Digital Green2025 - 2026

Scaling Agricultural Advisory Beyond Human Capacity

Making trusted agricultural expertise accessible, affordable, and useful at scale.

Product Manager – AI & Digital Transformation

  • AI
  • RAG
  • Agritech
  • Emerging Markets
  • 0→1
Farmers reached
200,000+Farmers reachedWithin the first year
Advisory interactions
1.5M+Advisory interactionsProcessed in year one
Lower cost per advisory
77%Lower cost per advisoryCompared to traditional human-led approaches
Activation rate
78%Activation rateImproved from 45%
The Context

Agricultural expertise could not scale fast enough

Millions of Ethiopian farmers rely on Development Agents for agricultural guidance on crop management, pest control, fertilizer use, and climate-related decisions.

The challenge is scale. A single Development Agent may support hundreds of farmers across multiple communities, making timely access to expertise difficult.

As demand for agricultural advisory services grew, traditional extension models struggled to keep pace.

FarmerChat was created to make trusted agricultural knowledge more accessible at scale.

The Challenge

Expertise existed. Access did not.

Agricultural expertise already existed across research institutions, agronomists, and government systems.

The problem was making that expertise available when and where farmers and Development Agents needed it.

Users often relied on low-end Android devices, limited data budgets, and inconsistent network quality. Agricultural recommendations also required a higher level of accuracy and trust than a typical consumer chatbot.

The challenge was making agricultural expertise accessible, trustworthy, and affordable at scale.

Product Strategy

We treated FarmerChat as an expertise access problem, not an AI problem.

The goal was not to introduce a chatbot. The goal was to help Development Agents and farmers access trusted agricultural knowledge faster and more consistently.

  • Trust before intelligence

    Agricultural advice had to be grounded in validated knowledge, not generic responses.

  • Accessibility before features

    The product had to work for low-end devices, low bandwidth, and real field conditions.

  • Adoption before scale

    A technically strong product with low usage would create little value.

Judgement Calls

Five judgement calls that decided whether FarmerChat reached real farmers or stayed a pilot

The most important product decisions were not about adding more features. They were about removing the barriers that would stop people from using the product.

  1. 01

    RAG instead of generic AI

    Agricultural advice required validated knowledge sources, not generic internet information.

  2. 02

    Optimize for low-end devices

    Many users relied on entry-level Android phones with limited storage and processing power.

  3. 03

    Reduce the cost of access

    The platform was optimized for low data consumption and supported by a Safaricom Ethiopia data partnership.

  4. 04

    Local language accessibility

    Agricultural guidance needed to feel understandable, relevant, and trusted.

  5. 05

    Continuous feedback loops

    Real-world interactions informed product improvements and advisory quality.

Access Strategy

Data affordability became a product decision

Connectivity was available for many users, but data affordability remained a major adoption barrier. I organized a partnership with Safaricom Ethiopia that reduced FarmerChat data costs by approximately 90% for Safaricom subscribers, with the remaining cost subsidized by Digital Green.

90%Data cost reduction
Execution

Building FarmerChat

FarmerChat combined large language models with a curated agricultural knowledge base to provide trusted advisory support through a conversational interface.

A Retrieval-Augmented Generation architecture connected user questions with validated agricultural content before generating responses.

Rather than maximizing features, the team focused on helping users access reliable agricultural guidance faster.

Execution

Removing friction at scale

Technology was only part of the challenge.

The team optimized the platform for low-end Android devices, low-bandwidth environments, and reduced data usage.

Removing access barriers proved just as important as building the product itself.

Execution

Driving adoption

Launching the platform was not enough.

We worked with field teams, Development Agents, researchers, and government stakeholders to integrate FarmerChat into existing advisory workflows.

Structured onboarding and continuous experimentation increased activation from 45% to 78%.

Results

AI-powered advisory at meaningful scale

FarmerChat demonstrated that trusted agricultural advisory could operate at scale in a resource-constrained environment.

Farmers reached
200,000+Farmers reachedWithin the first year
Interactions
1.5M+InteractionsAdvisory conversations processed
Cost reduction
77%Cost reductionLower cost per advisory
Activation
78%ActivationUp from 45%
Beyond The Metrics

The most important outcome was not the number of interactions. It was proving that trusted agricultural expertise could be delivered through AI while remaining accessible, affordable, and relevant to local realities.

The project demonstrated that AI adoption in emerging markets depends as much on distribution, trust, and accessibility as it does on model capability.

Lessons Learned

  • AI products rarely fail because of the model.

  • They fail because users cannot access, trust, or consistently use them.

  • FarmerChat succeeded because product decisions focused on removing barriers to adoption before adding complexity to the technology.

What I'd Do Next

Move from reactive answers to proactive advisory

I would invest earlier in proactive advisory experiences that anticipate farmer needs based on geography, seasonality, crop cycles, and historical interactions.