I work at the intersection of agri-tech and entrepreneurship, and in practically every conference I attend, every newsletter I open and every panel I sit on, I hear a vision of AI for agriculture that’s described as if it were already inevitable. The pitch goes something like this: “What if every farmer had a personal AI advisor telling them what to grow, when to plant and how much fertilizer or other inputs to apply?”
This approach is superficially compelling, and even sounds a bit like justice: democratising the knowledge that rich farmers pay consultants for, and delivering it for free to a smallholder in rural Kenya on a $40 smartphone.
But on the ground, it doesn’t work that way. Even if we had perfect data (we don’t), and even if every farmer had access (they don’t), some key questions remain: What does AI advising in agriculture actually solve, who will pay for it, and more importantly, why do we keep designing AI to make humans redundant instead of more powerful?
The Version of AI for Agriculture That Gets Demoed at Davos
There is a dominant mental model of AI for agriculture. It is built in San Francisco or Amsterdam or London. It is demoed at Davos or COP or the UN General Assembly. It is sleek, frictionless, and built on an assumption so quietly embedded that most people never notice it — namely, that the farmer is alone: alone with her phone, alone with her data, alone with the algorithm, a singular rational actor receiving personalised recommendations and optimising her decisions accordingly.
This assumption is not malicious; it is genuinely well-intentioned. However, it means AI models are designed to remove the human from the loop with fewer intermediaries. From my experience, humans cannot be automated out. Instead, AI models should be designed around making the human more powerful, not less necessary.
What We Learned from 5,000 Women Farmers and a WhatsApp Bot
In my work with Kuza, a certified B Corp social enterprise that supports 1.2 million farmers through a network of more than 6,000 youth advisors in seven African countries, we’ve learned a lot about how best to leverage both human and digital resources to support smallholders. Kuza’s model connects farmers to advisory services, inputs, finance and markets through human-trained young advisors (known as agripreneurs) and through our digital platform, the One Network. One lesson from our work stands out above all: A trusted human is the key, and technology infrastructure is the enabler. Remove the human and the technology fails to deliver.
We ran an AI advisory pilot, supported by the Gender & AI Livelihoods program, that worked with 5,000 women poultry farmers in Kenya. These were the kinds of farmers impact-focused organisations typically target: excluded, largely subsistence-level female smallholders. They work in groups, with around 50-200 birds per egg production cycle, buying feed at retail prices, and selling eggs and chickens with limited bargaining power into volatile local markets. These are the exact farmers AI for agriculture claims it will transform.
We gave them access to Kuza’s personalised, WhatsApp-based AI chatbot, Ask Nia, which provided real agricultural advice, available on demand, expressed colloquially in common local languages. Our first concern, adoption, turned out to be surprisingly easy. The women farmers used it, adoption was widespread, and data analytics showed that engagement was positive. They were resourceful about sharing devices within groups, maximising usage time by letting their fellow group members know when others were done using them.
Adoption, however, did not take the form many AI solutions assume. It wasn’t individual; it was social. In fact, uptake worked, in part, because we started with existing social groups, led by trusted agripreneurs. Use was embedded in these group dynamics. Recommendations from discussions on our AI chatbot sparked multiple in-person conversations among users, who questioned, interpreted and sometimes set this guidance aside, with input from the agripreneurs. These recommendations were one input among many, filtered through collective experience rather than followed at face value.
What did not change was where trust resided. It stayed with the agripreneur, the group champion with the most presence, credibility and accountability. The chatbot could explain topics like Newcastle disease (a highly contagious virus affecting birds) or good breeding practices, but when a farmer was faced with consequential decisions, like whether to switch feed suppliers, she turned to someone she knew, not a bot she queried.
Indeed, the question of input costs offers a prime example of the type of issues that AI can and cannot solve. Feed accounts for roughly 70% of poultry production costs in Kenya. That single fact affects farmers’ operations far more than any advisory ever could. Over 90% of the participating farmers in our pilot program cited feed cost as their biggest challenge. Even a working chatbot — and Ask Nia did work — cannot overcome challenges that emerge from this underlying cost structure. Better information itself is not enough, if it cannot translate into better outcomes, and the business model for farmers remains fragile.
This is where much of the current narrative around AI in agriculture falls short. It assumes that information is the primary constraint. But in most of the markets we work in, the main constraints are structural. Input costs remain high, market access is inconsistent, cold storage is lacking and infrastructure is unreliable. Farmers are operating within systems where even correct decisions based on accurate information do not guarantee viable returns.
Where AI Actually Helped: Making an Invisible Market Visible
While our AI chatbot could not directly solve the challenge that actually mattered most to the farmers, the cost of feed, it did something that was arguably more interesting, and that ultimately enabled collective solutions to this challenge. It made a market visible that had long been invisible.
Individually, each of the 5,000 poultry farmers was too small to matter to any serious input supplier: They represented tiny, fragmented, unreliable sources of demand that no feed company would restructure its logistics for. However, our data — aggregated, analysed and made accessible through AI — told a different story. These 5,000 women together managed tens of thousands of birds on coordinated production cycles. This represented predictable demand, genuine volume — and from a feed company’s perspective, a customer worth having. Without AI, few businesses would have the capacity or resources to collect and analyse this data or recognise this opportunity. But with it, a new local market came into view, enabling providers to offer more affordable, bundled services.
As implementation progressed, the frame of the conversation between farmers and Ask Nia shifted from questions like “How do I get better advice about feeding my 100 birds” to topics like “Can we negotiate bulk feed discounts as a group?” Many of the agripreneurs, who provide their advisory services to farmers for a fee as an entrepreneurial endeavor, started exploring new business opportunities, (e.g., launching a small feed mill enterprise to serve this new market). Interestingly, when we put the question of how to reduce the cost of feed to the group, the responses were immediate and creative. Some farmers shared how they were experimenting with black soldier fly larvae to convert food waste into protein-rich animal feed. One mentioned a neighboring farmer, outside of the pilot program, who was producing his own feed in bulk. Someone pointed out that the vegetative waste most people throw away, when sorted correctly, is already animal feed. A number of the circular economy approaches regularly discussed in conferences began operating in practice.
The resource that unlocked all of this wasn’t algorithmic. It was human. These women had always been part of the broader agricultural market, although individually they remained invisible to it. AI did not create the demand that sparked the emergence of more affordable inputs. It made that demand visible enough to incentivise providers to start acting on it.
The Design Choice Nobody Names
Much of today’s AI debate is based on a long-standing design default: Technology creates value by reducing or eliminating human intervention wherever possible. From factory automation to self-checkout, ATMs and now generative AI, a recurring question has been: How can this task be performed with less human involvement? This is not true of every technology or every society, but it has become a dominant assumption in much of Silicon Valley’s approach to innovation.
Yet as Daron Acemoglu and Simon Johnson argue in “Power and Progress,” technology does not advance through an inevitable march of progress, but rather produces outcomes shaped by prevailing economic and social priorities. For decades, these outcomes have overwhelmingly favoured labour-replacing automation over human augmentation. The drive to remove humans from the loop is therefore not a universal law: It is a specific design choice.
Our work with poultry farmers shows that there’s another way. The agripreneurs we work with are not a cost to be engineered away. The group champion is not an inefficiency to be disintermediated. The trusted local advisor is not a legacy artefact waiting to be replaced by a chatbot. These people are the system. They carry the trust, the context and the relationships that make any intervention, technological or otherwise, actually work. Optimising agriculture in this context is not primarily an information problem. It is a trust problem, a market access problem, a collective action problem and a financing problem. And above all, it is a relationship problem: It needs the human connection, the social element.
This dynamic is already visible in how rural economies organise capital through informal savings and investment groups, where members regularly pool money and rotate access to the funds for business, education or emergencies. In Kenya, these “chamas” are widely used: They function without contracts or platforms, relying instead on trust, proximity and accountability between members.
Chamas reveal something fundamental about how decisions are actually made in these systems. Even where money, risk and productivity are involved, the binding constraint is not information but trust and coordination. New tools, whether financial or digital or AI, do not replace that layer; they are absorbed into it.
This is the lens through which we saw our AI pilot unfold. AI has a role in this sort of work, but the design question has to be changed: from “How do we use AI to replace the human intermediary” to “How do we use AI to make this human intermediary 10 times more powerful?”
AI can process the data that turns a group of 5,000 individual farmers into a viable, unified market for a feed supplier. It can flag when a flock’s feed conversion ratio is drifting, allowing farmers to intervene before it becomes a crisis. It can help an agripreneur serve 10 times more farmers without losing the personal quality of the relationship. And it can identify, at scale, which interventions are actually changing behaviour, and which are just being politely ignored. AI is a multiplier, but it needs someone to multiply.
A Different Design Brief for AI for Agriculture
The question every AI-for-agriculture team should ask before they write a single line of code for Africa is, “Are we designing to empower the human, or to remove them?”
If the answer is “remove” we need to be honest about it and own the choice, rather than portray it as neutral. If the answer is “empower,” these are some of our learnings about what actually works on the ground:
- Design for the group, not just the individual: African farming decisions are social. The unit of adoption is often the savings group, the cooperative or the family compound. An AI tool that works for one isolated farmer but breaks down in a group setting has missed this context entirely.
- Design to amplify the trusted human: The agripreneur, the extension worker and the group champion have legitimacy that no algorithm can purchase. Don’t cut them out — instead, give them better data, faster answers and tools that make them genuinely more capable. The goal should be to enable them to serve 2,000 farmers instead of 200, with no drop-off in the quality of their services.
- Design for structural constraints, not just informational ones: If the problem is feed costs, information is not the solution — demand aggregation, local processing and waste valorisation are. AI can help model and facilitate these, but someone has to go knock on the feed supplier’s door. That someone is a human.
- Design for data equity: Every interaction these farmers have generates data. That data should make their lives better, not just feed a platform’s AI model while they remain as marginalised as before. If AI is extracting value from excluded communities instead of adding it to them, we have reproduced the problem we claimed to solve.
- Design for financial sustainability: Ask who pays from the start. While the generic answer to questions about a new technology’s costs is that it will get cheaper over time as it scales, the proof of affordability is in the pudding, and depends on whether the farmer will pay for it. And farmers will only pay if they see value: increased incomes, reduced costs — something tangible in their bottom line. Too often, we expect farmers to adopt technology based on the promise of future improved productivity or incomes. They will not typically take that risk, as they are rational, practical decision makers.
The 5,000 women in our pilot didn’t need a better algorithm. They needed their collective scale to be visible to suppliers, and they needed the humans in their network to have the tools and data necessary to help them optimise their operations. AI helped with all of that *through* people, not instead of them.
This is the design choice that matters: Africa does not need AI that replaces the agripreneur. It needs AI that makes her unstoppable. That version is messier. It doesn’t demo as well. It requires you to understand the group dynamics of a local women’s cooperative, the trust dynamics of a rural agrodealer network, and the exact moment in a production cycle when a farmer most needs someone to pick up the phone.
The opportunity, therefore, is not to substitute these systems with AI alternatives, but to build on them. The future of AI in Africa will not be individual-first. It will be community-driven, youth-led, and AI-enabled, where technology strengthens the social fabric rather than replacing it.
Sheena Raikundalia is the Chief Growth Officer at Kuza.
Photo credit: LENblR
Publisher: Source link
