Close
Research

Cooperatives face measurement, costly debt and untested AI

(26 min read)

Across the last weeks of 2025 the World Bank and the IMF put out a run of material that, read together, asks three questions. What do we count? What does money cost? And who gets to use the machines? The Bank put its Scorecard in the spotlight, published its International Debt Report 2025 and two reports on artificial intelligence, one of them on agriculture. The IMF’s podcast asked what AI cannot replace. A cooperative of smallholders lives inside all three: it is measured by funders, it borrows or avoids borrowing, and it is being offered chatbots and sensors. This article reads that material for what it means for the cooperatives Kilimora serves through AgriKonnekt.

It is the third in a series. Our companion pieces ask who pays when the rain fails, and who really gets the capital when a woman is given a grant or loan. This one asks what a cooperative should measure about itself, what it should ask about the money, and what to expect from AI.

Read this first. We read the Bank’s two Scorecard blog posts (insights and the data update), its debt data blog and press coverage of the debt report, the landing page of Harnessing AI for Agricultural Transformation, the fact sheet and an agriculture case study from the Digital Progress and Trends Report, and the IMF’s podcast page. We have not read the full debt or AI reports, we did not hear the podcast or retrieve Peña’s article, and two World Bank newsletters (28 November and 5 December 2025) exist only as emails. Two numbers differ between sources, and we say so where they do.
Four numbers from one season of reports

World Bank Scorecard, International Debt Report 2025, Digital Progress and Trends 2025 and the AI in agriculture case study

22results indicators on theWorld Bank Scorecard$741Bnet outflow from developingcountries, 2022 to 202427%internet use, low-incomecountries (93% in rich ones)$0.35projected chatbot advice costper farmer, down from $35
Seven weeks that framed the year end

Release dates of the material this article reads, October to December 2025

Oct 2025NovDecJan 2026Scorecard FY25 insights8 Oct8 OctScorecard Annual Meetings update14 Oct14 OctDigital Progress and Trends 202521 Nov21 NovIMF podcast, Pablo Peña24 Nov24 NovHarnessing AI for agriculture25 Nov25 NovInternational Debt Report 20253 Dec3 DecKenya Treasury on debt service21 Dec21 Dec

Eight findings

  • The Scorecard measures the institution. It tracks 22 results indicators and reports reach in the hundreds of millions: nearly 370 million people with health services, 325 million students, 244 million on safety nets.
  • Reach is a third to a half of what is expected. For transport, electricity, broadband and climate resilience, FY2025 results are 32% to 50% of what current projects expect to reach.
  • Poverty is concentrating. About 839 million people, 10.3% of the world, live in extreme poverty, 46% of Sub Saharan Africans among them.
  • Money flows the wrong way. Developing countries paid $741 billion more in principal and interest than they received in new financing in 2022 to 2024, the largest gap in at least 50 years.
  • Interest is crowding out the state. Countries spending over 20% of revenue on interest rose from 9 in 2017 to 21 in 2024; in the 22 most indebted countries, 56% of people on average cannot afford a minimum healthy diet.
  • AI for farming is promising and thinly proven. The Bank maps 60 use cases, yet its own case study counts about 15,000 farmers on its best known chatbot and finds fewer than 2% of US farm firms using any AI.
  • The gaps are basic. Internet use is 93% in high income countries and 27% in low income ones; basic digital skills are under 5% against 66%.
  • People stay in the loop. The IMF’s Pablo Peña argues that critical thinking and curiosity, not machines, keep people in charge, a point the evidence on generic AI advice supports.

Inside the Scorecard

The Bank calls its Scorecard an award winning website that tracks the Group’s work against 22 results indicators. Its latest insights, for the financial year ended June 2025, read like a list of very large numbers: nearly 370 million people accessing health services, 325 million students supported, 244 million people reached by safety nets, 217 million with broadband, 215 million with electricity and 136 million with strengthened climate resilience. The chart below uses those figures exactly as the Bank reports them.

What the Scorecard counts: people reached in FY2025

Millions of people. Blue = people, amber = digital, grey = infrastructure, green = planet. World Bank Group Scorecard insights, FY2025

Health servicesaccessednearly 370Mnearly 370MStudentssupported325M325MSocial safety netsbenefited244M244MBroadbandaccess217M217MElectricityaccess215M215MDigitally enabledservices used177M177MClimate resiliencestrengthened136M136M

The more useful chart is the second one. For four indicators the Bank also gives what current projects are expected to reach. Electricity stands at 215 million against 576 million expected, transport at 176 million against 524 million, and climate resilience at 136 million against 425 million. Only broadband is halfway, at 217 million of 431 million. These are projections for projects already under way, not targets, but they show how far the work still has to run.

Reached so far, against what current projects expect to reach

Millions of people. “Expected” is the Bank’s forward projection for projects under way, not a target

Reached in FY2025Expected to reachTransport176M176M524M524MElectricity215M215M576M576MBroadband217M217M431M431MClimateresilience136M136M425M425M

The Annual Meetings data update, published a week later, shows the problem the Scorecard exists to address. About 839 million people, 10.3% of the world’s population, live in extreme poverty, and Sub Saharan Africa has the highest rate at 46% in 2024. In 2000 more than a quarter of people in IDA and IBRD countries lacked electricity; by 2023 that share was below 10%. Mission 300 aims to connect 300 million people in Sub Saharan Africa by 2030, about half of those without power today.

Where extreme poverty sits

Left: share of the world’s population in extreme poverty. Right: Sub Saharan Africa’s rate, 2024

10.3%89.7%839Mextreme poverty10.3%About 839 millionin extreme poverty89.7%Everyone else(our arithmetic)46%54%46%Sub-Saharan Africa46%Sub-Saharan Africansin extreme poverty54%Everyone else(our arithmetic)

One result is good news with a catch. Account ownership among adults in the Bank’s client countries rose from 42% a decade ago to 76% today, and among women in South Asia from 27% to 83%. The Bank notes that gaps remain, especially for women and marginalised groups. Our companion piece on Kenya showed the same pattern: near parity at the account, large gaps in what flows through it.

Financial inclusion: the account is no longer the gap

Share with a financial account, per cent. Adults in World Bank Group client countries, and women in South Asia

A decade agoLatestAdults with an account,client countries42%42%76%76%Women with an account,South Asia27%27%83%83%

The Bank’s blog also cites a Zambian programme in which women got small grants with skills training and mentoring. It reports extreme poverty among participants down 30%, business profits up 80% and savings up by more than 230%. That is one programme, reported by the Bank, and we have not checked it. It fits, though, the pattern in the household level evidence we reviewed last week: cash for women works best when it comes with something that raises what the money can do.

Cash plus skills in Zambia, as the Bank reports it

Reported change among women who received small grants with training and mentoring. Single programme, reported by the Bank, not independently checked here

Extreme povertyamong participants−30%−30%Businessprofits+80%+80%Savingsmore than +230%more than +230%

The Scorecard is also candid about what lags. It says 56 countries still collect less than 15% of GDP in taxes, though 20 strengthened revenues, and that 60% of high risk countries put debt reforms in place. The Bank is consulting on a final indicator for jobs that would count people who move into better paid work because of its projects, not only jobs created. That is a harder test than reach, and the right one.

Where the Scorecard says it is lagging: money to run a state

Countries, FY2025

Collect under 15% ofGDP in taxes56 countries56 countriesStrengthened taxrevenues20 countries20 countries

Outside critics are less impressed. A Bretton Woods Project piece from October 2024 carries the headline that the new scorecard overlooks client priorities and ignores development impact. We have seen the headline and not the argument, so we cannot say how far it holds; we raise it because any list of indicators is a choice of what to leave out.

For Kilimora: three lessons. First, report what lags, not only what grows: a public page that lists plots that failed a check is more credible than one that lists only successes. Second, separate reach from benefit. Enrolling 1,500 farmers is reach; showing that their incomes rose is benefit, and the Bank’s own planned jobs indicator draws the same line. Third, publish definitions and update on a schedule. The table below translates five Scorecard habits to a cooperative.

What the Scorecard doesThe equivalent for a cooperativeWhy it matters
Counts people reached, then reports what lagsPlots enrolled and verified, then plots that failed a checkA list of failures is what makes the successes believable
Separates what was reached from what is expectedSeparates payments made from payments scheduledFunders read the gap between the two as delivery risk
Plans a jobs indicator that counts people moving into better paid workCounts members whose income rose, not only members trained or paidReach says who was touched; income says who benefited
Publishes definitions and updates on a scheduleA public indicator page with definitions, baselines and a season by season updateIt is the cheapest form of accountability to farmers
Lists indicators that were not updatedStates which indicators were not measured this season, and whySilence on a missing number looks like a good number

The price of money: the International Debt Report

The International Debt Report 2025, released in the first days of December, covers 120 low and middle income countries. Its headline is a reversal: those countries paid out $741 billion more in principal and interest than they received in new financing between 2022 and 2024, the largest negative transfer in more than five decades of data. Their external debt reached a record $8.9 trillion in 2024, of which the 78 mainly low income IDA eligible countries owed $1.2 trillion.

The interest bill

Interest paid by low and middle income countries on external debt, US$ billion

About ten yearsearlier*under $173B*under $173B*2024$415B (record)$415B (record)

* The report says interest was more than 2.4 times the level a decade earlier, so the earlier level was under $173 billion; that is our arithmetic.

The burden is not spread evenly. In low income countries, interest made up over half of external debt service. Interest consumed roughly 20% to 40% of government revenue in many countries. The number of countries spending over 20% of revenue on interest more than doubled from nine in 2017 to 21 in 2024.

How many countries are spending a fifth of revenue on interest

Number of countries in the World Bank debt data by share of government revenue spent on interest

2017: over 20%of revenue992024: over 15%of revenue33332024: over 20%of revenue21212024: over 25%of revenue1414

The price of new borrowing is the reason. Bonds sold before 2020 often carried coupons of 4% to 6%; many issuers now face 8% to 10%, and the Bank says new bond financing hovered around 10%. Bond markets did reopen: net flows swung from an outflow of about $75 billion in 2023 to an inflow of about $55 billion in 2024, according to the Bank’s data blog. The World Bank itself provided a record $18.3 billion more to IDA eligible countries than it received in repayments, plus $7.5 billion in grants, while bilateral creditors received $8.8 billion more than they disbursed.

What new borrowing costs

Typical interest rates on international bonds, per cent a year

0%4%8%12%Eurobond coupons4% to 6%4% to 6%Many issuers now8% to 10%8% to 10%
Markets reopened, but the net flow swung wide

Net bond flows to low and middle income countries, US$ billion

Net bond flows2023−$75B−$75B2024+$55B+$55B
Who sent net money, who took it, in 2024

Net transfers to IDA eligible countries, US$ billion. Negative = creditors received more than they disbursed

Net to countriesWorld Bank lending(net of repayments)+$18.3B+$18.3BWorld Bankgrants+$7.5B+$7.5BBilateralcreditors−$8.8B−$8.8B

Who holds the debt has changed too. Private bondholders account for nearly 60% of long term public external debt, and bondholders are the fastest growing group of creditors to IDA eligible countries. Paris Club creditors, which held nearly half of IDA countries’ public debt in the early 2000s, now hold about 10%. A fragmented creditor base is harder to bring to the table when a country cannot pay. Developing countries restructured $90 billion of external debt in 2024, the most since 2010.

A record stock, mostly owed to markets

Left: external debt of low and middle income countries, 2024. Right: private bondholders’ share of long term public external debt

$1.2T$7.7T$8.9Texternal debt$1.2T78 mainly low-incomeIDA-eligible countries$7.7TAll other low- andmiddle-income debt**“Nearly 60%”The rest≈60%bondholders“Nearly 60%”Private bondholdersThe restOfficial and othercreditors

** We assume IDA eligible countries’ debt sits inside the $8.9 trillion total; the $7.7 trillion remainder is our arithmetic.

The creditors changed

Paris Club creditors’ share of public debt owed by IDA eligible countries, per cent. Approximate

Early 2000snearly halfnearly halfTodayabout 10%about 10%

The human cost is stated bluntly. Among the 22 countries whose external debt exceeds 200% of export revenue, an average of 56% of people cannot afford the minimum daily diet for long term health, and in the 18 IDA eligible countries of that group nearly two thirds cannot. The Bank’s own December newsletter gives the figure as 50%, so the numbers differ by source; we use the report coverage’s 56% and flag the gap.

When debt service competes with dinner

Share of people who cannot afford the minimum daily diet for long term health, in countries whose external debt exceeds 200% of export revenue

56%44%56%cannot afford56%Average across the22 countries44%Can afford it(our arithmetic)“Nearly two-thirds”The rest≈66%cannot afford“Nearly two-thirds”In the 18 IDA-eligiblecountries of the 22The restApproximate

Kenya shows the squeeze up close. The Controller of Budget put national debt at Sh11.7 trillion in 2024/25, up from Sh10.5 trillion, with Sh1.6 trillion spent on servicing it. A Treasury official told a newspaper in December 2025 that 48.5% of ordinary revenue collected by the Kenya Revenue Authority goes to debt related payments, a figure that includes pensions and other consolidated fund services as well as interest, and that debt stands at about Sh11.5 trillion, or 64% of GDP. The two debt totals differ slightly by source, as do their definitions.

Kenya: the stock and the servicing

Left: public debt, Sh trillion, 2023/24 and 2024/25 (Controller of Budget). Right: Treasury official on share of ordinary revenue going to debt related payments

Sh10.5T+Sh1.2TSh11.7T2024/25 stockSh10.5T2023/24+Sh1.2TAdded in 2024/25(to Sh11.7T)48.5%51.5%48.5%of ordinary revenue48.5%Debt-related paymentsincl. pensions***51.5%Everything else(our arithmetic)

*** Includes consolidated fund services such as pensions, not interest alone.

The same pattern ran through the recovery from Hurricane Melissa. The Bank’s 5 December newsletter lists Jamaica’s secured package of US$6.7 billion over three years; as we noted in our companion piece, the Gleaner reported that most of it is financing and credit lines rather than grants.

For Kilimora: climate finance is not all alike, and a cooperative should know which kind it is taking. A payment for verified practice leaves the cooperative owing nothing; a loan, a bond or even input credit carries an interest cost that, as these numbers show, can swallow the benefit. AgriKonnekt offers organic fertiliser in flexible instalments, which is credit, and the real test is the instalment against a harvest time income. The table below gives five questions to ask of any money on offer.

Money on offerWhat a cooperative should askWhy
GrantIs it tied to verified practice? Who owns the data?No repayment, but conditions can still move cost onto members
Results based paymentWhat is measured, how often, who verifies, and what happens in a bad season?Payment depends on evidence, so evidence cost is part of the price
Concessional loanWhat rate, term and currency?The Bank’s debt data show official lenders’ new rates at a multi decade high
Commercial loan or bondWhat coupon and refinancing risk?Many bond issuers now face 8% to 10%, against 4% to 6% before 2020
Input creditWhat instalment schedule against harvest income?AgriKonnekt offers fertiliser in flexible instalments; the real cost is the instalment against income

AI for agriculture: what is proven, what is promised

The Bank’s new report, Harnessing AI for Agricultural Transformation, argues that AI could make smallholder farming more productive, more climate resilient and less labour intensive, if it is used where it truly adds value. It maps 60 use cases across the value chain, from faster seed breeding and pest detection to price forecasting, alternative credit scoring and climate indexed insurance. Small scale producers, which it says grow about a third of the world’s food, are its focus. Its recommendations are about foundations: open data, AI ready identity and payment systems, national AI strategies that cover agriculture.

AI for agriculture in four numbers

World Bank, Harnessing AI for Agricultural Transformation, and the Digital Progress and Trends case study on agriculture

60AI use cases mapped acrossthe agrifood value chain15,000farmers on Farmer.Chat infour countries, so far<2%of US farm firms used anyAI in a late-2023 survey47%of surveyed farmers nametech cost a top concern

The evidence base is the thinnest part. In the agriculture case study of the Digital Progress and Trends Report, fewer than 2% of US farm firms used any AI in a late-2023 survey, and the best known farmer chatbot has reached about 15,000 farmers across Ethiopia, India, Kenya and Nigeria. The chart below sets the reported yield effects side by side, and the colours matter: some come from published studies of simple tools such as text messages, one from a pilot, and some from the vendors themselves.

What digital advice has done to yields, and how well each is evidenced

Reported yield change, per cent. Blue = published studies, green = pilot, amber = vendor reported. Different crops, places and designs: do not average

Uganda maize:SMS reminders+10.5%+10.5%Kenya sugarcane:text-based advice+11.5%+11.5%India chili: AIadvisory pilot, per acre+21%+21%Banana farms:CropX, vendor-reported+27%+27%Trapview,vendor-reported+5%+5%

The cost claim is the most striking and the least tested. The case study says a chatbot could cut the cost of advice from $35 to $0.35 per farmer, but the effect on yields or incomes has not been tested. It also records that machinery and equipment cost 35% to 39% more in Sub Saharan Africa than in the US, and that 47% of surveyed farmers cite technology cost as a top concern. A related study found that ChatGPT’s weed control advice for Nigerian cassava farmers was often too generic or impractical for smallholders.

The promise on cost, and the missing test

Cost of extension advice per farmer, US$. The chatbot figure is a projection that has not been tested for impact

Extension advicetoday$35$35By chatbot,projected$0.35 (untested)$0.35 (untested)
Why chatbots tempt: too few human advisers

Farmers per extension agent. Bars are scaled to the largest value

Recommended(Ghana benchmark)1 per 7501 per 750Kenya1 per 9501 per 950Uganda1 per 2,5001 per 2,500Ghana1 per 10,0001 per 10,000

The motive for chatbots is a real shortage of human advisers: one extension agent for 950 farmers in Kenya, 2,500 in Uganda and 10,000 in Ghana, against a recommended 750. There are African examples that work at scale. Kenya’s Agricultural Observatory Platform gives real time data and forecasts to 1.1 million farmers, and Hello Tractor reports 3.5 million acres digitised, over 6,000 jobs and 5 million tonnes of additional food since 2014. Private agrifood tech investment in Sub Saharan Africa rose from under $10 million in 2014 to about $600 million in 2022. All of these figures are the programmes’ own or the Bank’s, not independently audited.

Digital advice that already works at scale in Africa

Figures as reported by the programmes and the World Bank; not independently verified

1.1Mfarmers on Kenya’s Agricul-tural Observatory Platform3.5Macres digitised by HelloTractor since 20146,000+jobs created, as reportedby Hello Tractor5M textra food output claimedby Hello Tractor
Money is arriving

Private agrifood tech investment in Sub Saharan Africa, US$ million

2014under $10Munder $10M2022about $600Mabout $600M

The Digital Progress and Trends Report names the reason the benefits may not reach smallholders. Internet use is 93% in high income countries, 54% in lower middle income and 27% in low income countries; the agriculture case study quotes about 40% for low income countries, so definitions differ and we use the report’s own figure. Computing is more lopsided: high income countries hold 77% of co location data centre capacity and low income countries under 0.1%. They also take 91% of AI venture funding with 17% of the world’s people. Basic digital skills reach under 5% of people in low income countries, against 66% in high income ones.

The first gap is connectivity

Share of people using the internet, by country income group. Digital Progress and Trends Report 2025

High-incomecountries93%93%Lower-middle-incomecountries54%54%Low-incomecountries27%27%
The second gap is computing, and who funds the models

Left: share of co location data centre capacity, June 2025. Right: share of AI venture funding going to high income countries

77%18%5%under 0.1%77%in rich countries77%High-income18%Upper-middle-income5%Lower-middle-incomeunder 0.1%Low-income91%9%91%of AI funding91%AI venture funding tohigh-income countries9%Everyone else(our arithmetic)
The third gap is skills

Share of the population with basic digital skills

Low-incomecountriesunder 5%under 5%High-incomecountries66%66%

The report’s answer is the “four Cs”: connectivity, compute, context (local languages and data) and competency (skills). It also backs “small AI”, tools that run on everyday devices without heavy infrastructure. That is close to how AgriKonnekt is built: lessons by USSD with no data plan, sensor readings sent over the GSM network, and AI running on the field sensor itself. The table below sets the four Cs against what AgriKonnekt is designed to do, and the status of each claim.

The Bank’s “four Cs”What AgriKonnekt is designed to doEvidence status
Connectivity: affordable, reliable internetLessons by USSD on any phone, data moved over GSM with no data planDesign claim; reach and completion rates not published
Compute: computing resourcesAI runs on the field sensor, so it works without a cloud connectionDesign claim; no accuracy results published
Context: local languages, data and realitiesTraining in Swahili, Kikuyu and other languages; cooperative led curriculumLanguage count is Kilimora’s statement; learning outcomes not measured
Competency: digital literacy and skillsFarmer educators who train their own networks (a cascade model)Target; educator numbers by county not yet published

Two platforms show what AI adjacent tools offer women specifically, with the honest caveat that neither has published outcomes we could find.

PlatformWhat it offers women farmersWhat we do not know
HerVest, Nigeria (launched 2020)A savings app built on rotating savings circles; buyer agreements, aggregation, training and a credit history from loan and repayment records; a USSD version was plannedNumber of women served, repayment results and impact: none reported in the 2021 profile we read
Lersha, Ethiopia (with Accion, Sidama Zone)Input loans, climate advisory and market links through an agent; account set up and financial trainingLoan volumes, repayment and women’s share of Lersha users: not reported; the project says over 5,000 women registered
Farmer.Chat (Ethiopia, India, Kenya, Nigeria)Conversational advice; the Bank projects extension cost falling from $35 to $0.35 per farmerImpact on yields or income is untested; about 15,000 farmers reached so far

A trade newsletter we receive also reports that the UK Agri Tech Centre and the Carbon Trust have launched an accelerator called FASTA to bring measurement, reporting and verification technology to market. We have not seen the programme itself, but it signals that the verification layer AgriKonnekt builds is becoming a recognised category.

The IMF’s counterweight: human capital

The IMF’s podcast this season features Pablo Peña, a University of Chicago economist and author of Human Capital for Humans, in conversation with Rhoda Metcalfe about his article in the December issue of Finance & Development. Going by the IMF’s page, his argument is that AI outperforms people in many tasks but depends on knowledge people already produced, so it recombines what we know rather than creating new understanding. He puts it as “a sophisticated remix of what we know already”. The conclusion he draws is to keep building people’s critical thinking and curiosity, rather than handing judgment to machines. We did not hear the audio or read the article, and the page gives no data.

For Kilimora: this is the design principle for any AI in AgriKonnekt. A farmer educator who knows the local soil should be able to overrule a recommendation, and a cooperative member should be able to ask why. The Nigerian cassava example is the warning: advice that is statistically sensible and locally useless. Curriculum and decisions staying with member led organisations, as Kilimora says they will, is the human capital half of the same bargain.

Three other signals

More chicken, more waste stream. Our World in Data reports that the world average has risen from about two chickens slaughtered per person a year in 1961 to about five in 1990 and nine in 2022, with birds also getting heavier. That is a food demand story, and a welfare story, since most farmed animals are raised in factory farms. It is also a supply story for anyone using abattoir by products: AgriKonnekt’s organic fertiliser uses blood and bone meal from slaughterhouses as a nitrogen source. We have not measured how much of that stream cooperatives can reach.

Chickens eaten per person per year

Chickens slaughtered for meat per person, world average. Our World in Data, from FAO

1961about 2about 21990about 5about 52022about 9about 9

A quiet average hides a bad place. Human Progress reports that no hurricane made landfall in the continental United States in 2025, the first such year since 2015, and that hurricane intensity across the Northern Hemisphere was 19% below average even though the North Atlantic’s was 9% above. Jamaica was hit by Melissa in the same season. National and hemispheric averages are poor guides to a single farm, which is the argument for plot level records.

A quiet year on average, a devastating one in Jamaica

Hurricane intensity in 2025 compared with the 1991 to 2020 average, per cent. Human Progress roundup of 7 December 2025

Change in intensityNorth Atlantic+9%+9%Whole NorthernHemisphere−19%−19%

Drones are arriving. The same roundup notes that agricultural drones are spreading, especially across East and Southeast Asia, mainly to spray agrochemicals, a job still done by hand in many poorer countries. For a cooperative promoting organic inputs, drones could cut exposure to chemicals; they could also lower the cost of conventional spraying. It is a trend to watch, not yet a result.

What it means for AgriKonnekt

Kilimora’s plan is ambitious in all three of the dimensions these reports discuss. It aims to grow from 1,500 farmers across six Kenyan counties in 2026 to 50,000 in three countries by 2030, from 24,000 verified tonnes of annual carbon potential to 180,000, and from $1.2 million of finance in 2025 to $50 million a year. Those are 33, 7.5 and 42 times today’s levels. Every one of them multiplies what has to be measured, and the money is only as good as the measurement behind it.

The scale AgriKonnekt plans to verify

Growth multiples implied by AgriKonnekt’s own stated numbers. Targets, not results

Farmers1,500 to 50,000×33×33Verified carbon24,000 to 180,000 t×7.5×7.5Finance a year$1.2M to $50M×42×42

The Gender Lens framework adds targets for women: more than 60% of AgriKonnekt participants, 7,500 women enrolled with individual payment accounts by 2028 and about 5,250 of them active, and women in over 60% of 5,000 direct roles. These are published targets, not results, and the framework promises independent annual evaluations with control groups and public reporting of all findings.

Who the targets are for

Left: women’s intended share of 5,000 direct roles. Right: women to be enrolled by 2028 and the modelled active share. Targets, not results

60%+Up to 40%5,000direct roles60%+Women, as targetedUp to 40%Everyone else(our arithmetic)~5,250~2,2507,500women enrolled~5,250Active users(about 70%, modelled)~2,250Enrolled, not yetactive (our arithmetic)

What the three reports add is a way to hold those targets to account. A Scorecard style page with a handful of indicators, a clear account of the money by instrument and its price, and a plain description of where AI is and is not used, would make the plan checkable. The table below lists what we can and cannot say today.

What we can and cannot claim

StatementEvidence status
Dual layer verification costs about 70% less on a 1,500 farmer pilotPilot estimate; not independently audited
96% documented order fulfilmentKilimora statement; supporting records not published
AI on the sensor works without a cloud connectionDesign claim; no published accuracy results
Farmers are paid directly within 24 hours of funder releaseDesign claim; no payment time report yet
More than 60% of participants and recipients are womenDesign statement; no payments by gender report yet
50,000 farmers, 180,000 verified tonnes and $50M a year by 2030Targets, not results
Reaching farmers by phone makes advice cheaper and as goodUntested for AgriKonnekt; the Bank also calls its own chatbot figures untested

What we would ask for

  1. Publish a Scorecard style page for AgriKonnekt: about ten indicators, with definitions, baselines, targets and a season by season update, including the ones that lag.
  2. Separate reach from benefit: farmers enrolled, farmers verified, and farmers whose income rose.
  3. Report the money by instrument (grant, results based payment, concessional loan, commercial loan, input credit) with the price of each.
  4. Show input credit instalments as a share of members’ harvest income, so the cost of credit is visible.
  5. Test any AI advice on local crops and languages before scaling, publish error rates, and keep farmer educators able to overrule it.
  6. Measure the data and time each lesson costs a member, and keep a USSD route for every digital service.
  7. Report women’s results by gender and household role, using the household level lens from our companion article.
  8. With members’ consent, share plot data openly enough to feed the public data exchanges the Bank recommends.

A reading list from the same month

The Bank’s November 2025 working paper list included several titles relevant to this work. We have not read them.

Working paper, November 2025Why it is on our listStatus
Women’s Mobility and Labor Supply: Experimental Evidence from Pakistan (Garlick, Field, Vyborny)Links to our companion article on household constraints on women’s workTitle only; not read
Labor Demand in the Age of Generative AI: Early Evidence from US Job Posting DataShows how AI is already changing which skills employers ask forTitle only; not read
Monitoring Global Aid Flows: A Novel Approach Using Large Language ModelsA direct parallel to measuring climate finance flows to cooperativesTitle only; not read
The Distributional Impacts of Climatic Variability on Welfare in ThailandWhose welfare moves when weather variesTitle only; not read
Sources & notes

SOURCES: World Bank, Scorecard insights, FY2025, Scorecard data update, October 2025, debt data blog, International Debt Report 2025 (landing document; we relied on the blog and coverage), Harnessing AI for Agricultural Transformation, Digital Progress and Trends fact sheet, agriculture case study, AI in Sub Saharan African farming; Bretton Woods Project on the Scorecard (headline only); Engineering News on the debt report; IMF, podcast page, Pablo Peña; Kenya, The Standard on public debt and People Daily on the Treasury interview; Accion on Lersha; AgFunder News on HerVest; Jamaica Gleaner on the recovery package; Our World in Data (chicken data insight); Human Progress (Doomslayer roundup, 7 December 2025); Kilimora, AgriKonnekt, Gender Lens Programming, Closing the verification gap, Who pays when the rain fails?. The World Bank newsletters of 28 November and 5 December 2025, the Policy Research Working Papers list, the IMF update email and the AgriTech Digest are emails without links; the newsletter claims we use are as stated there. Differences between sources: the Bank’s December newsletter says 50% of people in the 22 most indebted countries cannot afford the minimum diet, while report coverage gives an average of 56%; internet use in low income countries is 27% in the fact sheet and about 40% in the case study; Kenya’s debt is Sh11.7 trillion per the Controller of Budget and about Sh11.5 trillion per a Treasury official; the IMF page dates the podcast 24 November, the IMF email 2 December. Our own arithmetic: the debt level a decade earlier, complements in the doughnuts (everyone else, the rest), the IDA remainder of $7.7 trillion, the share of expected reach (32% to 50%), growth multiples, and the women and active user splits. The “nearly two thirds”, “nearly 60%” and “nearly half” doughnut and bar values are drawn at approximate sizes. Kilimora statements about its systems and targets are the company’s own.