Summary. AI can make farm inputs and credit fairer, but only if the people it measures can see its numbers and challenge them. Corruption does not vanish when a list becomes a map. It moves to the map's edits, thresholds and overrides, so those are what a responsible programme has to expose.
#The wrong question
The question asked of an AI allocation system is "how accurate is it?". For a system that decides how much fertiliser a farmer receives, or how large a loan they can get, accuracy is the wrong place to start.
Input programmes have always run on a list, and whoever controls the list holds the power. A name can be added, a farm can grow on paper, and a bag can be signed for and never delivered. Mapping appears to answer this. Use satellite imagery and field data to locate each farm and estimate its size, tie the allotment to that estimate, and the middleman loses their pen.
Take a system built on that idea. Geo-data and AI locate each farm and estimate its size. Extension agents use the result in the field, inputs are allotted to match each farm, and the same data supports micro-loans to farm businesses that have no collateral. This paper examines that design against the published evidence, because it is the shape many agritech programmes now take.
The better question is who can correct the map, and what each correction leaves behind.
A model does not remove the discretion that corruption feeds on. It moves that discretion into the data, the thresholds and the override button, and a programme is responsible only if an outsider can inspect all three.
#What the evidence supports
We looked for a rigorous before-and-after figure showing how much leakage a digital farmer registry removes. We did not find one, and anyone quoting a percentage should be asked for its source.
What exists points in several directions:
- Access can improve. A 2015 thesis on Nigeria's e-wallet scheme in Kaduna State reported that 60% of farmers had obtained subsidised inputs, against 30% reported under earlier schemes. The sample was 189 farmers and the comparison is with earlier reports, so treat it as suggestive. The same study asked for allocations suited to each farmer's scale of operation, which is the gap farm mapping is meant to close.
- Benefits can be real and still narrow. A study of Nigeria's e-voucher subsidy in World Development found large improvements in productivity and welfare for beneficiaries, with a benefit-cost ratio of 1.11, which its authors call marginally cost-effective.
- Digitising can make delivery worse. Zambia's move to e-vouchers, studied across nearly 53,000 household observations, more than doubled the distance to fertiliser and made it 30% more likely to arrive late. Fertiliser use fell. Some farmers got their cards only after planting, and misspelled names failed to match national IDs.
The Zambian authors blame implementation and political will more than the concept. They also report possible deliberate efforts to derail the scheme by civil servants who had benefited from the old one. Leakage has owners, and they do not leave because the software changed.
| What the programme wants | The reflex | What goes wrong | What actually works |
|---|---|---|---|
| No ghost farmers | Register every field by satellite | Imagery shows a field, never who farms it | Link field, person and agent visit, then have someone else re-check a random sample |
| No inflated allotments | Allot per mapped hectare | The boundary becomes the thing to falsify | Log every edit with name, time, before and after, and flag edits clustering around one agent |
| Inputs matched to need | Trust the model's area | Small plots carry the largest proportional error | Publish a confidence band and measure on the ground below a set threshold |
| Credit without collateral | Score on farm data | A flood year reads as bad character | Separate shocks from behaviour and give every borrower a human review |
| Faster delivery | Digitise the voucher | Inputs arrive after planting | Measure arrival against the local planting date and keep a manual fallback |
#Where the risk actually sits
#An accurate map can still be an unfair one
Accuracy is a transfer of inputs, so someone has to choose who gains. World Bank research using Nigerian survey data found that farmers over-report the size of small plots and under-report large ones, while GPS measures are more reliable. Replace self-report with accurate measurement, keep a formula based purely on area, and the correction runs one way: smallest plots shrink, largest plots grow. That is our inference rather than the study's finding, but the arithmetic is hard to avoid.
The model's own error compounds it. One transfer-learning study covering northern Mozambique and southwest Nigeria reported a mean site-level error of 0.071 hectares in field size, which is a strong result. Applied as an illustration, that is about 14% of a half-hectare plot and about 2% of a three-hectare one. The same error is not the same harm, and errors on individual fields can exceed a site-level figure.
So this gets filed as a data science problem when it is really a policy decision. Whether allotments follow area alone, or carry a minimum for the smallest farms, should be decided by an accountable person and published. It should never emerge from a formula nobody reviewed.
#Corruption moves to the override
Every correction tool is also a falsification tool. Field agents must be able to fix a bad boundary, because models are wrong. An agent who can fix a boundary can also enlarge one for a price, and the result looks like model-verified data, which is harder to challenge than a handwritten list.
The cost is asymmetric. A forged paper record invites suspicion, while a forged polygon inherits the credibility of the satellite. Extension agents carry this pressure with little support: experts estimate one agent for every 2,500 to 10,000 farm families in Nigeria, depending on the state. Overstretched agents will take shortcuts, so judge the design by what it does on a bad day.
#A map is also a list of what can be taken
Visibility is not security. Customary land rights are often unwritten, and a boundary on a screen can look like a settled claim. A review of mapping initiatives in sub-Saharan Africa found they can help secure women's rights, but also warned of mapped land becoming more exposed to exploitation and of old boundary conflicts resurfacing.
A programme map should therefore state that it records farming activity for allocation and does not establish ownership. It should not be handed to anyone else by default, and the farm-level coordinates deserve the protection you would give a bank record.
#A credit score can launder the weather
A default in a flood year is a fact about the flood. Micro-loan models learn from repayment history. Train one on seasons that include drought or flood, and it can learn that a village is a bad risk when that village was simply underwater. Climate damage then becomes a permanent mark against the people who suffered it.
The opportunity is real. Research on KhetScore in Odisha, India, which scores farmers using satellite and smartphone data, found women more willing to apply and to seek larger loans, with stronger effects than for men. The researchers also worried about backlash, given women's lower digital literacy. That evidence comes from one Indian state and does not transfer automatically, but it shows that land records need not be the only gateway to credit.
#What this does for farmers and the climate
The strongest evidence concerns matched advice. A three-year randomised trial with 792 maize farmers in northern Nigeria tested a tool that lets extension agents generate field-specific fertiliser recommendations. Good management practices rose by up to 116%. Where farmers also received price information, nutrient rates, yields and revenues rose by up to 18%. Nutrient use efficiency improved and the greenhouse gas emitted per unit of output fell after two years.
Read that carefully. It shows lower emission intensity, not lower total emissions, and in part of the trial fertiliser use went up. Many of these soils need more fertiliser, not less. The honest climate claim is that each bag does more work, so farmers get more from what they apply. A claim of carbon savings would need measurement this trial did not attempt.
The map adds value in three other ways, each of which we regard as promising rather than proven:
- Timing. Where rainfall is growing harder to predict, the planting window matters more, and a bag that arrives after planting is wasted however precisely it was sized. Zambia shows the failure. Measuring delivery against local planting dates turns timeliness into a number someone must answer for.
- Response to shocks. Nigeria's 2022 floods damaged more than 650,000 hectares of farmland by a UN count. With fields mapped, a flood extent can be laid over the register to show which farmers were hit and which loans should be rescheduled, without waiting for each farmer to prove the loss.
- Protection of the record. The same overlay can mark a season as a shock, so that a flood-year default never reaches the credit model as evidence about the farmer.
#The record is the anti-corruption control
The anti-corruption value comes less from the AI than from the record it makes possible. Nigeria's Auditor General reported ₦629.04 billion outstanding under the Anchor Borrowers' Programme at the end of 2022, citing weak internal controls and inadequate disclosure of beneficiaries. The central bank replied that records of beneficiaries and cultivated hectares were "available on request", and the auditors found that unsatisfactory. The dispute is over what can be shown, and nothing here says who was at fault for the unpaid loans. A programme should be able to answer without being asked.
The shape of that answer is one record per farmer per season:
Allotment
farm, person, season
area: estimate, confidence, source (model | ground)
proposed_by: model version, rule version
decided_by: named person
edits[]: who, when, before, after, reason
delivered: quantity, date, confirmed by farmer
dispute: raised, reviewed by, outcome
This buys three things. An auditor can sample records and visit fields without asking anyone's permission. Unusual patterns, such as one agent's farms growing every season, become a query rather than a tip-off. And a farmer's complaint has somewhere to attach. The record should be append-only, because a log the agent can edit is just another list.
The limits on the model belong in writing too:
| The model may | The model may never |
|---|---|
| Estimate field area and report its confidence | Remove a farmer from the programme |
| Propose an allotment for a named person to approve | Refuse a loan with no human review |
| Flag unusual edits and patterns for checking | Settle a boundary dispute |
| Prioritise which farms an agent visits first | Record who owns the land |
This is also where the law is heading. Section 37 of Nigeria's Data Protection Act 2023 gives people a right not to be subject to significant decisions based solely on automated processing, with exceptions, and calls for safeguards including human intervention and the right to contest the decision. A contest right means little to a farmer who cannot read the screen, which is why the channel has to be a person, in the farmer's language, with the number in hand.
#Where this design is weakest
The weakness we would worry about most is that an appeals process costs a farmer more than it costs anyone else. It takes a trip, a day's work and the nerve to dispute an official. If few people appeal, the programme will read that as accuracy. Low complaint rates are poor evidence of fairness, so the error rate has to be measured independently, by ground-checking a random sample, and never inferred from silence.
#Before you deploy: questions for your own programme
| Question | If yes | If no |
|---|---|---|
| Has a named person chosen how allotments treat the smallest farms? | The trade-off is a decision you can defend | Your formula has made it for you |
| Can someone with no stake in the result sample records and visit fields? | You have an audit | You have a report |
| Do you measure error by farm size, gender and district? | You can see who carries it | You only know the average |
| Can a farmer see their number and challenge it through a person? | The right is usable | It exists only on paper |
| Is every edit logged in a record the editor cannot change? | Overrides are evidence | Overrides are the new list |
| Do you measure delivery against planting dates? | Timeliness has an owner | A perfect allotment can still arrive too late |
| Are shock seasons separated from borrower behaviour? | The score reflects conduct | The score punishes bad weather |
| Is the map barred from use as proof of ownership, and from sharing by default? | The risk to land rights is contained | You may have built a tool for taking land |
| Would the programme survive a week with no network or power? | Farmers are insulated from outages | The failure lands on them, in planting season |
The row teams skip is the first. It feels like policy, so engineers leave it alone, and it feels like engineering, so policy people never see it.
#When to skip this
- The real bottleneck is supply or funding. If fertiliser arrives late because money is released late, a better map only describes the delay more precisely. Fix the calendar first.
- Plots are very small, intercropped or shifting. Boundaries there are hard to see and models struggle. A trained agent with a GPS device may be cheaper and more accurate, and it still gives you the audit trail.
- You have no one to run the appeals. Without a staffed route for disputes, do not automate the allotment. Use the model only to choose which farms to verify first.
- Land is actively contested. Where mapping could inflame a dispute, consent from the community comes first, and sometimes the right answer is not to map.
- A flat allotment is good enough. Where farms are similar in size, a simple per-farmer allocation with strong delivery checks gets most of the anti-leakage benefit, without the model. It will not match need on mixed farms, and that is the honest limit.
#Key takeaways
- Stop asking only how accurate the model is. Ask who can correct it and what each correction leaves behind.
- Decide in writing how allotments treat the smallest farms, because accurate measurement alone can shift inputs towards the largest.
- Log every edit in a record the editor cannot change, then audit by random field visit.
- Report error by farm size and gender, never as a single average.
- Remove shock seasons from credit models, so that nobody's score carries the weather.
- Claim lower emissions per unit of output only where you measured it, and do not present it as a fall in total emissions.
- Measure delivery against the planting date. A well-sized bag that arrives late has still failed the farmer.
- State that the map is not evidence of land ownership, and restrict who may see it.
Based on published evaluations from Nigeria, Zambia and India, and on the author's product and technology work across agritech and financial services. The design described is a general pattern, not an account of one programme's results, and the claims about risk are arguments from that evidence unless a study is cited.