value (how much does this move the business metric that matters most right now?) and data and technical feasibility (how ready is the organisation to actually build this, given existing data assets and team capability?).
Opportunities that score high on both dimensions are your immediate priorities. Opportunities that score high on strategic value but low on feasibility are your investment targets — the things worth building data infrastructure or team capability to unlock. Opportunities that score low on strategic value regardless of feasibility should be deprioritised, even if they are technically interesting.
A common mistake is to prioritise based on technical feasibility alone, which produces a portfolio of AI features that are easy to build but strategically marginal. This is the path to the 'AI feature' that the team is proud of but users do not notice.
Sequencing matters as much as prioritisation. Some AI capabilities are foundational — they produce value on their own and also unlock future capabilities. An example: building robust mobile money transaction data parsing is foundational to automated income estimation, which is foundational to thin-file credit decisioning, which is foundational to personalised credit limit management. Building the foundational layer first creates compounding value; building the downstream feature first creates a dependency that is difficult to retrofit.
The sequencing principle for African fintech AI specifically: trust before automation, coverage before optimisation. Users and regulators in African markets are often appropriately sceptical of automated systems that affect access to financial services. Building trust through transparency, human oversight, and demonstrated accuracy in lower-stakes applications before deploying fully automated high-stakes decisions is not just good ethics — it is good product strategy.
Worked example: Credable loan pre-screening
Let us apply the five-question framework to a concrete fintech context: a multi-market embedded lending platform operating across East and West Africa, evaluating whether to build an AI-powered automated loan pre-screening feature.
Problem: Loan officers spend an average of 45 minutes per application manually reviewing mobile money statements (M-Pesa, Airtel Money) and bank statements to assess basic income and repayment capacity. This creates a processing bottleneck that limits the volume of applications that can be reviewed, extends time-to-decision to 3–5 days, and introduces significant inconsistency across loan officers with different experience levels. Applications from thin-file customers with no formal employment records are disproportionately declined without full review because officers lack the time and frameworks to evaluate non-standard income signals.
Why AI: The pre-screening task involves processing semi-structured financial documents (mobile money PDFs, bank statements in multiple formats) to extract income signals, identify recurrence patterns, and flag anomalies. This is precisely the kind of unstructured-data-at-scale problem that LLMs and structured extraction pipelines handle well. A rule-based approach would require exhaustive formatting rules for every possible statement format across every market and bank — a maintenance nightmare. An AI approach can generalise across formats and learn from new patterns as they emerge.
Data available: Three years of historical applications with associated mobile money statements and outcomes (approved/declined/defaulted) across two markets. Approximately 85,000 processed applications. Statements are stored as PDFs in an S3 bucket. PII has not been stripped, which creates a compliance workstream. Consent for data use in model development was not explicitly captured in earlier cohorts and needs legal review.
Success definition: Pre-screening accuracy within 5% of loan officer accuracy on a held-out validation set of 500 applications; processing time from document submission to pre-screening output under 90 seconds; loan officer agreement rate (proportion of pre-screening recommendations they accept without modification) above 75%; measured within 60 days of production deployment on a randomly selected 30% traffic slice.
Primary failure hypothesis: Statement format diversity is higher than anticipated, particularly for informal mobile money printouts and WhatsApp-forwarded statement screenshots, causing extraction accuracy to degrade on the long tail of formats. Mitigation: build a format confidence score into the system; route low-confidence extractions to human review rather than generating a pre-screening output.
This worked example becomes the basis for Session 1's deliverable. The pattern — problem → why AI → data → success → failure hypothesis — should become habitual before building any AI product.
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.
- 1Name the three strategic failure pathways for AI products and give one example of each from an African market context.
- 2What distinguishes an augmentation AI opportunity from an automation AI opportunity? When would you choose one over the other?
- 3A colleague proposes building an AI-powered credit scoring model for a market where you have no historical loan data. Walk them through the five-question framework.
- 4What does 'trust before automation' mean as a product sequencing principle, and why is it particularly important in African fintech?
Session deliverable
AI Product Brief
Complete a structured one-page AI product brief for a real or hypothetical AI feature you are considering. The brief forces precision on problem, solution, data, success, and failure — the five questions that determine whether an AI project is worth starting.
Show worked example ↓
Mobile money statement pre-screening engine
Show worked example ↓
Loan officers spend 45 min/application reviewing M-Pesa and Airtel Money PDFs to extract income signals. This caps application throughput at ~12/day per officer, extends time-to-decision to 3–5 days, and causes inconsistent treatment of thin-file applicants whose informal income is present in transaction data but not evaluated systematically.
Show worked example ↓
Statement formats vary across 6+ mobile money providers and 3 markets. A rule-based parser would require exhaustive format mappings and break on edge cases. LLM-based extraction generalises across formats and can reason about income patterns (salary recurrence, business cash flow, remittance streams) in a way that lookup tables cannot.
Show worked example ↓
85,000 historical applications with associated PDFs and outcomes across KE/TZ. Stored in S3. PII not stripped — compliance workstream needed. Consent for model development not captured in cohorts pre-2022 — legal review required before use.
Show worked example ↓
Pre-screening accuracy within 5% of loan officer decisions on 500-application validation set; processing time <90 seconds; loan officer acceptance rate >75%. Measured 60 days post-deployment on 30% traffic slice.
Show worked example ↓
Statement format diversity (especially WhatsApp screenshots, handwritten printouts) causes extraction accuracy to degrade on the long tail. Mitigation: build format confidence scoring and route low-confidence cases to human review rather than generating a pre-screening output.
Show worked example ↓
Automation — replacing a manual task