Capital often fails because its terms, timing and risk model do not fit the business.
A research proposition
The
Context
Tax
How AI can hallucinate infrastructure in African markets without inventing a single fact.
Nairobi, Kenya / Photo: Marie Frank
Imagine asking an AI system to design a market-entry strategy for a digital health company expanding across Africa.
It recommends a smartphone app, an AI chatbot, prepaid subscriptions, automated onboarding and one playbook for several countries.
Each recommendation has worked somewhere. The answer may contain no false citation, invented statistic or impossible technology.
It may still fail.
In my research for Founders Factory Africa, one healthtech operator described a less tidy system. Reaching patients meant using an app, a website, a chatbot, SMS and WhatsApp.
Delivering care also required relationships with clinics, pharmacies, laboratories, doctors, diagnostic centres and telecom companies.
Some patients did not own smartphones. Some markets had no clear telemedicine framework. Others required the company to operate through licensed clinical facilities.
The useful product was not simply software. It was software combined with distribution, trust, regulation, payment design and offline delivery.
The hypothetical AI plan did not have to hallucinate a fact. It hallucinated the surrounding system.
The context tax is the loss in usefulness, feasibility or economic value when advice lacks the institutional, cultural, market and infrastructure context needed to act on it.
Context is not decoration
Data is only one layer.
The World Bank frames AI readiness through four foundations. Its definition of context focuses on locally relevant data and content.
That is essential. Field evidence suggests an operational extension.
01Connectivity
Energy, devices and digital infrastructure.
02Compute
Chips, data centres and cloud access.
03Context
Local data, content and model adaptation.
04Competency
Skills to adopt, adapt and innovate.
Observed in the research
Three forms of context tax
Founders Factory Africa was a client of mine, later known as 54 Collective. Across two research projects, I interviewed more than 100 people connected to African business and technology ecosystems.
The participants ranged from micro and small business owners to founders, regulators, support organisations, institutional investors and European funds of funds focused on Africa.
Founders privately build infrastructure that other markets treat as a shared service.
Training is abundant while procurement, working capital and market access remain scarce.
Local context changes the business model, not simply the language used to sell it.
Trust and regulatory relationships behave like operating infrastructure.
Tax 01 / Infrastructure
The company has to build the market around the product.
A software plan can treat payments, logistics, identity, connectivity and regulation as external services.
In the research, founders often had to assemble those functions themselves. Some needed power and connectivity before they could build the product investors thought they were funding.
What looks like inefficiency can be a private payment for missing public or market infrastructure.
Software plan→Missing rails→Different capital and pace
Tax 02 / Intervention mismatch
A better score does not create a buyer.
Generic agricultural finance advice might focus on alternative credit scoring. That addresses information, but not necessarily the economic system.
One operator financed smallholders after securing an off-taker, a buyer for their produce. Prompt payment at delivery was also critical.
When procurement moved slowly, invoice finance filled the gap. The viable product connected credit, demand, procurement and cash flow.
Credit score→Guaranteed demand→Cash at the right time
Tax 03 / Market flattening
Reusable software does not mean a reusable market.
An investor described fragmentation across borders, geography, purchasing power, legal regimes and languages.
A product may travel. Its customer journey, regulatory relationships, payment behaviour and distribution network may not.
Population size is not an integrated addressable market. Every new country can add another layer of translation.
One product→Many institutions→Country-specific entry
System map
Where the tax enters the decision
The model can be capable and the recommendation plausible while the assumed surrounding system is wrong.
Capable modelBroad knowledge and reasoning
Plausible taskMarket, policy or investment advice
infrastructure
institutions
trust
capital
regulation
market structure
The context taxPlausible advice, wrong operating system
Adapted adviceOr a better question before action
Operational value ≈ model capability × context quality × institutional fit. Conceptual relationship, not a measured equation.
A conceptual relationship
Operational value ≈ model capability × context quality × institutional fit
This is not a measured equation. It is a warning that a more capable model can generate a more detailed version of the wrong plan.
The constructive proposal
A proprietary context layer
The answer is not to upload confidential transcripts into a chatbot.
Selected findings could become small, permission-safe evidence cards. At query time, a system would retrieve only the cards relevant to a specific country, sector and decision.
This resembles retrieval-augmented generation. The model is not retrained. Relevant evidence is placed into its working context before it responds.
The key design choice is restraint. A card should record where an observation applies, how confident we are and where it should stop travelling.
- Context
- Smallholder finance / represented East African market
- Perspective
- Operator financing agricultural production
- Mechanism
- Finance becomes viable when demand and prompt payment are secured
- Implication
- Do not recommend credit scoring as a complete intervention
- Confidence
- High within the observed case, not region-wide
- Boundary
- Validate crop cycle, buyer structure and payment timing locally
Evidence card / specimenEC-AG-014
Beyond a product feature
The policy question inside the product question
Founders are not the only ones who pay. Public institutions, investors and citizens also absorb the cost of advice that misreads the system around the technology.
Capacity
Local AI capacity should include institutional knowledge, not only compute, datasets and technical talent.
Evaluation
Measure decision quality, not fluency. A polished answer that assumes missing infrastructure should count as failure.
Value
Researchers and operators supplying the context that makes AI useful should share in the value it creates.
A follow-up, not evidence already claimed
From argument to test
The context tax is a research proposition. This piece does not claim that a frontier model has already failed a controlled African market evaluation.
A later study could compare the same market scenario under three conditions.
Scenario only
A realistic prompt with no added country or sector evidence.
Public dossier
The same prompt plus credible public country and sector sources.
Local evidence layer
Public sources plus anonymised evidence cards from field research.
African practitioners should grade feasibility, institutional awareness, uncertainty and actionability. A null result should remain publishable.
Model intelligence and situated understanding are not the same thing.
Some AI mistakes will not look like fabricated facts. They will look like elegant plans for systems that do not exist. The question is who notices, and who pays the tax.