MYPE AsesorIA

In lower and middle income countries, a significant proportion of the workforce is self-employed (Gindling & Newhouse 2014), and these Micro, Small and Medium-sized Enterprises (MSMEs) suffer persistently low productivity. There are well-defined business and management practices that, when adopted, cause firms to be more productive and grow faster (Bloom and van Reenen 2010, McKenzie and Woodruff 2017). But firms fail to adopt them. 

For this reason, governments and aid agencies spend at least $1b to train 4-5 million MSME entrepreneurs per year (McKenzie 2020). Traditional classroom-based training increases their profits by 10% on average, while approaches incorporating psychological focus and heuristics increase profits by 15% (McKenzie 2020), and impacts of consulting services are even larger (McKenzie & Woodruff 2023). These are direct, sustainable increases to incomes for the self-employed. For these reasons, two of Innovation for Poverty Action’s 14 “Best Bets for Impact at Scale” (IPA 2023) are soft-skills training to boost business profits and sales and consulting services to support SMEs.

Yet 4-5 million MSME entrepreneurs is barely 1% of the total number of MSMEs who could benefit from such training. This limited scale is due to costs of scaling in-person, human-administered training, typically $150-$1500 per participant (Woodruff 2020). As noted by both IPA (2023) and Woodruff (2020), the most critical need to unlock this best bet is “further experimentation with alternative delivery methods”.

MYPE AsesorIA empowers owners with customized, interactive, ongoing support leveraging LLMs. It is a contact in WhatsApp that they can ask questions to (via text, voice memo, or images) any time, and which periodically sends them tips and special exercises, based on their own history and common questions from other users.

The natural-language interface is easier to use and allows for a much richer understanding of and customization to the particular entrepreneur and context compared to other technological solutions to training like structured bots or online courses. This unlocks situated, experiential learning, and uses a front-end platform that owners are already familiar with (WhatsApp). The tool can rapidly draw from a library of existing entrepreneurial training and materials, retrieving information that is most relevant to the entrepreneur’s particular situation and priorities at that moment, and allows users to interrogate that knowledge. This approach is 1/100th to 1/1000th the cost of the status quo, and highly scalable.

Moreover, because of this cost advantage, training and mentorship can be continuous and open-ended, using active learning and delivering the most relevant information for the problem at hand over time rather than trying to transmit the entire corpus of training materials in a limited session. This interactive nature can assist directly, step-by-step, with implementation of the knowledge, similar to a consultant. For this reason, beyond the massive cost and scale advantages, impacts could actually be larger than for traditional training.

Our technology stack is designed to be highly transparent and as easy to copy as possible, so that we can scale not only directly, but also indirectly by empowering other entities that reach MSMEs to copy and learn from our work. We use OpenAI models (GPT4, GPT4V, and Whisper), share all model instructions and RAG resources, and also provide a clone of our middleware system to all pilot partners we work with (pilot partners are suppliers, associations, and other entities that serve large numbers of MSMEs). 

Our solution serves the 99% of approximately 500 million MSMEs in developing countries that aren't reached by traditional methods of entrepreneurial training and consulting every year. Specifically, we are starting in our own community (Peru) and focusing on some MSME-intensive sectors to support queries with sector-specific training materials for RAG. Our first vertical is Bodegas (small-scale retail shops) and Cafetaleros (coffee growers). In Peru there are approximately 535,000 Bodegueros and 200,000 Cafetaleros. The largest entrepreneurial training program in the country reaches 8000 MSMEs per year. This is a massive gap that we can fill.

We have worked extensively on MSME productivity, including a number of studies in Peru with these particular sectors. These studies have included extensive field work and focus groups, which was part of the genesis for this project, and has provided a representative group that has been at the center of all aspects of solution design and beta testing from day 1. This proximity is crucial because we are essentially building an 'on-ramp' to LLMs for a community with limited technology comfort (with the exception of WhatsApp), and the key barrier to adoption is to what degree they see the tool adapting to and improving their daily priorities. Our users provide continuous input, not only feedback on tool features, but sharing their own ideas for new features which guide our product roadmap.

We began an invitation-only pilot in November 2003 of approximately 15  participants (bodega owners in Peru), and since then have expanded availability to an additional +/- 30 users (some of which are testing the tool on behalf of potential pilot partners). The tool is consistently available to them and used on an on-going basis.

We would most benefit from partners who can help us advance our technology and distribution. On the technology front, this includes expanding model capabilities, improving user experience, and further work on reliability. In terms of distribution, ideal partners are other organizations that work with MSME entrepreneurs who would benefit from AI-based training and support.

Our solution is innovative because it applies frontier AI models and their emerging capabilities (such as image analysis) to a solution that itself is already proven (MSME entreprneuerial training), but suffers from high costs and limited scale. In addition to solving the cost and scaling problem, LLMs should actually increase impact compared to traditional approaches, as they can offer training that is more consistent with how adults learn best (contextualized, experiential learning over time rather than up-front classroom-based). Unlocking this potential requires further work on the technology side, but even more, it requires work on user experience (UX) to make the solution accessible and impactful.

Our approach is both unique from others in this space and could catalyze broader impacts because we are building the leanest and most transparent approach possible that can effectively solve this problem, and actively sharing it broadly under a creative commons license.

The type of training, mentoring & consulting provided by our tool is already extensively studied, and proven to have an impact on MSME productivity ("Training Entrepreneurs", McKenzie & Woodruff 2023. "Best Bets: Emerging Opportunities for Impact at Scale", Innovations for Poverty Action 2023). Our theory of change is that this proven impact can be achieved in a lower-cost, more scalable way by LLMs that are correctly instructed, draw on reliable best-practice training materials, and built with a UX that is an effective 'on-ramp' for those with little comfort with technology and no experience with LLMs. 

We directly draw from the extensive literature on the impacts of entrepreneurial training referenced above, and will measure impacts on productivity and incomes using the same methods in McKenzie & Woodruff (2023) so that results are comparable to the broader literature. Our goal is to make those established impacts of traditional training larger, but most importantly, more broadly distributed beyond the 1% of MSMEs they currently reach each year. We supplement these longer-term financial indicators with short-term measures of users' perceived usefulness of the tool, measured indirectly through tool usage and measured directly through WhatsApp button-reply surveys.

Our core technology is OpenAI's GPT4 and GPT4V delivered through the WhatsApp Business Platform. These are connected by a low-cost intuitive WYSIWYG platform (built with make.com, formerly integromat), which manages the user experience and can be easily cloned and shared with partners.

The team currently dedicated to this solution is 1 full-time and 3 part-time

We have been working on MSME productivity in Peru for decades, across all dimensions including knowledge & capacity building, access to finance, and integration into value chains. We have been working on this particular tool since October 2023.

Our tool has benefitted from close and intensive work with the population we serve, beginning with small-scale Peruvian retailers. This was possible because our team has extensive experience working with these populations, as well as family in these activities directly.

Our product is always provided for free to end users, who we reach through partnerships with organizations that already work with MSME entrepreneurs, and give them an introduction to and access to our tool. Examples include major suppliers, co-operatives and trade associations, and government agencies. Our product from the point of view of those partners is both the tool itself, and through the partnership, an introduction to adapting AI tools for MSME entrepreneurs and learning from this experience to inform their own future initiatives. Funding can be provided by those partners, or by third parties such as donors. It is important to note that unlike others in this space, we are not building a large team and a highly engineered proprietary product. We are building a fully transparent system with a very lean team, provided at-cost to partners, who have the option to scale it themselves directly at any time. This makes our funding needs very small, pilot partners easy to find, and the scale-per-dollar substantial, particularly when taking into account indirect scaling through imitation. Finally, if we can obtain in-kind support from the technology companies that are our major cost drivers (WhatsApp & model inference) our costs virtually disappear. We have received early in-kind support from OpenAI. 

We are in advancing stages of multiple grant programs (some of which suggested SOLVE to us). We have agreements with pilot partners that are ready to launch when funding is secured (these first pilots are zero cost to pilot partners, but we will migrate that approach over time). Our major cost drivers (particularly model inference) are falling rapidly over time.