Who is agricultural AI really working for?

Artificial intelligence could help extension services respond to more farmers with more relevant advice. But a technically sound answer has no value to a farmer who cannot reach the system, afford the interaction, use its language or influence the data from which it learns. Before asking how well agricultural AI performs, we need to ask who gets to use it.

Exclusion begins before the first answer

Agricultural AI is being asked to solve a genuine problem. There are too few extension workers to give every smallholder farmer timely, field-specific advice, while changing rainfall, pest pressure and input prices make generic recommendations less useful. An AI service that can interpret local evidence and answer questions through a phone appears to offer advice at a scale that conventional extension cannot reach on its own.

Discussions about responsible AI often focus on whether a model gives a correct answer. Agricultural advice does need to be accurate: a wrong recommendation about a pesticide, planting date or livestock treatment can carry a direct cost. Yet accuracy tests normally begin after a user has reached the system and submitted a question.

For an advisory service, the chain begins earlier. The intended user must know the service exists, have access to its channel, be able to afford the interaction, understand how to use it, trust it with a real question and receive the answer in a form that can inform a decision. Failure at any one of these points makes the quality of the underlying model irrelevant to that person.

This is why mobile-network coverage is not the same as access to agricultural AI. The International Telecommunication Union reports that mobile broadband covered 86 percent of Africa's population in 2024, while only 38 percent used the internet. Cost, digital skills and reliable rural connectivity all sit between technical coverage and actual use.

A farmer may be excluded because she does not control the household phone. Another may own a basic handset but not a smartphone. An older farmer may be able to make a call but struggle with a menu designed around typing, passwords and small icons. If these people are missing from the service, they are also missing from the questions, feedback and usage data that shape its future development.

How access becomes unequal

The phone in a household is not automatically a woman's phone

The GSMA Mobile Gender Gap Report 2026 makes an important distinction. In sub-Saharan Africa, 72 percent of women owned a mobile phone in 2025, while 35 percent used mobile internet. The regional gender gap was 10 percent for mobile ownership and 26 percent for mobile internet use.

These are regional estimates, not figures for every East African country. They nevertheless show why a service cannot infer equal access from national mobile penetration. Ownership also says little about who controls a device when advice is available, who has money for data, whether use is private, or whether a phone can be used freely for questions about finance, land or household production.

Agricultural programmes can deepen the gap unintentionally. Registration may be tied to the person recognised as the household head, even where someone else performs much of the crop management. A service may send advice to the registered number and later count the household as reached. That does not establish who received the message, understood it or had the authority and resources to act.

The Kenyan data shows how representation becomes a model problem

A current initiative in Kenya makes the problem unusually visible. The iShamba SMS and WhatsApp advisory platform found that only about 15 percent of its users were women. Researchers working with AICCRA, the Alliance of Bioversity International and CIAT analysed more than 9,000 farmer questions submitted between 2020 and 2025.

They identified English-dominant responses, more frequent escalation of women's questions and weaker replies to short questions with little context. These differences matter because a farmer facing time, literacy or data constraints may ask a shorter question not because the problem is simple, but because the channel makes explanation difficult.

The project is responding with a human-in-the-loop approach: agricultural experts review and improve AI-generated answers using real farmer questions, while the evaluation explicitly considers language and gender. The broader lesson is that training data reflects access. More data does not automatically correct an imbalance if it comes from the same narrow user base.

Older farmers need better design, not a stereotype

Age is often treated too casually in digital agriculture. One version of the story assumes that younger people welcome technology while older people resist it. That explanation is both incomplete and unhelpful.

Research on agricultural technology adoption in Ethiopia has found that age interacts with education, extension contact, distance to markets and household resources. Work on the EzyAgric platform in Uganda similarly identifies unequal access to devices, skills and support, and recommends personalised training for groups that face greater difficulty using the platform.

The design implications are practical. A service can create unnecessary barriers through small text, unfamiliar symbols, multi-step registration, passwords, frequent software updates or a requirement to type a detailed question. Agricultural vocabulary may not have a familiar written equivalent in a local language even when the underlying practice is well understood in speech.

None of this means older farmers cannot use digital services. It means a programme should not make individual smartphone fluency the price of receiving agricultural advice. Voice, assisted use, demonstrations, farmer groups and a reliable human support route can make a service easier to use without reducing the quality of the information.

The smartphone assumption selects the audience

The rural connectivity gap makes the choice of channel an inclusion decision. According to the ITU, 57 percent of Africa's urban population used the internet in 2024, compared with 23 percent of the rural population. The median price of an entry-level two-gigabyte mobile broadband plan was 4.2 percent of monthly gross national income per person, more than twice the UN Broadband Commission's affordability target.

Those figures cover Africa as a region and should not be substituted for a country assessment. They do show why an Android application cannot be treated as a neutral delivery mechanism for smallholder agriculture. Requiring a recent smartphone, regular data and an individual account narrows the audience before agronomic relevance is considered.

Smartphones can still be useful. A camera can support pest identification, location can improve weather advice, and messaging applications can enable a useful conversation. The mistake is making one device the only entrance. A multi-channel service could combine SMS, interactive voice, community radio, extension workers, farmer organisations and smartphone applications around the same governed knowledge base.

Shared access can help, but it is not automatically equal access

Malawi offers a small but useful example. In 2021, FAO and the Government of Malawi distributed 78 smartphones through Farmer Field School networks and cooperatives in Mzimba and Kasungu. Thirty-four went to women and 44 to men. The phones were intended to support advisory services, market information and communication with value-chain actors, using video, audio and documents as well as text.

Working through a group can lower the cost barrier and avoids assuming that every farmer needs to own a smartphone. It can also connect digital information to an existing place for discussion and collective learning.

However, placing a device in a group does not prove that every member can use it equally. Who keeps it? Who decides which questions are submitted? When is it available? Can a woman ask a question privately? Is the person operating it able to explain uncertainty rather than simply repeat an answer? These questions need observation and feedback, not an assumption that shared ownership has solved inclusion.

For Malawi, the East African evidence is therefore a design lesson rather than a model to copy. Local languages, household decision-making, phone access, farmer organisations and extension capacity need to be studied in the communities a service intends to support. Our earlier article on digital farm advice in Malawi examines the wider service chain behind that work.

What we are evaluating at our test site

At ulimi.labs, we are currently evaluating this question through our own test site: who can actually use AI-supported agricultural advice when it meets the conditions of a real smallholder farming environment?

The purpose is not simply to demonstrate that an AI tool can produce an agronomically plausible answer. We are examining the wider service around that answer: who is able to reach it, whether the explanation is understandable, how people without their own smartphone can participate, when human support is needed and whether the advice can inform a practical farm decision.

The test site also gives us a place to compare a digital recommendation with field observation and discussion. That matters for practices such as intercropping, agroforestry, soil improvement and water management, where a useful decision depends on the plot, available labour and materials, the season and the farmer's own experience. A generic answer may sound convincing while missing one of those constraints.

What inclusive agricultural AI would require

An inclusive service does not need to offer every possible channel from its first day. It does need to identify whom its initial design excludes and create a credible route for reaching them. Six requirements stand out.

1. Start with access research, not a preferred interface

Before choosing an app, identify who owns which devices, who controls them, which channels people already use, what an interaction costs and which languages work in speech and writing. This should be investigated separately for women and men, different age groups, remote communities and people with disabilities.

2. Keep a basic-phone and assisted route

SMS, USSD and interactive voice response remain relevant. So do extension workers, call centres, community radio, farmer organisations and lead farmers. Assisted access is not a temporary failure to become digital; it can be a deliberate part of the service.

3. Design for speech and local meaning

Translation is not enough when a technical term has no familiar local equivalent or when the same word is used differently between districts. Farmers and extension workers need to test whether an answer is understandable, actionable and appropriate to the local production system.

4. Preserve a human escalation path

IFPRI's Generative AI for Agriculture work stresses reliable agricultural knowledge and human oversight. A service should be able to recognise uncertainty, ask for missing information and refer high-risk or ambiguous questions to a qualified person.

5. Give farmers a return for their data

Questions, field observations, crop images and corrections all help improve an advisory system. Farmers should know what is collected, how it is used and what they receive in return. Consent and data governance matter especially where a shared phone or intermediary submits information on someone else's behalf.

6. Test outcomes by group, not only in total

An overall satisfaction score can conceal a service that works well for young smartphone owners and poorly for everyone else. At minimum, programmes should examine repeated use, comprehension, trust, referrals and reported action by gender, age group, location, device and channel. Claims about yields, income or resilience require stronger evaluation than user analytics.

Five questions before calling a service inclusive

1. Who is absent from the registration and usage data compared with the farming population the programme intends to serve?
2. Can someone without a smartphone receive advice and ask a follow-up question at a reasonable cost?
3. Have women and older farmers tested the service themselves, rather than having an interface demonstrated to them?
4. Does the system know when to express uncertainty and involve an extension worker or specialist?
5. Are results reported separately enough to reveal who benefits and who drops out?

If these questions cannot be answered, the programme may still be a useful pilot. It should not yet be described as inclusive.

Inclusion is part of accuracy

Agricultural AI is often evaluated as if inclusion and technical quality were separate. In practice they reinforce one another. A system trained and tested with a narrow user group has less evidence about other farmers' crops, vocabulary, constraints and decisions. A farmer who cannot explain a problem in the available interface is more likely to receive an answer based on missing context. An unaffordable recommendation may be agronomically plausible and still be unusable.

The objective is not to place an AI chatbot in every farmer's hand. It is to make reliable agricultural knowledge easier to reach, question and use. Sometimes the appropriate channel will be a smartphone. Sometimes it will be a voice call, a radio programme, a farmer-group meeting or an extension worker using an AI-supported knowledge service on behalf of several communities. Our test-site evaluation is intended to help us distinguish between those possibilities in practice rather than assume the answer in advance.

Agricultural AI is inclusive only when farmers can reach it, understand it, question it and act on its advice. A service used mainly by connected, literate and commercially stronger farmers may be efficient. It is not yet working for everyone it claims to serve.

Frequently asked questions

Why might agricultural AI exclude women farmers?

Women may have lower rates of phone ownership and mobile internet use, less control over a shared household device, less money for data, less time to interact with a service or less representation in the data used to design it. These barriers vary by place and should be measured rather than assumed.

Does every farmer need a smartphone to use AI-supported advice?

No. AI can support services delivered through SMS, interactive voice calls, radio scripts, call centres, farmer organisations or extension workers. A smartphone can add useful functions, but it does not need to be the only channel.

Are older farmers unwilling to use digital tools?

That is too broad a conclusion. Use is shaped by interface design, prior experience, literacy, training, cost, trust and available support as well as age. Voice, assisted access and practical demonstrations can remove many avoidable barriers.

How can a programme tell whether its AI service is inclusive?

It should compare its users with the population it intends to serve and report repeated use, comprehension, trust and outcomes by gender, age, location, device and channel. Registrations or messages sent are measures of activity, not proof of equal access or benefit.

What can Malawi learn from East African experience?

The main lesson is not to copy a specific application. It is to combine locally grounded advice with several access channels, existing farmer organisations, human extension support and evidence that shows which groups are reached or missed. ulimi.labs is evaluating these questions at its test site; the work is ongoing and no results are claimed here.

Sources and further reading

1. The Mobile Gender Gap Report 2026 (opens in a new tab) — GSMA, 2026
2. State of digital development and trends in the Africa region: Challenges and opportunities (opens in a new tab) — International Telecommunication Union, 2025
3. Harnessing gender-responsive artificial intelligence for agricultural advisories in Kenya (opens in a new tab) — AICCRA, Alliance of Bioversity International and CIAT, 2026
4. Digital agricultural platforms and inclusivity: Access needs for user groups of the EzyAgric Platform in Uganda (opens in a new tab) — IFPRI, 2023
5. The effect of age on agricultural technology adoption by smallholder farmers in Ethiopia (opens in a new tab) — Advances in Agriculture, 2025
6. Generative AI for Agriculture, Phases I and II (opens in a new tab) — International Food Policy Research Institute, 2023–2026
7. Strengthening market-led digital agriculture platforms in Malawi (opens in a new tab) — Food and Agriculture Organization of the United Nations, 2021