This is the capstone session. Everything from sessions 1 through 8 — the product thesis, the mental models, the data strategy, the model selection, the architecture, the UX, the safety and security, the MLOps — connects here, to the question that ultimately determines whether your AI product survives: do the economics work?
An AI product that does not have viable unit economics is not a product — it is a pilot. This session gives you the tools to build the economic model for your AI product, integrate it into a credible product roadmap, and present it as a fundable narrative. The deliverable is a capstone pitch that synthesises the work of all nine sessions.
Unit economics for AI products
Unit economics for an AI product measures the revenue and cost attributable to a single unit of the product — typically a single inference, a single application processed, or a single active user. Unlike traditional software where marginal cost approaches zero at scale, AI products have meaningful marginal costs that must be explicitly modelled.
The unit economics equation for a lending AI feature is:
Revenue per unit = (revenue generated per loan × AI-attributable approval rate improvement) or (cost savings per manual process replaced × automation rate)
Cost per unit = inference cost + data storage cost + human oversight cost (for flagged cases) + model maintenance cost (amortised over volume) + infrastructure cost
The contribution margin per unit is revenue minus cost. The product is economically viable when contribution margin per unit is positive, and strategically valuable when the total contribution margin across volume is large enough to justify the fixed investment.
Common unit economics errors:
Ignoring human oversight cost. Most AI systems route some percentage of cases to human review. If 20% of applications are flagged for human review, and each human review takes 20 minutes, the human cost per 100 applications is 400 minutes of loan officer time — a significant cost that must be included in the unit economics.
Omitting model maintenance cost. Models require ongoing evaluation, retraining, and iteration. The engineering cost of this should be amortised over the volume processed and included in the per-unit cost. A team that spends 2 engineer-months per quarter on model maintenance is spending approximately $30,000 per quarter (at loaded cost). At 10,000 applications per quarter, that is $3 per application in model maintenance cost — material for a micro-loan product.
Misattributing revenue. When AI is one component of a larger product, attributing all revenue improvement to the AI feature is optimistic. Use a counterfactual approach: what would revenue be without the AI feature, using the best available non-AI alternative? The delta is the AI-attributable revenue.
AI feature roadmap sequencing
An AI product roadmap answers two questions: what will be built, and in what order? The 'what' follows from the strategy work in session 1. The 'in what order' requires understanding dependencies, compounding value, and risk sequencing.
Dependency mapping. Some AI features are prerequisites for others. Data collection is a prerequisite for model training. Model training is a prerequisite for automated decision-making. Automated decision-making is a prerequisite for personalisation. Mapping these dependencies prevents the common mistake of attempting to build a sophisticated capability on a foundation that does not yet exist.
Compounding value. Some AI investments compound — they create value on their own and also make future investments more valuable. A high-quality customer data platform is valuable for current ML models and also reduces the cost of training future models. An evaluation dataset built for one model can be reused (with modification) for successor models. Prioritise investments that compound over those that are point solutions.
Risk sequencing. Build proof of value before building for scale. The first AI feature a team deploys is a learning experience as much as a product release — it will surface unexpected challenges in data quality, model performance, user behaviour, and operational complexity that inform all subsequent AI development. Starting with the highest-risk, highest-complexity AI capability is a mistake; starting with the most tractable problem that has credible business value is the right approach.
A practical roadmap structure for a lending platform AI product: Phase 1 (0–6 months): document extraction and structured data pipeline — the foundational data layer; Phase 2 (6–12 months): automated pre-screening with human-in-the-loop — value demonstration; Phase 3 (12–18 months): thin-file alternative data model — market expansion; Phase 4 (18–24 months): dynamic credit limit management and personalisation — optimisation.
AI as competitive moat vs. commodity
Not every AI capability is a moat. As frontier models become cheaper and more capable, and as AI tooling matures, the 'we use AI' positioning that differentiated products in 2022 is becoming table stakes in 2025. The question for any AI product is: what specifically about our AI capability is defensible?
Defensible AI moats in fintech come from one or more of four sources:
Proprietary data. A model trained on data that competitors cannot access — your customers' transaction history, your markets' specific default patterns, your borrowers' mobile money behaviour across years — is genuinely differentiated. The model is only as unique as the data. This is the most durable moat.
Distribution advantage. In African markets, reaching customers through established MNO relationships, agent networks, or embedded banking partnerships creates a data collection advantage over a challenger that is building distribution from scratch. More distribution → more data → better models → more value → stronger distribution. This is a compounding loop.
Operational excellence in model management. The ability to iterate models quickly, deploy them safely, and learn from production is a competitive capability that goes beyond the model itself. An organisation with strong MLOps can outperform a competitor with a nominally better initial model because they improve faster.
Regulatory trust. In regulated sectors in African markets, demonstrating to regulators that your AI systems are explainable, fair, and governed appropriately creates a licence to operate that competitors who have not invested in governance cannot easily replicate. Regulatory trust is becoming a moat as regulators become more active in AI oversight.
The commodity AI trap is building capabilities that any team can replicate with an API key and a good prompt. API-integrated chatbots, standard RAG implementations, off-the-shelf OCR — these add value but do not create moats. The strategically valuable investment is in the data, the relationships, and the operational capabilities that underlie the AI, not the AI layer itself.
Building a fundable AI product narrative
Fundraising for an AI product in an African market context requires a narrative that connects three things: the market opportunity (why this problem is large and addressable now), the AI advantage (why AI creates a solution that is categorically better than alternatives and specifically defensible), and the execution credibility (why this team can build and operate AI in this context).
Market opportunity framing. Do not start with the AI. Start with the problem. The size and clarity of the problem determines the total addressable market. 'Only 18% of Sub-Saharan African adults have access to formal credit' is a compelling problem statement. 'We are building an AI credit scoring engine' is a solution statement that most investors cannot evaluate without first understanding the problem.
AI advantage framing. The AI advantage should be stated in terms of outcomes, not technology. 'Our AI model allows us to lend profitably to 40% more customers than a bureau-based approach, as demonstrated in a 6-month pilot in Kenya with 8,200 applicants' is credible. 'We use machine learning to assess creditworthiness' is not differentiated. The unit economics model is the evidence behind this claim — be ready to walk through it.
Execution credibility. Investors in African market AI products are appropriately sceptical of AI claims that cannot be substantiated with production data. The most credible narratives include: specific evaluation results from production deployment (not demos), evidence of data infrastructure that supports ongoing model improvement, and a team with demonstrated experience operating AI systems in the specific market context. If you have not yet deployed, be explicit about the pilot plan and the specific hypotheses you are testing.
Risk honesty. Sophisticated investors know that AI products have risks. Acknowledging the primary risk hypothesis (data readiness, regulatory evolution, model performance on edge cases) and explaining your mitigation strategy is more credible than pretending these risks do not exist. An investor who discovers a significant risk after investment that was not disclosed — and could have been — is a problem; an investor who funds you knowing the risks and believing in your mitigation is a partner.
Capstone: integrating the nine sessions
This final section is not new content — it is an integration map of how the nine sessions fit together as a system.
Session 1 (Product Strategy) defines the problem, the AI thesis, and the success criteria. Everything else serves this foundation.
Session 2 (Mental Models) gives you the conceptual vocabulary to evaluate technical decisions without requiring engineering depth. It is the translator between strategy and architecture.
Session 3 (Data Strategy) determines what is actually possible, given the data assets available. The data strategy either validates or constrains the product thesis from Session 1.
Session 4 (Model Selection) determines how the AI capability will be built, at what cost, and with what dependencies. It operationalises the data strategy into a concrete build decision.
Session 5 (Architecture) designs the system that wraps the model — the pipeline, the orchestration, the fallbacks, the monitoring hooks. It is the blueprint that engineering implements.
Session 6 (UX Design) determines how users interact with the system's outputs — whether they trust them, act on them, and return. Even a technically excellent AI system fails if users do not understand or trust it.
Session 7 (Safety and Security) identifies the ways the system can be misused, can cause harm, or can fail in ways with regulatory consequences — and builds the governance structures to prevent and respond to those failures.
Session 8 (MLOps) ensures the system remains reliable, fair, and performant after deployment — the ongoing operational discipline that separates a production system from a pilot.
Session 9 (Economics and Strategy) connects all of the above to the commercial case — the unit economics, the roadmap, the moat, and the narrative that attracts capital and partners.
The capstone deliverable asks you to synthesise all of this into a 5-minute pitch. Five minutes is the discipline — it forces you to identify what is essential and discard what is peripheral. If you cannot explain your AI product's problem, solution, data advantage, architecture approach, safety posture, operating model, and economics in five minutes, the product is not yet strategically clear.
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.
- 1Walk through the unit economics of an AI-powered loan pre-screening system at 50,000 applications per month. What are the revenue-side and cost-side components?
- 2What is the 'commodity AI trap' and how would you assess whether your AI capability is defensible?
- 3A Series A investor asks: 'What is the single biggest risk to your AI product, and how are you mitigating it?' What would you say for your product?
- 4Draw the dependency map between the nine sessions — which sessions are prerequisites for which? What does this tell you about the sequencing of AI product development?
Session deliverable
Capstone AI Product Pitch
Build a 5-minute pitch for your AI product, integrating all nine sessions. Cover the problem, the AI thesis, the data advantage, the architecture approach, the safety posture, the economics, and the 90-day plan. This is the capstone deliverable for the entire programme.
Show worked example ↓
82% of adults in Sub-Saharan Africa lack access to formal credit because they have no credit bureau history. Traditional lenders require formal employment records and bureau data — requirements that structurally exclude informal sector workers, market traders, and smallholder farmers who represent the largest economic segment in every African market. In Kenya alone, this represents 28 million creditworthy adults who cannot access the credit they need to invest in their businesses and smooth income volatility.
Show worked example ↓
We can assess creditworthiness for thin-file customers using mobile money transaction data — three years of M-Pesa, Airtel Money, and MTN MoMo transaction records demonstrate income, cash flow patterns, and financial discipline as reliably as a credit bureau record, for the 80% of adults who have a mobile money account. Our proprietary advantage: 85,000 labelled outcomes paired with transaction data across Kenya and Tanzania — a training asset that cannot be replicated without three years of lending operations in these markets.
Show worked example ↓
Core system: LLM-based structured extraction pipeline that parses mobile money PDFs into standardised income features, feeding a gradient-boosted credit decision model. RAG system for policy compliance — decision rationale grounded in current product policy. Fine-tuned model for East African market; separate model version under development for West Africa. All AI decisions logged with full audit trail. Human-in-the-loop for low-confidence cases (< 20% of volume).
Show worked example ↓
Cost per application: $0.085 (LLM inference $0.015 + human review share $0.04 + infra/ML ops $0.03). Revenue attributable to AI: +$1.20 per application through improved approval accuracy (12% lift in approval rate, 8% reduction in default rate on approved applications). Contribution margin: +$1.12 per application. Break-even volume: 45,000 applications/month against fixed model development cost. Current run rate: 28,000/month. Break-even in month 7 at current growth trajectory.
Show worked example ↓
Top risks: (1) bias against informal sector customers — mitigated by quarterly disaggregated accuracy audit, mandatory exploration budget for thin-file segment; (2) regulatory data permission exposure — mitigated by explicit ML consent in revised application flow, PII anonymisation pipeline deployed Q1. Regulatory compliance matrix maintained for KE, TZ, NG — legal review scheduled quarterly.
Show worked example ↓
Phase 1 (current): document extraction + pre-screening, Kenya/Tanzania. Phase 2 (Q3): thin-file model expansion — target 40% approval rate lift for informal sector; Phase 3 (Q4/Q1): Nigeria market entry with West Africa model training; Phase 4 (18 months): dynamic credit limit management — personalised limit increases based on repayment behaviour. Dependency logic: each phase requires the data and model infrastructure of the prior phase.
Show worked example ↓
Day 90 target: production deployment of v2 extraction pipeline with format confidence scoring, reducing manual review queue from 28% to < 20% of applications. Three priorities: (1) complete PII anonymisation pipeline and legal consent review — unblocks use of full historical dataset for model retraining; (2) build labelled evaluation dataset to 500 examples with disaggregated informal-sector subsample; (3) deploy v2 extraction pipeline with Mistral fine-tune to canary (10% traffic). Success at day 90: pre-screening accuracy ≥88% on evaluation set, manual review rate < 20%, data compliance workstream complete.