Blog

MAIA started in Peru, which remains our largest country, but one year ago we started working in Panama, in partnership with <a data-mce-href="https://ampyme.gob.pa/" href="https://ampyme.gob.pa/" target="_blank" rel="noopener">AMYPME </a>and the ILO. As in all countries where we operate, the first stage is a detailed proecss of adaptation with local experts. In the case of Panama, our initial focus was small-scale producers of coffee and vegetables, and the tool was adapted and validated with sector experts. With them we collected a series of key country-specific manuals, guides, and rules, ending with country-sector specialized prompts and knowledge bases for RAG. MAIA @AMPYME was born.After multiple rounds of testing, in the face of very positive user feedback the tool was made generally available. <a data-mce-href="https://www.instagram.com/reel/DP4L1NeDlZ3/" href="https://www.instagram.com/reel/DP4L1NeDlZ3/" target="_blank" rel="noopener">Here you can see Jackeline Munoz, Panama's minister of labor and workforce development, along with leadership from AMPYME and the ILO at their official launch of Maia</a> <figure id="figure-of710qks13mi7olx5l"><img id="of710qks13mi7olx5l" src="https://cdn.durable.co/blocks/3aPOnZBjZgu8MNUusiZVgKUcnXeuzusOp8wZP8ZkxLJqrhz4He5GYJDOuatGTpFU.jpg" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/3aPOnZBjZgu8MNUusiZVgKUcnXeuzusOp8wZP8ZkxLJqrhz4He5GYJDOuatGTpFU.jpg" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure> Panama has quickly grown to become MAIA's second largest country with thousands of users, and we are entring new rounds of validation and expansion with the ILO and AMPYME!

MAIA @AMPYME Panama

Here is the latest MAIA overview, all data as of mid-oct<figure id="figure-5b2bfb0ylmhdn7qol" class=""><img id="5b2bfb0ylmhdn7qol" src="https://cdn.durable.co/blocks/3achi9WTMClfHY6KtCzRmPi0XvHixpDeVKKD1GDEDrwNWQnjRLdWpMFTNvif61xi.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/3achi9WTMClfHY6KtCzRmPi0XvHixpDeVKKD1GDEDrwNWQnjRLdWpMFTNvif61xi.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-6prnpn70njmhdn9ylr" class=""><img id="6prnpn70njmhdn9ylr" src="https://cdn.durable.co/blocks/20AcUW0ag6AIPXvXY8nPE7tMnnFKsGApYsEK9zY2P3cG6nPMqyIZuTfYlN7sexIp.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/20AcUW0ag6AIPXvXY8nPE7tMnnFKsGApYsEK9zY2P3cG6nPMqyIZuTfYlN7sexIp.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-hvg4xh6tkdmhdnal63" class=""><img id="hvg4xh6tkdmhdnal63" src="https://cdn.durable.co/blocks/5Ytvm0C6nO2q1tXTJ7QoKmgAhOkgfmXhCamLgax9qX3AyVT6n09ZAisKmDTMUsap.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/5Ytvm0C6nO2q1tXTJ7QoKmgAhOkgfmXhCamLgax9qX3AyVT6n09ZAisKmDTMUsap.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-n86b8xmaxmhdnaypo" class=""><img id="n86b8xmaxmhdnaypo" src="https://cdn.durable.co/blocks/7DStsQ0OXEMMobEG5n9RttsfTe2UsOpHvclNWSEEsMnYlx8TYIl00ZxL5FjYISiD.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/7DStsQ0OXEMMobEG5n9RttsfTe2UsOpHvclNWSEEsMnYlx8TYIl00ZxL5FjYISiD.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-7upurcrhw7mhdnba1p" class=""><img id="7upurcrhw7mhdnba1p" src="https://cdn.durable.co/blocks/8EAKEWkQ6ztkDmWEDWfjWnbJPptqyehEZP7VIGAa7sC1ZgNrtLUAzL4Rq44ruPf9.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/8EAKEWkQ6ztkDmWEDWfjWnbJPptqyehEZP7VIGAa7sC1ZgNrtLUAzL4Rq44ruPf9.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-irmovvfuqomhdnbohd" class=""><img id="irmovvfuqomhdnbohd" src="https://cdn.durable.co/blocks/aIcKJKXZYTuqpVxlf0lZlMjtsAOpdCwCRJa8VQkvCKeOiDcXAe1bH3fAQ9eDTiHI.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/aIcKJKXZYTuqpVxlf0lZlMjtsAOpdCwCRJa8VQkvCKeOiDcXAe1bH3fAQ9eDTiHI.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-rh1zlere9hmhdnc4ra" class=""><img id="rh1zlere9hmhdnc4ra" src="https://cdn.durable.co/blocks/c0ozxOaJC3bv3tzyxJTnzwLGTgMWykmvJjpHUMGDcrFSCPSJz84clvpW9OxD7DDi.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/c0ozxOaJC3bv3tzyxJTnzwLGTgMWykmvJjpHUMGDcrFSCPSJz84clvpW9OxD7DDi.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-p8w8bfpcilmhdncefo" class=""><img id="p8w8bfpcilmhdncefo" src="https://cdn.durable.co/blocks/dxZIguzkjesNquf0snPsRFYzts6LX6DYpG5hwRJsNsUXiFkAc6RfUS8cSViUFUjM.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg selected" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/dxZIguzkjesNquf0snPsRFYzts6LX6DYpG5hwRJsNsUXiFkAc6RfUS8cSViUFUjM.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-d3iq6oioimhdncocx"><img id="d3iq6oioimhdncocx" src="https://cdn.durable.co/blocks/fwSsIGt1q1BcsrP9J3J1tAYa1VRGhL9bBpa7EixfwMt16n6G3MUOql6Ulawpv4sm.png" alt="" title="" class="object-cover bg-gray-100 rounded-sm md:rounded-md lg:rounded-lg" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionX="undefined" data-positionY="undefined" data-fullWidth="false"/></figure>

Maia Overview Presentation

<a data-mce-href="https://mypeasesoria.org/blog/maia-s--intellectual--architecture--situated-entrepreneurial-coaching" href="https://mypeasesoria.org/blog/maia-s--intellectual--architecture--situated-entrepreneurial-coaching" target="_blank" rel="noopener">The previous post</a> described Maia's appraoch to situated entrepreneurial coaching through a continuous coaching cycle, as well as its basis in the literature. How do we actually implement that in practice?About 90%+ of the time, the cue is Maia's tip of the week*. This used to be human-curated and fixed by sector, but now our coaching engine derives a customized tip for each user, the next idea/best practice with the highest marginal impact on their business. This is derived based on the user's own priorities and history as reflected in their entrepreneurial profile**, Maia's 8-dimension diagnostic of their business***, the country-sector specific knowledge bases we've built with local experts, and also considers emerging and seasonal themes via search. These tips are expressed via a set of WhatsApp template shells with a few different approaches.Interaction with the tool following the cue quickly takes the user towards the action plan, which is calibrated to the businesses' level of development and priorities, both as expressed in the conversation and in the profile. Along with the plan, Maia provides every asset it can to make taking the next step as easy as possible, such as templates, images, checklists, if-then scenarios, etc.The guided reflection and reward stage is implemented through a second weekly message. Unlike the cue, which brings in new information and ideas, this follow-up message looks backwards, encouraging and rewarding action and reflecting on results to overcome obstacles and find new opportunities. Again, implemented though a series of WhatsApp template shells.Interaction with the tool on this follow-up message leads directly to adjusting plans and taking corrective action when things worked poorly, and building habbits when things worked well. It naturally leads to a new cue, to start the cycle again.* The other 10% are inbound user questions. Current cadence is fixed weekly, our roadmap is to move to adapting the cadence to user preferences** A structured profile of each user, including key business characteristics and goals. Will be discussed in a future post.*** This is synthesized based on diagnostic tools and training programs from a variety of organizations like the ILO, the IFC, the FAO, and consulting companies. Will also discuss in a future post.

Part 2: Implementing this Architecture - MAIA's Coaching Engine

The Opportunity: Closing the "Adoption Gap"Traditional entrepreneurial training faces a well-documented "Adoption Gap." After multi-day courses teaching 20-30 new business practices, entrepreneurs typically implement only one or two. This leads to modest average profit gains of ~12.1% and leaves significant potential untapped. Maia is designed to close this gap, moving beyond one-time knowledge transfer to build lasting, high-impact habits through a continuous, AI-powered coaching cycle. The Continuous Coaching Cycle The architecture is a virtuous loop that turns insight into action and action into habit. It is powered by a dynamic, AI-driven core that builds an evolving profile of each entrepreneur and their business, ensuring every interaction is firmly based in their context and working towards their goals.Stage 1: Personalized Insight (The Cue)How it Works: Rather than only responding to inbound questions, the system proactively scans best practices including detailed country-specific knowledge bases, crosses them with the user’s situation and priorities, and proactively delivers new personalized ideas. These are single, targeted micro-insights framed as a simple "rules-of-thumb" to avoid cognitive overload.&nbsp; Basis: - Heuristic-Based Training: Simplified rules have been shown to increase profits by 8.1% in microenterprises, proving more effective than complex training. - Andragogy (Adult Learning Theory): Adults are motivated by solving immediate, real-world problems. This stage directly addresses their "need to know" and problem-centered orientation.Stage 2: Action Plan &amp; Asset (The Routine)How it Works: It then co-creates a short low-effort action plan with "if-then" scenarios to form a clear implementation intention. Friction is minimized by providing accompanying ready-made digital assets when possible (e.g., templates, schedules, content) that makes the behavior easy to perform. Basis: - Fogg's Behavior Model (B=MAP): Lasting behavior change comes from maximizing Ability (making it easy), in addition to building Motivation and providing the prompt.&nbsp; - Situated Learning Theory (Lave &amp; Wenger): Learning is most effective when it occurs in the context of application. Maia delivers the tool and the task directly into the entrepreneur's daily activities.Stage 3: Guided Reflection &amp; RewardHow it Works: Continuous structured follow-up prompts a supportive conversation to analyze results, learn from the experience, and celebrate the effort. This provides the immediate positive reinforcement that solidifies the actions as habits. Basis: - Personal Initiative (PI) Training: A focus on proactive mindset, error-learning, and opportunity-seeking has been shown to increase profits by 30%. The reflection stage is designed to coach this mindset - Kolb's Experiential Learning Cycle: This stage operationalizes "Reflective Observation," the critical step where experience is transformed into durable knowledge. - Duhigg's Habit Loop: The feeling of success and progress serves as the immediate "Reward" that tells the brain the entire loop is worth remembering and repeating.Stage 4: Adjustment &amp; AdaptationHow it Works: Insights from the reflection stage directly update the user's dynamic profile and inform the next cue. This ensures the next micro-insight is even more personalized, relevant, and impactful, creating a virtuous cycle of compounding gains. Basis: - High-Impact Consulting: The most effective entrepreneurial interventions evaluated in the literature are tailored and adaptive. Maia emulates this personalization at scale, a model proven to drive significant growth in SMEs.

Maia’s (Intellectual) Architecture: Situated Entrepreneurial Coaching

With Maia's new backend, we've shifted from the user's entire conversation history as context (truncated to fit into the LLM's window) to building a more intentional business &amp; entrepeneurial profile and providing that with a limited number of recent conversation turns to the LLM. This profile is the heart of the LLM's context, and has an number of advantages:1) As conversation histories become long (we now have users with &gt;1yr of history) the truncation can cut important business information provided in early turns.2) Research shows <a data-mce-href="https://arxiv.org/abs/2505.06120" href="https://arxiv.org/abs/2505.06120" target="_blank" rel="noopener">LLMs can get lost in long multi-turn convesations</a>. Reducing the conversation history helps reduce those episodes.3) Even for users with shorter conversation history, our testing shows better results with this two-stage process: first, a dedicated profile builder (with scheduled and episodic rebuilds), then second, the main LLM using this highly efficient and strucutred profile to respond.This approach also shows how a tool dedicated to helping MSMEs can perform better than a generalized tool. <a data-mce-href="https://www.forbes.com/sites/johnkoetsier/2025/08/09/gpt-5s-system-prompt-just-leaked-heres-what-we-learned/" href="https://www.forbes.com/sites/johnkoetsier/2025/08/09/gpt-5s-system-prompt-just-leaked-heres-what-we-learned/" target="_blank" rel="noopener">GPT 5's sytem prompt reveals a similar structure</a>: it builds up a 'user bio' over time to contextualize future responses. But that bio is structured to give contextualized responses about health, coding, cooking, family travel, and everything else ChatGPT is used for. We have the luxury of building a business and entrepreneurial profile highly targeted on a single purpose: to help business owners sell more, spend less, and grow. Since this is the only thing we do, we can explicitly keep out all PII and all the stuff advertisers care about but a business mentor would not. This is also where the research into business training and MSME productivity interventions becomes highly relevant- we include key fields like what is the businesses' stage of formalizastion, is it banked, and what are the entrepreneur's key stressors, etc. And we can get the responding LLM to fill in key parts of the profile with targeted questions over time. There are also key fields for particular sectors, like coffee growers and restaurants. The resulting profile, along with a more efficient record of past conversations, actions, and outcomes, greatly improves the tool's responses in our internal testing. We continue to refine the profiles and its integration into conversation history in a phased roll-out, and will share the structure in the coming weeks when that is complete.

The Business & Entrepreneurial Profile

We built a new backend for Maia! This was for two reasons: 1) OAI signalled they would be depricating the Assistants API and 2) new features were increasingly difficult to implement on the previous stack. After a couple months building and testing, it is now live and serving all of our traffic. Here are a few of the more notable changes:1) Model changes: we've long wanted to deploy a few different models, particularly for certain sector-questions that are complicated and require multi-step calculations and multi-turn online research. In addition to GPT 4.1 we are now using Gemini 2.5 Pro.2) Business profiles: now that we're managing conversation history directly instead of using OAI's threads, we're implementing <a data-mce-href="https://mypeasesoria.org/blog/can-an-llm-run-a-vending-machine" href="https://mypeasesoria.org/blog/can-an-llm-run-a-vending-machine" target="_blank" rel="noopener">improvements in context engineering</a> to make sure important business information is easily surfaced to the responding model as a detailed business profile rather than burried in a long transcript. These LLM-optimized business profiles are only available for that user's conversation and <a data-mce-href="https://mypeasesoria.org/" href="https://mypeasesoria.org/">follow our strict privacy policy</a>.3) Usability improvements: our users aren't familiar with AI and expect the tool to function like a human conversation on WhatsApp. That menas processing multi-message bursts as a single message and queueing messages rather than forcing users into the usual one message - wait for response before next message structure of AI chatbots.4) Image generation: we're adding LLM-generated images for particular business cases where its useful (eg ideating on logos and branding, social media promotions) while teaching users other tools for use cases that are less appropriate for LLM images (eg adding the Yape logo to their sign).Along with these and other improvements for users, this backend is also highly scalable and modular (implemented as a set of google cloud functions), which makes it easier to share with white-label partners.

Blog

MAIA @AMPYME Panama

Maia Overview Presentation

Part 2: Implementing this Architecture - MAIA's Coaching ...

Maia’s (Intellectual) Architecture: Situated Entrepreneur...

The Business & Entrepreneurial Profile

Maia's New Backend!