Nubank, based in Brazil, is the largest online bank in the world. This example demonstrates that developing markets can build global-scale financial technology companies, not just adopt technology from developed countries.

Introduction of cell phones to Kerala fishermen created market convergence where all fishermen converged on a single market price through text messaging about prices, leading to more efficient markets, expanded operations, and better supply chains for restaurants—demonstrating that simple information availability can have powerful economic effects.

The AI advisor provided phenomenally good answers to entrepreneur questions, including detailed responses on dairy cow purchase budgets (covering costs, veterinary care, land maintenance, and risks from disease and price volatility) that Brian from Kenya verified as accurate for current prices, and strategic advice on competition that includes Porter's Five Forces-style differentiation analysis with specific tactical options (loyalty programs, community involvement, menu differentiation).

The study's control group received a PDF 'course note' similar to what HBS distributes to students. This is an appropriate placebo because empirically, when HBS distributes course notes, students do not read them—they read cases (because they know they will be cold-called on cases), and they ignore module notes and PDFs.

Sarvam AI, based in India, is building generative AI tools for all 22 scheduled languages of India. Rem predicts: 'In two years, for any language with over 500,000 speakers, we're going to have basically perfect generative AI for all those languages.' Recent cost reductions (10x drop for Japanese) suggest private market incentives will solve many language problems.

Mino Health.ai, funded by the Gates Foundation, is building foundation models for health on the African continent, starting with radiology and expanding to help doctors make better decisions and take better medical notes. This represents applying AI to healthcare in emerging markets with adaptation to local context.

There is a 'peak bubble' in AI investment and attention. Rem received an email offering $20 million for generative AI research funding, and jokes that it might be from a scammer. He states: 'Definitely a peak bubble. Like let me just say that.'

The 10% decrease in revenue/profits for low-performing entrepreneurs who received AI advice (vs. control) was surprising and unexpected. Rem states: 'all that work and nothing?' and describes the team hitting 'the accelerator' in May 2023 and working nonstop until November before discovering this negative result.

The study did not find meaningful demographic differences between low and high performers across industries, gender, or age. The difference between groups was not demographic but behavioral and problem-focused.

Data drift has occurred since the AI model's training: prices for Friesian dairy cows have changed following Kenya's currency devaluation, meaning the advice given reflects outdated price information. This is a technical limitation of static models applied to dynamic emerging market contexts.

Developing the prompt that powers the business advisor required 3 months of iterative testing with entrepreneurs and local experts (including RMI lab member Brian from Kenya) to refine tone, vocabulary (removing Harvard jargon), and contextual appropriateness. The prompt works by instructing the AI to 'use simple, non-technical English' and 'avoid discussing health, religion, politics, or current events.' This demonstrates that effective AI tool design for emerging markets is fundamentally a writing/humanistic problem, not a coding problem.

There is potential for the AI tool to improve significantly beyond its current V1 version by adding contextual data: GPS location information to provide regionally tailored advice, integration with Shopify or other platforms to provide store-specific data (products, sales numbers), or integration with health records and doctor data to provide medically contextualized advice. However, this raises inequality concerns because data availability varies dramatically between developed and developing markets.

In the study, entrepreneurs asked high-quality, strategic business questions of the sort that would place in the top decile at Harvard Business School, such as: 'I have a fast food joint adjacent to a matatu terminus with stiff competition—how can I overcome this?' and 'I wash cars and motorbikes alone and find it tiresome; how do I hire staff and monitor them to maximize profit?' These questions demonstrate that entrepreneurs in emerging markets have genuine, sophisticated business challenges, not trivial ones.

A fifth open question is whether effects of AI mediated through humans (AI-to-human-to-human loops) can persist. Rem gives the example: teachers use AI to improve their teaching, which then improves student outcomes. Can impact chain through this human intermediary, and does it compound or diminish?

In a randomized controlled trial of 640 Kenyan entrepreneurs (311 control, 306 treatment), providing access to a GPT-4-powered WhatsApp business advisor had no statistically significant effect on revenue and profit in the full sample (approximately 5% increase but within error bars), contrary to expectations.

The divergent effects between high and low performers stem from applying AI to fundamentally different problem types: low performers focus on problems of competition and survival (e.g., 'my business is facing drastic change and incurring losses'), while high performers focus on problems of growth (e.g., 'I have capital to expand to a new location'). The AI provides useful tactical advice for growth problems but provides unhelpful generic advice (like 'go get financing') for survival problems, causing low performers to become discouraged and make fewer business changes than the control group.

Many entrepreneurs in the study do not have alternative sources of trusted business advice. When surveyed, entrepreneurs say they either live in suburban Nairobi or rural areas, don't trust people around them, or feel embarrassed to ask questions. This explains willingness to use an AI advisor: it provides privacy and removes social risk from asking what they perceive as naive questions.

GPT-4 was trained on massive internet data including World Bank reports, HBR articles, newspaper articles from Kenya, and other global sources. This training data means the model has surprisingly good knowledge of emerging market contexts like Kenya, knowing terms like 'matatu terminus' (bus terminus) and being able to answer questions about Kenyan dairy prices, though with some data drift since training.

Women are significantly less likely to use ChatGPT and other generative AI tools compared to men, across multiple countries and data sources. This gap persists even when controlling for profession—software developers who are women are 20-50% less likely to use AI tools than male software developers with the same job title in Denmark. This suggests a barrier beyond professional necessity.

Entrepreneurs in the treatment group showed high engagement with the AI advisor over WhatsApp: some explicitly requested more reminder text messages because they forgot the tool was available ('Could you send me more texts cuz I forget to talk to the AI? It gets pushed down on my WhatsApp'), and approximately a dozen users sent 'happy new year' messages to their chatbot at the turn of 2024 ('It has been amazing working with you').