Karlan et al. (2014) Insurance, credit and agricultural investment Impact evaluationPaper detailsFull title: Agricultural Decisions after Relaxing Credit and Risk Constraints Review coverage: Existing detailed local extraction retained; source version unchanged. Read the paper or report ↗ |
Finance Insurance and other risk sharing Cash grants and rainfall-index insurance offers, separately and together. No AI in interventionIntervention detailsAI in this study: No AI evaluated; insurance and cash transfers. Possible AI contribution: While no AI was used, AI could help identify similar farming areas for insurance pricing and crop-loss assessment. Effects on premiums and protection would need testing. Other relevant problems: |
Northern Ghana Smallholder farm households Sample: 502 households initially; expanded to 1,406. Population detailsSample: 502 households initially; expanded to 1,406. Cultivation outcomes cover two seasons. Who was included: Predominantly male decision-makers in the study. To assess: access, uptake, and protection for women, poorer farmers, and farmers without adequate coverage or data. Study period: Study timeframe: 2009–2011; cultivation outcomes cover two seasons. Further follow-up details: Not yet extracted. |
Randomized trial Finding: Rainfall insurance increased cultivation investment; cash grants alone did not significantly increase total investment. Key limitation: Profit gains were not statistically established; results cover two cultivation seasons and predominantly male decision-makers. Study details |
Delivery: Insurance was offered in a randomized field study over two cultivation seasons. Main gap: Uptake, full delivery costs and performance beyond the study setting need further review. Scalability details- Sustained use Not yet reviewed
The sample counts households assigned to the study, not sustained insurance users. Uptake and renewal evidence still need review. Pages: pp. 604–615 - Costs and funding Not yet reviewed
A complete account of insurance administration, data collection and claims costs has not yet been reviewed. - Delivery Reported
The study delivered insurance offers and grants. Wider provision would require premium collection, claims administration and reliable rainfall data. Pages: pp. 604–615 - Use in other settings Not yet reviewed
Performance under different rainfall risks, farming systems and insurance arrangements has not yet been reviewed. - Who benefits Measured
Study decision-makers were predominantly men; the findings do not establish equal access or protection for women. Pages: pp. 604–615 - Lasting benefits and wider effects Measured
Cultivation outcomes cover two seasons. Profit gains were not statistically established; longer-term protection and investment remain uncertain. Pages: Tables IV and VI, pp. 625–633
|
Burke et al. (2019) Seasonal credit and maize sales Impact evaluationPaper detailsFull title: Sell Low and Buy High: Arbitrage and Local Price Effects in Kenyan Markets Review coverage: Existing detailed local extraction retained; source version unchanged. Read the paper or report ↗ |
Finance Seasonal liquidity and credit Seasonal loans allow farmers to store maize and choose when to sell. No AI in interventionIntervention detailsAI in this study: No AI evaluated; seasonal lending. Possible AI contribution: While no AI was used, AI could help lenders assess repayment risk for farmers with limited credit histories. Effects on access, interest rates and exclusion would need testing. Other relevant problems: |
Western Kenya · maize Existing NGO clients who farm maize Sample: 1,589 farmers in year 1; 1,019 in year 2, including 602 returning participants. Population detailsSample: 1,589 farmers in year 1; 1,019 in year 2, including 602 repeat participants. Seventeen sublocations and 52 markets. Who was included: Existing NGO clients in two districts. To assess: eligibility and loan use among non-clients, women, and farmers without usable phone data. Study period: Study timeframe: 2012–2014. The existing record reports two participant cohorts and notes that longer-run revenue gains were not statistically established. Further follow-up details: Not yet extracted. |
Randomized trial Finding: Seasonal credit increased storage and net maize revenues. Effects on consumption and longer-run revenues were not statistically established. Key limitation: Participants were existing clients in two districts. Net maize revenue does not deduct all farm production costs. Study details |
Delivery: Loans were tested through an existing service network over two years. Main gap: Delivery costs and access for farmers outside the existing client network need review. Scalability details- Sustained use Not yet reviewed
Trial participation and repeat survey coverage are documented. Loan take-up and sustained borrowing have not yet been separately reviewed. Pages: pp. 796–801 - Costs and funding Not yet reviewed
Complete loan administration and servicing costs have not yet been reviewed. - Delivery Reported
The experiment used an existing NGO client network. Replication requires lending capacity, appropriate seasonal terms and access to grain storage. Pages: pp. 796–801 - Use in other settings Not yet reviewed
Results beyond the two study districts and outside the existing client network have not yet been reviewed. - Who benefits Reported
Recruitment covered existing NGO clients; it does not establish access for farmers outside that network. Pages: pp. 796–801 - Lasting benefits and wider effects Measured
Longer-run revenue gains were not statistically established. The paper also estimates local price effects, which need fuller review before assessing wider impacts. Pages: p. 822; Table VI, p. 825
|
Abate et al. (2023) Video-supported agricultural extension Impact evaluationPaper detailsFull title: Accelerating technical change through ICT: Evidence from a video-mediated extension experiment in Ethiopia Review coverage: Existing detailed local extraction retained; source version unchanged. Read the paper or report ↗ |
Institutions and public delivery Extension and service-delivery capacity Public extension agents screen locally produced farming videos and lead discussions. No AI in interventionIntervention detailsAI in this study: No AI evaluated; agent-facilitated video extension. Possible AI contribution: While no AI was used, AI could help translate and prepare locally relevant video scripts, potentially reducing preparation time and cost. Agronomic review would remain necessary. Other relevant problems: |
Ethiopia · teff, wheat and maize Farm households reached through public extension Sample: 2,345 households retained in 342 villages; outcome sample sizes vary. Population detailsSample: Initially selected: 350 kebeles and 2,450 households. First survey: 347 kebeles and 2,422 households. Retained panel: 342 kebeles and 2,345 households (77 lost between surveys). Adoption regressions use 2,345 observations in year 1 and 2,119 in year 2; the main text does not explain the additional second-year exclusions. Who was included: Approximately 90% of sample households were male headed. Separate spouse sessions improved some early access/knowledge measures but did not establish additional adoption gains. Videos were not tailored to women; transfer to other excluded groups remains untested. Study period: Two meher seasons, 2017/18 and 2018/19, with household surveys in January–March 2018 and January–March 2019. Video delivery continued in both seasons; there is no post-withdrawal persistence test. Long-run effects and possible novelty effects remain unresolved. |
Randomized trial Finding: Adding videos increased reported adoption of recommended practices by about 6–7 percentage points over two seasons. Sustained yield gains were not established. Key limitation: Delivery continued in both seasons; adoption was self-reported, and selected districts limit generalization. Study details |
Delivery: Videos were delivered through public extension across four regions for two seasons; lower costs at wider coverage were projected. Main gap: Continued benefits after support ends and complete service costs remain uncertain. Scalability details- Sustained use Measured
Videos were screened in 85% of treatment villages in year 1 and 83% in year 2. These percentages measure village delivery, not individual viewing. Pages: p. 6, footnote 13 - Costs and funding Projected
The paper reports $16–30 per additional adoption during the trial and projects $3–6 at full coverage, above existing extension costs. Detailed cost calculations were not available for review. Pages: pp. 10 and 12; Appendix H unavailable - Delivery Reported
Delivery relied on trained public agents, local videos, farmer groups and working projectors. Projector failures and staff turnover disrupted implementation. Pages: pp. 4–7 - Use in other settings Measured
Implementation covered four regions, but districts were selected for expansion. Transfer depends on local content, staffing and input access. Pages: pp. 4–7 - Who benefits Measured
About 90% of sampled households were male-headed. Separate sessions for spouses did not establish additional adoption gains. Pages: Table 1; Table 4, p. 8 - Lasting benefits and wider effects Measured
Adoption gains continued during two seasons of delivery. The study did not test continuation after videos and support were withdrawn. Pages: Table 4, p. 8; Table 7, p. 11
|
Fabregas et al. (2025) Agronomic SMS: six experiments Impact evaluationPaper detailsFull title: Digital Information Provision and Behavior Change: Lessons from Six Experiments in East Africa Review coverage: Existing detailed local extraction retained; source version unchanged. Source note: Published paper metadata verified online; detailed claims cross-checked against existing extracted records. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge Human-written SMS recommend locally relevant lime, fertilizer and farm-management practices. No AI in interventionIntervention detailsAI in this study: No AI evaluated; agronomic SMS. Possible AI contribution: While no AI was used, AI could tailor SMS messages to crops, local conditions and farmers’ questions, and translate them into local languages. Added benefits and costs would need testing. Other relevant problems: |
Kenya and Rwanda · crop inputs Farmers recruited through public agencies, agrodealers and existing service networks Sample: Six experiments; the largest direct-effect analysis includes 82,873 farmers. Population detailsKALROSample: 832 farmers randomized; 773 in the final Table 2 analysis sample. The lime-purchase logit uses 674 observations; the linear-probability model uses 773. Who was included: Eligibility required phone ownership and responsibility for farming; roughly two-thirds of participants were women. Effects for farmers outside this eligible population remain uncertain. Study period: 2015 short-rains survey and 2016 long-rains coupon purchases. No second-season advisory texts were sent. IPA/PxD1-KSample: 1,897 farmers in the trial; the initial lime-compliance logit uses 1,854 purchase-record observations and the linear-probability model uses 1,897. Who was included: Existing NGO or contract-farming contacts able to participate by phone; approximately 37% were women. This is not a representative sample of all smallholders. Study period: Two seasons: 2016 short rains and 2017 long rains. Treatment continued in the second season, so that comparison is not untreated persistence. IPA/PxD2-KSample: 5,890 farmers in the trial; the main lime-compliance logit uses 5,732 purchase-record observations and the linear-probability model uses 5,890. Who was included: Agrodealer customers with eligible farms and available local soil data; approximately 34% were women. Farmers outside input markets or without usable data may face different barriers. Study period: Purchase records cover the 2017 long rains; surveys also collected subsequent-season behavior. No new advisory texts were sent in that following season. 1AF1-KSample: 4,884 farmers in the initial experiment: 3,325 assigned to messages and 1,559 controls. A linked second-season randomization covered 2,931 returning clients. Who was included: Previous One Acre Fund clients with usable phone contacts; approximately 65% were women. Nonclients and farmers outside the service network were not the target population. Study period: Input records for 2017 and 2018 long rains; 2,931 returning clients were rerandomized. Follow-up estimates therefore concern a selected, linked sample. 1AF2-KSample: 32,572 previous One Acre Fund clients in the initial analysis sample, including controls. Later observations follow returning clients rather than a new independent cohort. Who was included: Previous service clients; approximately 69% were women. Performance among nonclients, farmers without usable phone contacts, and those unable to obtain inputs remains uncertain. Study period: 2018 and 2019 long-rains input records. All returning farmers received lime messages for 2019, ending the untreated lime comparison; fertilizer persistence can be assessed separately. 1AF3-RSample: 82,873 farmers in 17,850 groups in the direct-effect analysis. The broader frame contained 202,972 farmers in 20,944 groups; 110,400 had registered phones. These are distinct denominators. Who was included: The main direct estimate is restricted to registered phone holders. Farmers without phones were considered in separate spillover analyses, not included in that direct-effect denominator. Study period: 2018 and 2019 main-season input purchases. Returning 2018 clients were rerandomized; follow-up is selected and linked to the original experiment. |
Six randomized trials Finding: Across six experiments, following input recommendations increased by about 2 percentage points (95% confidence interval: 1–3). Effects varied across trials. Key limitation: The paper measures input decisions, not income gains. More detailed advice and telephone assistance did not clearly improve purchases beyond simpler SMS. Study details (6 experiments) |
Costs: The paper models SMS costs and benefits; trials document delivery through established farmer and input-service networks. Main gap: Sustained reading, standalone service costs and measured income gains remain unresolved. Scalability details- Sustained use Not reported
Trial samples include control farmers and are not counts of active readers. The study could not directly observe whether texts were read. Pages: Tables 1–2, pp. 534–535; Appendix J - Costs and funding Projected
Benefits and benefit–cost ratios are modeled from input decisions, assumed agronomic responses and prices. Low SMS transmission costs do not cover the full cost of a standalone service. Pages: pp. 557–559; Appendix G, pp. 25–29 - Delivery Reported
Delivery used farmer phone lists, agronomic content and established agencies or service networks. Several trials also provided input discounts, credit or distribution. Pages: pp. 533–541; Appendix J, pp. 39–51 - Use in other settings Measured
Six experiments span Kenya and Rwanda, with different recruitment and delivery arrangements. Effects vary; pooled gains do not guarantee results in each new setting. Pages: Figure 2, p. 547; Tables 1–2 - Who benefits Measured
Direct effects generally concern farmers with usable phone contacts. The Rwanda experiment separately studies information sharing to farmers without registered phones. Pages: p. 557; Appendix J.3.3, pp. 49–51 - Lasting benefits and wider effects Measured
Repeated messages helped, but continued purchase effects without renewed messages were not clearly established. Follow-up often involved returning service clients. Pages: Section IV; supplemental Tables D4 and F1
|
Mrisho et al. (2020) Nuru smartphone diagnosis of cassava diseases Technical validationPaper detailsFull title: Accuracy of a Smartphone-Based Object Detection Model, PlantVillage Nuru, in Identifying the Foliar Symptoms of the Viral Diseases of Cassava–CMD and CBSD Review coverage: Existing detailed local extraction retained; source version unchanged. Source note: Full publisher text inspected; matches existing local record. Read the paper or report ↗ |
Information and knowledge Diagnosis and interpretation A smartphone vision model recognizes visible cassava disease symptoms, including through multiple-leaf checks. AI usedIntervention detailsAI in this study: Smartphone computer vision for visible cassava symptoms. Other relevant problems: |
Tanzania and Kenya · cassava Cassava plants, farmers and extension officers Sample: 170 benchmark images; field tests: 50 plants (300 leaves). Population detailsSample: 170-image benchmark; 60 human assessors in six groups. Methods describe 50 plants/300 leaves for field testing. The separate teaching assessment included 30 extension officers and 50 farmers. Who was included: To assess: performance across phone quality, varieties, symptom severity, languages, literacy, gender, and farmers outside the evaluated groups. Study period: The teaching exercise lasted two weeks. The model comparison used 2018 and 2020 releases, not a two-year panel of farmer outcomes. |
Field validation Finding: On a 170-image benchmark, the 2020 Nuru version averaged 65% accuracy, compared with 49% for trained extension officers. Key limitation: Technical accuracy does not establish better disease management, yields or incomes. Absence of visible symptoms cannot certify absence of infection. Study details |
Field performance: Image and field tests support symptom recognition; the teaching exercise lasted two weeks. Main gap: Delivery costs need further review; sustained use and farm benefits were not established. Scalability details- Sustained use Not reported
Benchmark and teaching samples are evaluation observations, not counts of regular users. No sustained-use estimate is established. Pages: Methods, pp. 3–5; teaching assessment, p. 9 - Costs and funding Not yet reviewed
A complete estimate of service costs has not yet been reviewed. Offline operation alone does not establish low total costs. - Delivery Reported
The app can work offline. Practical delivery still needs compatible phones, training, usable images and access to appropriate disease management. Pages: Methods, pp. 3–5 - Use in other settings Measured
Performance varies by symptom and assessment method. Each model version needs validation in the intended crop, disease and field setting. Pages: Tables 1–4, pp. 6–9 - Who benefits Not yet reviewed
Access and performance by phone quality, literacy, gender and language have not yet been systematically reviewed. - Lasting benefits and wider effects Not reported
The two-week teaching exercise does not establish lasting use or farm benefits. Effects on disease spread to neighboring farms were not evaluated. Pages: Teaching assessment, p. 9
|
Adjéi et al. (2024) Continued use of Nuru after a supported pilot Implementation studyPaper detailsFull title: Perception and adoption by cassava farmers of the PlantVillage Nuru application disseminated in the agricultural environment of Côte d’Ivoire: a case study in the departments of Dabou, Bouaké, and Man Review coverage: Existing detailed local extraction retained; source version unchanged. Read the paper or report ↗ |
Information and knowledge Diagnosis and interpretation Nuru smartphone diagnosis supported by training, extension services and some supplied smartphones. AI usedIntervention detailsAI in this study: Observed use of an AI diagnostic application; no causal AI effect identified. Other relevant problems: |
Côte d’Ivoire · cassava Farmers previously exposed to a Nuru pilot Sample: 146 farmers across three departments; six staff interviews. Population detailsSample: 146 farmers: 44 in Dabou, 46 in Bouaké, and 56 in Man; six staff interviews. Sampling used the platform user list and depended on availability and willingness to participate. Who was included: Pilot selection favored literate farmers; 40% of surveyed farmers owned smartphones. To assess access for excluded farmers and whether device and support arrangements narrow participation gaps. Study period: Survey in August 2023. Platform trends cover six semesters in Dabou and Bouaké and three in Man; this is not a controlled longitudinal impact study. |
Implementation study Finding: Use differed across three pilot areas and became less frequent after an initial rise. Farmers described phone-access and support barriers. Key limitation: Participants were selected from pilot user lists. Reported continuation rates contain inconsistencies, and causal farm benefits were not assessed. Study details |
Sustained use: Usage trends and interviews identify declining use and barriers related to phones and continuing support. Main gap: Delivery costs need further review; population-wide adoption and causal farm benefits were not established. Scalability details- Sustained use Measured
Platform trends show an initial rise followed by declining usage frequency. The surveyed pilot users do not represent all farmers in the study departments. Pages: Sections 3.3 and 3.5, pp. 5–10 - Costs and funding Not yet reviewed
Complete costs of phones, training, technical support and continued operation have not yet been reviewed. - Delivery Reported
Farmers described barriers involving phone access, installation, training and continued assistance. Some pilot participants received phones. Pages: Sections 3.7–3.8, pp. 10–12 - Use in other settings Not reported
Three pilot areas show different use patterns. The study does not establish national adoption under routine delivery conditions. Pages: Sections 2.1–2.3 and 3.3 - Who benefits Measured
Pilot selection favored literate farmers, and 40% of surveyed farmers owned smartphones. These features limit conclusions about access for excluded farmers. Pages: Sections 2.1–2.3 and 3.7–3.8 - Lasting benefits and wider effects Measured
Platform trends cover several semesters, but the survey is not a controlled long-term impact evaluation. Farm outcomes and effects on nonusers were not evaluated. Pages: Sections 3.5–3.8, pp. 7–12
|
Singh et al. (2024) Farmer.Chat: multilingual AI agricultural advice Implementation studyPaper detailsFull title: Farmer.Chat: Scaling AI-Powered Agricultural Services for Smallholder Farmers Review coverage: Existing detailed local extraction retained; source version unchanged. Source note: Version 2 primary full text inspected and compared with existing extraction. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge A generative-AI chatbot provides multilingual advice using curated agricultural information and user support. AI usedIntervention detailsAI in this study: Generative AI with retrieval, speech and translation in an advisory service. Other relevant problems: |
Kenya, India, Nigeria and Ethiopia Farmers, lead farmers and agricultural service staff Sample: 15,000+ users reported; detailed experience research centers on selected Kenyan users. Population detailsSample: Authors report 15,000+ users across four roles and more than 300,000 queries by the 2024 report. Kenyan experience studies include 199 users in January focus groups, 20 women lead farmers in April, seven interviewees and seven usability participants; overlap is not resolved. Context-retrieval evaluation uses over 1,000 queries. These are separate denominators. Who was included: Kenyan qualitative samples were selected among users, often active users; they cannot represent excluded farmers. Participants reported smartphone access, data affordability, language, and digital-literacy barriers. Women-focused research and voice interfaces inform design, but do not establish equal access or benefits. The paper notes remaining local-dialect limitations (Sections 5.1–5.3, pp. 11–18; Section 8, p. 32). Study period: Kenya launched in October 2023. User studies took place in January, February, April, and July 2024; biweekly phone feedback began in August 2024. Platform analysis covers differing periods, with no uniform observation window or longitudinal farmer-outcome follow-up established (Sections 4.1 and 5, pp. 10–18). |
Implementation study Finding: The report documents use across four countries and evaluates answer quality and user experience. It does not establish effects on yields or incomes. Key limitation: Implementers authored the study. Selected users and internal quality measures do not establish population-wide benefits or independent agronomic accuracy. Study details |
Reported reach: Implementers report deployment in four countries and describe local content, language services and user support. Main gap: Sustained use, full costs and farmer benefits are not established. Scalability details- Sustained use Reported
The report states 15,000+ users and over 300,000 queries. Accounts and queries do not equal unique farmer beneficiaries; country-level use is uneven. Pages: Sections 4–4.2, pp. 9–11, v2 PDF - Costs and funding Not reported
No full Farmer.Chat operating budget or cost per user is provided. Cost figures for SMS and video extension refer to other services. Pages: Section 2; Sections 3–4, v2 PDF - Delivery Reported
Delivery requires updated local content, AI and speech services, mobile access, onboarding, support and answer-quality monitoring. Their combined recurring costs are not estimated. Pages: Sections 3–4, pp. 4–10, v2 PDF - Use in other settings Reported
Deployment spans four countries, but detailed user research centers on Kenya. Independent validation across crops, seasons and languages remains necessary. Pages: Sections 5–6 and 8, v2 PDF - Who benefits Reported
Users described phone, data-cost, literacy and language barriers. Women-focused research and voice interfaces inform design but do not establish equal access or benefits. Pages: Sections 5.1–5.3, pp. 11–18; Section 8, p. 32 - Lasting benefits and wider effects Not reported
The paper does not establish long-term farmer outcomes or continued operation after external support. Advice sharing is described but indirect benefits are not measured. Pages: Sections 4–5, pp. 9–18, v2 PDF
|
Masiwal et al. (2026) AI forecasts of Indian monsoon onset Technical validationPaper detailsFull title: Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon Review coverage: Existing detailed local extraction retained; source version unchanged. Read the paper or report ↗ |
Biophysical hazards and resource base Weather and climate exposure Compare six AI weather models for predicting when the Indian monsoon begins. AI usedIntervention detailsAI in this study: Technical benchmark of AI weather forecasts for monsoon timing. Other relevant problems: |
India · monsoon rainfall Weather-model forecasts; farmers are the intended users Sample: Headline results cover ten 4° × 4° grid cells; evaluation periods differ by model. Population detailsSample: Forecasts for India at 4° × 4° resolution, with headline summaries over ten core-monsoon-zone cells. Main recent period: 2019–2024; extended period adds 1965–1978; further checks use 1965–2024 and 2004–2021. Available years differ by model; AIFS training extends through 2022. These are weather observations, not a farmer sample. Who was included: Regional forecasts are not plot-specific. The dissemination domain excluded locations with frequent post-onset dry spells or high within-cell onset variability. Differential access, comprehension and benefits by gender, poverty or phone access were not evaluated. Study period: Historical hindcasts and a 2025 operational forecast case; no longitudinal farmer follow-up. Training overlap and missing years differ across models, so the recent period is not wholly out of sample for every model. |
Forecast validation Finding: Most models improved onset predictions over historical averages at 1–15-day lead times. Performance weakened at longer horizons; no model was consistently best. Key limitation: Evaluation years and training overlap differ by model. The study does not measure farmers’ use or benefits. Study details |
Forecast performance: The benchmark informed model choice for a government forecast service. Main gap: Full service costs, sustained use and farmer outcomes were not assessed. Scalability details- Sustained use Not reported
The paper refers to the same 2025 government dissemination as Aitken et al. It supplies no separate user sample or measured rate of reading or continued use. Pages: pp. 2–6; Extended Data Table 2, p. 17 - Costs and funding Not reported
Computational demands are discussed, but full service costs and costs per active farmer are not established. Pages: Methods, pp. 6–11 - Delivery Reported
Operational forecasts need accessible models, current atmospheric inputs, local rain observations, computing capacity and institutions able to communicate usable forecasts. Pages: Methods, pp. 6–11 - Use in other settings Measured
Forecasts are tested across Indian regions and periods. Performance varies with lead time and local climate; each new rainfall regime needs validation. Pages: Figures 2–4, pp. 14–16 - Who benefits Not reported
Forecasts are regional rather than plot-specific. Access, comprehension and benefits by gender, poverty or phone access were not evaluated. Pages: Methods, pp. 6–11 - Lasting benefits and wider effects Not reported
Historical forecast tests and one operational season do not establish continued farmer benefits or effects on nonrecipients. Pages: Extended Data Table 2, p. 17
|
Aitken et al. (2026) Blending AI forecasts for monsoon advice Technical validationPaper detailsFull title: Designing probabilistic AI monsoon forecasts to inform agricultural decision-making Review coverage: Existing detailed local extraction retained; source version unchanged. Version: Survey located v2; numerical study extraction and source URL remain v1. Do not combine versions without re-extraction. Read the paper or report ↗ |
Biophysical hazards and resource base Weather and climate exposure Blend two AI weather models with updated historical rainfall expectations to produce weekly onset probabilities. AI usedIntervention detailsAI in this study: AI forecasts blended with updated climatological expectations. Other relevant problems: |
India · monsoon rainfall Weather forecasts linked to government farmer messaging Sample: 32 selected 2° × 2° grid cells; approximately 38 million forecast recipients reported separately. Population detailsSample: 32 selected 2° × 2° Indian grid cells used to train the blend; AIFS forecasts and 30-member NeuralGCM ensembles, initialized twice weekly from May until local onset. Evaluation periods: 2000–2024, 1965–1978 and 2025. Approximately 38 million recipients is the reported dissemination population, not the statistical evaluation sample. Who was included: The 32 cells were selected partly for model skill, dry-spell risk and within-cell onset variability; forecasts use weekly bins and 2° cells. The evolving-expectations baseline models shared information, not every farmer’s private beliefs or constraints. Distributional effects and access gaps were not measured. Study period: Historical cross-validation and earlier hindcasts, plus one season of operational forecasts in 2025. No farmer panel or multi-season impact follow-up is reported. |
Forecast validation and deployment Finding: Blended forecasts improved onset predictions and supported a 2025 government dissemination program. The reported approximately 38 million recipients are not measured active users or beneficiaries. Key limitation: No causal farmer-outcome evaluation. The reported dissemination is shared with Masiwal et al.; it must not be counted twice. Study details |
Reported reach: Government dissemination reportedly reached approximately 38 million farmers in 2025. Main gap: Reading, sustained use, farmer benefits and full costs remain not yet established. Scalability details- Sustained use Reported
Approximately 38 million forecast recipients are reported for 2025. This does not measure reading, understanding, adoption or repeated use, and overlaps with Masiwal et al. Pages: pp. 2–3; p. 12 - Costs and funding Not reported
The paper does not establish a full operating budget, cost per active user or cost per farmer benefited. - Delivery Reported
The service combined weather data and computing with government messaging channels, with collaboration from India’s agriculture ministry, DIL and PxD. Pages: pp. 2–3 and 9–12 - Use in other settings Measured
The blend was evaluated across selected Indian cells and historical periods. Results do not establish performance in other countries or rainfall regimes. Pages: pp. 4–8; Figures 2–4, pp. 14–15 - Who benefits Not reported
Forecast areas were selected partly for model skill and local rainfall patterns. Access gaps and distribution of benefits were not measured. Pages: Methods, pp. 6–8 - Lasting benefits and wider effects Not reported
The paper covers one operational season and historical forecasts. It does not establish multi-season farmer outcomes or effects on nonrecipients. Pages: p. 12; Extended Data Figures 2 and 4, pp. 17–18
|
Fabregas et al. (2019) Digital agricultural advice: evidence and opportunities ReviewPaper detailsFull title: Realizing the potential of digital development: The case of agricultural advice Review coverage: Full text and bibliographic metadata inspected, 2026-11-10. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge Review mobile and other digital approaches to delivering agricultural advice. AI discussedIntervention detailsAI in this study: Digital information; does not isolate AI effects. Possible AI contribution: AI could help tailor and translate advice and respond to farmers’ questions. These are potential extensions, not effects isolated by this review. Other relevant problems: |
Low- and middle-income countries Smallholder farmers in digital-advisory evaluations |
Review and meta-analysis Finding: The synthesis estimates average yield gains of 4% and 22% higher odds of adoption across the included digital-advice studies. Key limitation: Average effects combine different services and settings. Underlying studies overlap with other entries in this table. Study detailsStudy design: Synthesis of digital advisory evaluations; not an additional independent intervention. Findings: The synthesis estimates average yield gains of 4% and adoption odds 22% higher across its included digital-advice evidence; it motivates testing timely, localized advice and feedback systems. Limits: Averages combine heterogeneous interventions. Historical mobile-access statistics are not current access estimates. Do not count underlying experiments again as additional impact evidence. Pages: https://pmc.ncbi.nlm.nih.gov/articles/PMC10859166/; abstract and discussion. Advisory-panel research seed. |
Review: Synthesizes several advisory services. It highlights costs, customization and public support, rather than assessing one intervention’s scalability. |
Cole and Fernando (2021) Voice-based agricultural advice Impact evaluationPaper detailsFull title: ‘Mobile’izing Agricultural Advice: Technology Adoption, Diffusion and Sustainability Review coverage: University author repository metadata and abstract inspected; full methods remain to extract. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge A mobile voice service delivers agricultural advice and answers farmers’ questions. No AI in interventionIntervention detailsAI in this study: Non-AI voice information service. Possible AI contribution: While no AI was used, AI could help answer routine questions and translate voice advice, potentially reducing staff time and reaching more farmers. Accuracy and total costs would need testing. Other relevant problems: |
Gujarat, India · cotton Smallholder cotton farmers |
Randomized trial Finding: Farmers changed information sources and adopted recommended cotton inputs. Systematic yield or profitability gains were not established. Key limitation: One crop and region; detailed methods and outcome estimates still need review. Study detailsStudy design: Two-year randomized evaluation of voice-based advisory. Counterfactual: Farmers not offered the voice-advisory service in the randomized evaluation. Findings: Farmers changed information sources and adopted recommended cotton inputs; systematic yield or profitability gains were not established. Limits: Study-specific crop, location and delivery conditions constrain transfer. Positive adoption and stated trust should not be translated into farmer-income effects. Pages: https://dash.harvard.edu/entities/publication/45ab9abb-1d3b-4e18-9465-b8eebb99d674; Economic Journal 131(633), 192–219. |
Costs: Average willingness to pay was below study operating cost; lower costs with more users were projected. Main gap: A financially sustainable service at wider scale was not demonstrated. Scalability details- Sustained use Not yet reviewed
The evaluation lasted two years. Detailed use, retention and renewal measures have not yet been reviewed. - Costs and funding Projected
The paper reports willingness to pay below operating cost and anticipates lower costs with more users. The latter is a projection, not observed financial sustainability. Pages: Abstract; Economic Journal 131(633), pp. 192–219 - Delivery Reported
Advice was delivered by a mobile voice service. Staffing, content preparation and support requirements need fuller review. Pages: Abstract - Use in other settings Not yet reviewed
Evidence concerns cotton farmers in Gujarat. Performance across other crops and locations has not yet been reviewed. - Who benefits Not yet reviewed
Access and benefits by gender, phone ownership and literacy have not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
The two-year design is documented, but continued use and outcomes after support ended have not yet been reviewed.
|
Cole et al. (2025) Digital rice advice and weather shocks Impact evaluationPaper detailsFull title: The impact of digital agricultural extension service: Experimental evidence from rice farmers in India Review coverage: Primary full text inspected; version February 5, 2025. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge An established voice-advisory service supplies crop and weather information to farmers. No AI in interventionIntervention detailsAI in this study: Advisory is non-AI; machine learning estimates yields for evaluation, not treatment delivery. Possible AI contribution: While no AI was used to deliver the advice, AI could help tailor messages and answer farmer questions. Machine learning was used to estimate yields for the evaluation. Other relevant problems: |
Odisha, India · rice Rice farmers in five districts Sample: 13,675 farmers in two trial cohorts. |
Randomized trial Finding: Knowledge and adoption improved. Crop-loss and production effects varied with weather shocks; average profit effects were imprecisely estimated. Key limitation: Profit data cover one season. Projected benefits depend on how often weather shocks occur. Study detailsStudy design: Randomized service rollout among 13,675 farmers, two cohorts; survey and satellite measures. Counterfactual: Farmers not assigned to the service rollout in the randomized evaluation. Findings: Knowledge and practice adoption improved; crop-loss and production responses varied by weather shock. Average profit effects were imprecise. Limits: Not evidence that AI advisory improves profits. Cost-benefit extrapolation depends on shock frequency and one season of profit data. Pages: PDF pp. 0–5, 17–21, 26–28. Related, overlapping report: Harigaya and Zhu (March 26, 2025), https://precisiondev.org/wp-content/uploads/2025/03/Customized-Advisory-Services-in-India-PxD-March-2025.pdf. Count the same experiment once. |
Delivery and costs: The evaluation links to a service with millions of registered users and documented program budgets. Main gap: Registration, actual listening and benefits must be assessed separately. Scalability details- Sustained use Reported
The service had millions of registered users. Registration counts are separate from the 13,675 trial farmers and do not by themselves establish regular listening. Pages: PDF pp. 0–5 and 17–21 - Costs and funding Projected
Program budgets are documented, but benefit–cost calculations extrapolate from one season of profit data and assumed weather-shock frequency. Pages: PDF pp. 26–28 - Delivery Reported
The trial was linked to an existing agricultural advisory service. Budget and delivery arrangements provide operational information beyond a stand-alone pilot. Pages: PDF pp. 0–5 - Use in other settings Measured
The trial covered five Odisha districts and two cohorts. Effects varied with weather shocks, so expected gains depend on local exposure. Pages: PDF pp. 17–21 - Who benefits Not yet reviewed
Access, listening and effects by gender, literacy and phone ownership need further review. - Lasting benefits and wider effects Not yet reviewed
Longer-term farm benefits and continuation after project support have not yet been fully reviewed. Related reports describe the same experiment. Pages: PDF pp. 26–28
|
Arouna et al. (2021) Personalized digital fertilizer advice Impact evaluationPaper detailsFull title: One Size Fits All? Experimental Evidence on the Digital Delivery of Personalized Extension Advice in Nigeria Review coverage: Full publisher text and metadata inspected; AJAE 103(2), 596–619; online 2020, issue year 2021. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge Extension agents deliver personalized RiceAdvice recommendations, with or without fertilizer grants. AI role not establishedIntervention detailsAI in this study: Digital decision support; a specific ML component is not established in this source. Possible AI contribution: AI could help update recommendations using local soils, weather and farmer feedback. The paper does not establish a specific machine-learning component or its additional value. Other relevant problems: |
Kano State, Nigeria · rice Rice farmers |
Randomized trial Finding: Advice alone increased rice yields by about 7% and study-defined profit by 10%. Advice plus fertilizer grants produced larger gains. Key limitation: The profit measure excludes labor and equipment costs. Combined treatment effects cannot all be attributed to personalized advice. Study detailsStudy design: Cluster RCT: personalized advice, advice plus fertilizer grant, versus standard extension advice. Counterfactual: Standard extension advice; a separate treatment adds fertilizer grants to personalized advice. Findings: Advice alone increased rice yields by about 7% and study-defined rice profit by 10%; this profit measure excludes labor and equipment costs. Advice with fertilizer grants produced larger gains. Average fertilizer quantity did not increase. Limits: Does not identify incremental AI value. Input subsidy is a complementary intervention, so combined effects cannot all be attributed to personalization. The reported profit measure is not net returns after all production costs. Pages: https://onlinelibrary.wiley.com/doi/10.1111/ajae.12151; experimental design and Tables 2–3. |
Delivery: Extension agents delivered personalized recommendations; fertilizer grants increased treatment gains. Main gap: Complete delivery costs and long-run financing need further review. Scalability details- Sustained use Not yet reviewed
Repeat use of recommendations beyond the evaluation period has not yet been reviewed. - Costs and funding Not yet reviewed
Complete agent, software, training and maintenance costs and long-run financing have not yet been reviewed. - Delivery Reported
Extension agents delivered RiceAdvice recommendations. The fertilizer-grant arm shows that complementary input access affects outcomes. Pages: Experimental design; Tables 2–3 - Use in other settings Not yet reviewed
Advice needs local agronomic calibration and trained agents. Performance outside the study setting has not yet been reviewed. - Who benefits Not yet reviewed
Access and outcomes by gender, wealth and farm size have not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Persistence after advice or fertilizer grants end and wider environmental effects need further review.
|
Fafchamps and Minten (2012) SMS market, weather and crop information Impact evaluationPaper detailsFull title: Impact of SMS-Based Agricultural Information on Indian Farmers Review coverage: Publisher abstract and World Bank published-PDF provenance inspected; full extraction pending. Read the paper or report ↗ |
Information and knowledge Price and market information Free Reuters Market Light subscriptions send market prices, weather and crop information by SMS. No AI in interventionIntervention detailsAI in this study: Non-AI information service. Possible AI contribution: While no AI was used, AI could tailor SMS information to a farmer’s crop, location and trading options, and help answer follow-up questions. Effects would need testing. Other relevant problems: |
Maharashtra, India Farmers in study villages Sample: 100 villages. |
Randomized trial Finding: Some small changes in trading across locations and crop grading were reported. Average prices, crop value added and rainstorm losses did not significantly improve. Key limitation: The bundled service does not separate the effects of market, weather and crop information. Study detailsStudy design: Randomized free subscriptions to Reuters Market Light market, weather and crop SMS. Counterfactual: Farmers without the free SMS subscription in the randomized trial. Findings: Some small effects on spatial arbitrage and grading were reported; average prices, crop value added and rainstorm losses did not significantly improve. Limits: A null average result is not proof that market information never helps. The bundled service does not separately identify each information component. Pages: https://openknowledge.worldbank.org/server/api/core/bitstreams/8c37abc7-12bc-5f9b-b24b-ed217ea7221c/content; World Bank Economic Review 26(3), 383–414. |
Take-up: The paper reports low commercial take-up and limited average gains from the information service. Main gap: Cheap messages alone did not establish a viable route to wider farmer benefits. Scalability details- Sustained use Reported
Low commercial take-up is reported. Detailed subscription, readership and renewal measures still need review. Pages: Abstract; World Bank Economic Review 26(3), pp. 383–414 - Costs and funding Not yet reviewed
Full service costs, willingness to pay and financing requirements have not yet been reviewed. - Delivery Reported
The trial delivered free subscriptions through an SMS information service. Operational requirements beyond messaging need fuller review. Pages: Abstract - Use in other settings Not yet reviewed
The trial covers 100 villages in Maharashtra. Transfer to other crops, markets and information needs has not yet been reviewed. - Who benefits Not yet reviewed
Eligibility, readership and effects by phone access, literacy and gender need further review. - Lasting benefits and wider effects Not yet reviewed
Continued use and benefits after free subscriptions ended have not yet been reviewed.
|
AIEP Initiative (2026) Building AI advisory services: implementation lessons Implementation studyPaper detailsFull title: Building AI-based advisory services for smallholder farmers: Technical learnings from the AIEP Initiative Review coverage: Full HTML inspected. ArXiv identifier is 2601.11537, while displayed submission date is November 27, 2025; retain this date discrepancy in bibliographic review. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge Five prototypes combine agricultural information with AI chat, translation and voice services. AI usedIntervention detailsAI in this study: LLM advisory, retrieval, speech recognition, translation and speech synthesis. Other relevant problems: |
Kenya and Bihar, India Users of five AI-advisory prototypes Sample: More than 800 surveyed users; not a deployment count. |
Implementation study Finding: The report identifies slow responses, language-processing problems and the work needed to maintain agricultural information. Satisfaction does not establish yield or income gains. Key limitation: Implementer learning; component-level performance does not establish the quality of the complete advice service. Study detailsStudy design: Implementation synthesis and user research; no causal farm-outcome comparison. Counterfactual: No causal comparison with untreated farmers or a non-AI advice service. Findings: Documents barriers in response latency, low-resource language pipelines, and maintaining agricultural corpora. User satisfaction does not establish productivity or income effects. Limits: Implementer learning; performance of individual language components does not establish end-to-end advice quality. Pages: Sections 3 and 7. Gooey participates in FarmerChat and Opportunity International/Safaricom cohorts; avoid double-counting these deployments. |
Delivery: Five prototypes identify practical requirements for language quality, maintained content and institutional responsibility. Main gap: Operating costs and sustained use need further review; causal farmer outcomes were not assessed. Scalability details- Sustained use Not yet reviewed
More than 800 surveyed users are a research sample. Comparable active-user and retention measures across prototypes need review. Pages: Sections 3 and 7 - Costs and funding Not yet reviewed
A complete, comparable cost per active user across the five prototypes has not yet been reviewed. - Delivery Reported
Implementers identify response delays, weak language support, maintained local agricultural information and clear institutional responsibility as practical requirements. Pages: Sections 3 and 7 - Use in other settings Reported
Five prototypes in Kenya and India document different operating conditions. Lessons do not establish equivalent performance in other languages or countries. Pages: Sections 3 and 7 - Who benefits Reported
Local-language and voice-service performance affects accessibility. Comparative access and benefits for excluded farmer groups have not been established. Pages: Sections 3 and 7 - Lasting benefits and wider effects Not yet reviewed
Long-term operation, benefits and effects on nonusers need further review. Some deployments overlap with Farmer.Chat and Gooey accounts. Pages: Sections 3 and 7
|
Precision Development (2026) AI versus human-written and voiced advice Workflow evaluationPaper detailsFull title: Paddy AI A/B test 1: Artificial Intelligence versus Human-Generated Advisories Review coverage: Primary registry inspected, last modified October 13, 2026; no peer-reviewed publication claimed. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge Compare AI-written and AI-voiced telephone advisories with human-written and human-voiced advisories. AI usedIntervention detailsAI in this study: AI drafts and voices agronomist advisories. Other relevant problems: |
Karnataka, India · coffee Coffee farmers served through the Coffee Board of India Sample: 28,796 farmers randomized. |
Randomized workflow test Finding: The test assesses engagement and comprehension. Qualitative review found that some AI scripts omitted actionable recommendations or technical details. Key limitation: Farm outcomes were not assessed. The registry gives conflicting descriptions of script length, which require resolution before quantitative extraction. Study detailsStudy design: 28,796 farmers randomized across human content/human voice, AI content/AI voice, and AI content/human voice; November 2025–February 2026. Counterfactual: Human-written, human-voiced advice; a third arm combines AI-written content with a human voice. Findings: Assesses engagement and comprehension; qualitative review found that some AI scripts omitted actionable recommendations or technical details. Limits: No farm-outcome evidence. Registry summary says scripts were longer whereas detailed results say shorter: resolve before quantitative extraction. Pages: Registry Research Design and Results; source located from precisiondev.org. This is an AI-versus-human workflow comparison, unlike older non-AI advisory trials. |
Delivery: AI workflows were tested within an existing telephone-advisory service. Main gap: Cost savings, reliably actionable content and sustained farm benefits remain not yet established. Scalability details- Sustained use Measured
The randomized test measures engagement and comprehension during November 2025–February 2026. Assignment of 28,796 farmers is not a count of regular listeners. Pages: Registry: Research Design and Results - Costs and funding Not yet reviewed
Changes in writing, recording, review and total delivery costs still need review. AI production should not be assumed cheaper. Pages: Registry: Results - Delivery Reported
The test used the existing Coffee Board advisory channel. Some AI scripts omitted actionable recommendations, indicating a need for content review. Pages: Registry: Research Design and Results - Use in other settings Not yet reviewed
The trial concerns coffee advice in Karnataka. Performance for other crops, languages and services remains to be reviewed. - Who benefits Not yet reviewed
Engagement and comprehension by gender, literacy and other access barriers have not yet been reviewed. - Lasting benefits and wider effects Not reported
The registry does not establish long-term adoption, yields, income or continued operation after the test. Pages: Registry: Results
|
Gooey.AI (undated) AI advice through agricultural institutions Provider descriptionPaper detailsFull title: How Gooey.AI works with Agricultural Institutions Review coverage: Official agriculture page and Sean profile read 2026-11-10. Read the paper or report ↗ |
Information and knowledge Actionable agronomic and husbandry knowledge A provider combines institutional agricultural information with multilingual AI chat and voice channels. AI usedIntervention detailsAI in this study: Multilingual conversational agents using vetted institutional documents, messaging and voice interfaces. Other relevant problems: |
African and Asian advisory deployments Farmers and agricultural organizations |
Provider description Finding: The provider describes how institutional agricultural information can support multilingual advice through messaging and voice channels. Key limitation: Reach and cost claims are not independently verified; this page is not a farmer-impact evaluation. Study detailsStudy design: Descriptive provider account; not an independent evaluation. Findings: Provides a concrete implementation pathway: convert institutional agricultural knowledge into multilingual advice with familiar farmer communication channels. Limits: Marketing reach and cost claims are unverified; do not use them as causal impact or audited cost-effectiveness estimates. Pages: https://gooey.ai/Sean/ identifies Sean Blagsvedt, founder and CEO. Profile is a discovery seed, not agricultural impact evidence. Cross-check technical claims against AIEP and Farmer.Chat papers. |
Delivery: The provider describes institutional partnerships, local content, onboarding and familiar communication channels. Main gap: Active use, full costs and farmer benefits have not been independently verified. Scalability details- Sustained use Reported
The provider describes deployments and reach. These claims have not been independently verified or reconciled with overlapping Farmer.Chat and AIEP reports. Pages: Agriculture page - Costs and funding Reported
The provider makes cost claims, but an independently checked account of full operating costs is not available in the current review. Pages: Agriculture page - Delivery Reported
Partnerships supply agricultural information, while chat and voice tools deliver advice. Content updates, user onboarding and quality checks are part of delivery. Pages: Agriculture page - Use in other settings Reported
The provider describes multiple contexts and languages. This is not independent validation of advice quality across those settings. Pages: Agriculture page - Who benefits Not yet reviewed
Verified access and benefit differences by gender, literacy, phone access and language have not yet been assessed. - Lasting benefits and wider effects Not yet reviewed
Continued financing, maintained service quality and long-term farmer outcomes have not yet been assessed.
|
Ganesh et al. (2026) A staged approach to crop-photo diagnosis Review pendingPaper detailsFull title: Configurable Multi-Stage Vision Pipeline for Crop Disease and Pest Diagnosis Review coverage: Metadata and primary arXiv abstract retrieved; full-text review pending. Read the paper or report ↗ |
Information and knowledge Diagnosis and interpretation Combine image-quality checks, crop identification and disease diagnosis in a configurable AI workflow. AI usedIntervention detailsAI in this study: Vision-language and specialist computer-vision models. Other relevant problems: |
Ethiopia, India, Kenya and Nigeria Farmer.Chat crop-photo submissions |
Technical study · abstract reviewed Finding: The abstract describes separate checks for unusable photos, crop identity and diagnosis using farmer-submitted images. Key limitation: Full methods and data splits have not yet been reviewed; accuracy and farmer benefits cannot yet be assessed. Study detailsStudy design: Production-image analysis and comparisons of image-quality, crop-identification and diagnosis pipelines. Counterfactual: Alternative image-processing and diagnosis workflows; details still need full-text review. Findings: A recent lead explicitly separates unusable-photo rejection, crop identification, and diagnosis—useful subproblems for field deployment. Limits: Abstract only inspected. Do not import headline accuracy, assert independent validation, or infer changed crop losses until methods and splits are checked. Pages: https://arxiv.org/abs/2609.21651; September 18, 2026. Links to the same Farmer.Chat platform, not a new independent reach estimate. |
Field images: The study uses photos submitted through an operating farmer-advice platform. Main gap: Independent validation, processing costs and farmer outcomes need full-text review. Scalability details- Sustained use Not yet reviewed
Farmer-submitted images come from Farmer.Chat; the study does not provide a separately verified reach estimate in the material reviewed so far. - Costs and funding Not yet reviewed
Computing costs, review costs and costs per usable diagnosis require full-text review. - Delivery Reported
The abstract describes checks for unusable images before crop identification and diagnosis. Their operational performance still needs review. Pages: Abstract - Use in other settings Not yet reviewed
Training and test splits, external validation, and performance by crop, country and image quality have not yet been reviewed. - Who benefits Not yet reviewed
Performance and access by phone quality, language and user characteristics have not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Changes in model performance over time, disease-management decisions and farm benefits have not yet been reviewed.
|
Srinivasagan et al. (2025) Poultry welfare monitoring with low-power machine learning Technical validationPaper detailsFull title: Edge intelligence for poultry welfare: Utilizing tiny machine learning neural network processors for vocalization analysis Review coverage: Full publisher text inspected; published 16 January 2025. Read the paper or report ↗ |
Biophysical hazards and resource base Livestock health, stress and welfare A low-power device uses a neural network to classify chicken vocalizations. AI usedIntervention detailsAI in this study: TinyML convolutional network classifies poultry sounds on low-power hardware. Other relevant problems: |
Saudi Arabia · poultry Hens at a university poultry facility Sample: 500 hens; recordings collected January–June 2024 |
Technical validation Finding: The system classified poultry sounds on low-power hardware, demonstrating a possible way to monitor welfare without continuous cloud processing. Key limitation: One facility; effects on animal health, production and farm income were not established. Study detailsStudy design: Vocalization classification with within-dataset training/validation splits and edge-device deployment. Counterfactual: Alternative classification models and recordings with and without noise reduction. Findings: Demonstrates low-power acoustic classification for poultry welfare monitoring. This is not a crop pest-control study. Limits: No LMIC field validation, mortality, production or income effect established. Single-site validation does not establish performance across farms. Abstract and results report slightly different post-denoising accuracy figures; do not repeat a headline accuracy without reconciliation. Pages: User-provided DOI; PLOS sections II.2.1–2.4, III and abstract. |
On-device operation: The classifier runs on low-power hardware, reducing dependence on a continuous connection. Main gap: Performance across farms and the cost of continued operation are not yet established. Scalability details- Sustained use Not reported
Sustained use by farmers or commercial poultry operators is not reported. Pages: Sections II–III - Costs and funding Not reported
Full purchase, maintenance and operating costs are not reported. Low power use alone does not establish affordability. Pages: Sections II–III - Delivery Measured
The classification system was implemented on low-power hardware. Pages: Sections II.2.1–2.4 and III - Use in other settings Not reported
Testing across other farms is not reported. Recordings came from one university facility, with training and validation drawn from that dataset. Pages: Sections II.2.1–2.4 - Who benefits Not reported
Access and usefulness for farms with different resources were not evaluated. Pages: Sections II–III - Lasting benefits and wider effects Not reported
Longer-term operation and effects on welfare, production or income were not evaluated. Pages: Sections II–III
|
Jia et al. (2026) Machine learning for poultry processing and food safety ReviewPaper detailsFull title: Machine learning applications for postharvest poultry processing: a review Review coverage: Metadata/abstract verified through PubMed PMID 41955127; publisher search extract available, direct full-text fetch returned 403. Published online 9 April 2026. Read the paper or report ↗ |
Biophysical hazards and resource base Food contamination and safety Reviews machine learning applied to imaging, spectroscopy, sensors and genomic information. AI discussedIntervention detailsAI in this study: Reviews ML applied to imaging, spectroscopy, sensors and genomic information. Other relevant problems: |
International literature · poultry processing Studies of poultry processing, quality and safety |
Review Finding: The review identifies uses for machine learning in processing, contamination detection and microbial-risk assessment. Key limitation: This review does not establish performance or producer benefits in low- and middle-income settings. Study detailsStudy design: Review of processing, quality, safety and risk-assessment applications. Findings: Identifies microbial-risk assessment and contamination detection as distinct uses alongside processing and quality control. Limits: Not a primary field evaluation. Reviewed technologies and industrial throughputs do not establish LMIC suitability or producer benefits. Trace promising claims to original studies. Pages: User-provided DOI; https://pubmed.ncbi.nlm.nih.gov/41955127/; publisher-indexed article sections. |
Not applicable: Review of several technologies; scaling needs to be assessed for each application using its original studies. |
Savary et al. (2019) Crop losses from pathogens and pests Background evidencePaper detailsFull title: The global burden of pathogens and pests on major food crops Review coverage: Publisher metadata and PubMed abstract inspected; publisher full-text access failed. Read the paper or report ↗ |
Biophysical hazards and resource base Crop pests and diseases Estimates crop losses from expert assessments of pathogens and pests. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help identify pest symptoms from images and target crop surveillance. |
Global · five major food crops Wheat, rice, maize, potato and soybean production Sample: 137 pathogens and pests assessed |
Expert assessment Finding: Estimated losses vary substantially across crops and regions; the global maize estimate is 22.5%. Key limitation: Expert estimates do not establish how much loss a particular intervention could prevent. Study detailsStudy design: Expert assessment of losses associated with 137 pathogens and pests. Findings: Estimated global losses vary substantially by crop: the maize estimate is 22.5%, with a reported range of 19.5–41.1%. Supports the relevance of crop-health constraints. Limits: Expert estimates, not a representative farm-loss census. Ranges are not presented here as confidence intervals. Burden is not the economically recoverable share and cannot be attributed to missing AI. Pages: General search; https://pubmed.ncbi.nlm.nih.gov/30718852/; Nature Ecology & Evolution 3, 430–439. |
Not applicable: Estimates the size of a problem; it does not evaluate an intervention to expand. |
Gachoki et al. (2025) Predicting aflatoxin risk in maize Technical validationPaper detailsFull title: Predicting aflatoxin risk in maize using machine learning and satellite data in East and Southern Africa Review coverage: CGIAR institutional repository metadata and detailed abstract verified; publisher/full methods not obtained. Read the paper or report ↗ |
Biophysical hazards and resource base Food contamination and safety Machine learning combines satellite, soil, weather and terrain data to classify aflatoxin risk. AI usedIntervention detailsAI in this study: ML classifies aflatoxin risk using soil, weather, topographic and satellite covariates. Other relevant problems: |
Kenya, Uganda, Malawi and Tanzania · maize Preharvest maize samples Sample: 907 samples collected in 2009–2022 |
Technical validation Finding: In external tests, the model identified low-risk samples more reliably than medium- or high-risk samples. Key limitation: Weak identification of higher-risk samples limits operational use; reduced contamination or illness was not demonstrated. Study detailsStudy design: Environmental predictors, space-time cross-validation and external 2020 validation. Counterfactual: Predicted risk categories were checked against measured aflatoxin results, including external 2020 data. Findings: Provides an LMIC food-safety prediction example. External validation F1 scores were 71% for low-risk samples but only 5% and 14% for medium/high-risk samples. Limits: Small high-risk samples and weak class-specific validation limit operational claims. Prediction is not laboratory certification or evidence of reduced exposure, losses or illness. Pages: General search; https://cgspace.cgiar.org/items/2dded65d-ad8c-4e69-a8e0-30f24bb0fb4c. |
External validation: The model was tested with external data, but identification of medium- and high-risk samples was weak. Main gap: Costs and performance in an operating surveillance service need review. Scalability details- Sustained use Not yet reviewed
Use by farmers, inspectors or surveillance services has not yet been reviewed. - Costs and funding Not yet reviewed
Data, sampling, laboratory confirmation and recurring service costs have not yet been reviewed. - Delivery Not yet reviewed
How predictions would guide sampling or inspection has not yet been reviewed. - Use in other settings Measured
Data span four countries and several years. External validation showed weak performance for higher-risk samples. Pages: Abstract - Who benefits Not yet reviewed
Coverage of different farm sizes and locations has not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Prospective performance and effects on food exposure or health have not yet been reviewed.
|
Hengl et al. (2017) Mapping soil nutrients across sub-Saharan Africa Technical validationPaper detailsFull title: Soil nutrient maps of Sub-Saharan Africa: assessment of soil nutrient content at 250 m spatial resolution using machine learning Review coverage: Full publisher text inspected; project metadata corroborated by ISRIC. Read the paper or report ↗ |
Biophysical hazards and resource base Soil fertility and degradation Machine-learning models combine soil observations and spatial data to predict nutrient distributions on a 250-metre grid. AI usedIntervention detailsAI in this study: Predicts nutrient distributions on a 250 m grid from soil observations and spatial covariates. Other relevant problems: |
Sub-Saharan Africa · soils Soil observation sites Sample: Approximately 59,000 locations |
Technical validation Finding: The study produced continent-wide soil nutrient maps, with prediction performance varying across nutrients. Key limitation: Wide coverage and fine map resolution do not guarantee farm-level accuracy or improved yields. Study detailsStudy design: Random-forest/gradient-boosting ensemble, cross-validation and comparison with crop field-trial data. Counterfactual: Predictions were checked against soil observations and crop field-trial data. Findings: Demonstrates continental soil mapping with variable performance across nutrients; provides an information input for targeting soil management. Limits: Clustered samples and coarse geological inputs constrain mapping. Historical pooled samples cannot establish nutrient-depletion rates. Grid resolution is not guaranteed farm-level accuracy; no farmer impact established. Pages: General search; publisher Abstract, Materials and methods, Discussion; https://isric.org/projects/soil-nutrient-maps-of-sub-saharan-africa-at-250-m-resolution/. |
Geographic coverage: The study produced nutrient maps across sub-Saharan Africa; local accuracy and farm benefits remain separate questions. Main gap: Use in farm decisions and the costs of keeping maps current are not yet established. Scalability details- Sustained use Not reported
Use of the maps by farmers or advisers, and continued use over time, are not established. Pages: Discussion - Costs and funding Not yet reviewed
The recurring cost of new soil observations, model updates and delivery has not yet been reviewed. - Delivery Reported
The mapping system combines soil observations and spatial data as a shared information product. Pages: Materials and methods - Use in other settings Measured
Predictions were assessed by cross-validation and comparison with crop trials; clustered samples and coarse inputs constrain accuracy. Pages: Materials and methods; Discussion - Who benefits Not reported
Differences in access, use or benefits across farmer groups are not established. Pages: Discussion - Lasting benefits and wider effects Not reported
Effects on yields, income and sustained soil management are not established. Pages: Discussion
|
Rosa et al. (2020) Water scarcity and unused irrigation potential Background evidencePaper detailsFull title: Global agricultural economic water scarcity Review coverage: Full open-access article inspected. Read the paper or report ↗ |
Biophysical hazards and resource base Water availability and management Combines monthly water availability and crop-water needs to estimate unused irrigation potential. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help predict local water needs and tailor irrigation schedules where water and infrastructure are available. Other relevant problems: |
Global · cropland around 2000 Cropland assessed for water availability |
Model-based assessment Finding: The baseline model classifies 15% of cropland as rain-limited but without irrigation despite renewable water being available in the model. Key limitation: Modelled water availability does not establish affordable irrigation, investment returns or current local feasibility. Study detailsStudy design: Monthly agro-hydrological accounting separating rainfall limitations, physical scarcity and unused irrigation potential. Counterfactual: Modelled rainfed and irrigation scenarios under water-availability constraints. Findings: Baseline estimates classify 15% of cropland as economically water-scarce: rain-limited land without irrigation despite modeled renewable-water availability. Deficit-irrigation scenarios extend the estimated area. Limits: Modeled potential, not observed affordability, adoption or investment returns. Financial and institutional causes are inferred; current local feasibility requires separate assessment. Pages: Existing rationale bibliography plus web verification; https://pmc.ncbi.nlm.nih.gov/articles/PMC7190309/; Results: Exposure to agricultural EWS, Discussion, Materials and methods. |
Not applicable: Assesses water constraints and potential; it does not evaluate expansion of an irrigation service. |
Ortiz-Bobea et al. (2021) Climate change and agricultural productivity growth Background evidencePaper detailsFull title: Anthropogenic climate change has slowed global agricultural productivity growth Review coverage: Existing local rationale and bibliography checked; publisher metadata/search verified; direct publisher full text unavailable in this run. No new numerical estimate added. Read the paper or report ↗ |
Biophysical hazards and resource base Weather and climate exposure Estimates the contribution of human-induced climate change to agricultural productivity growth. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help translate weather forecasts into location-specific advice on planting and other farm decisions. |
Global · agriculture since 1961 Country-level agricultural productivity records |
Econometric analysis Finding: Agricultural productivity is estimated to be lower than it would have been without human-induced climate change, with larger effects in warmer regions. Key limitation: The estimated shortfall is relative to a counterfactual; it is not an observed fall in agricultural output. Study detailsStudy design: Econometric weather–productivity estimates combined with counterfactual climate scenarios. Counterfactual: Estimated productivity under climate conditions without human-induced warming. Findings: Estimates lower agricultural total factor productivity relative to a world without anthropogenic climate change, with larger estimated effects in warmer regions. Limits: Counterfactual productivity shortfall is not an observed decline in output. Existing extraction notes 2016–2020 projections. No inference about AI benefits or uniformly higher loss in every LMIC. Pages: planning/share/challenge-rationale.qmd and .bib; publisher article; author manuscript listed at https://arxiv.org/abs/2007.10415 (earlier version; avoid mixing numbers). |
Not applicable: Estimates climate impacts; it does not evaluate an intervention or its expansion. |
IPCC (2022), Chapter 5 Climate risks to food production and livelihoods ReviewPaper detailsFull title: Food, Fibre, and Other Ecosystem Products Review coverage: Existing rationale source retained; chapter link checked, HTML access blocked in this search. Read the paper or report ↗ |
Biophysical hazards and resource base Weather and climate exposure Synthesizes research on climate impacts, vulnerability and adaptation. AI role not establishedIntervention detailsAI in this study: None evaluated. Possible AI contribution: Potential AI use: combine forecasts and local data to support decisions about weather, water and disease risks; this chapter does not evaluate an AI intervention. |
Global · crops, livestock and other food systems Agricultural producers and food systems |
Assessment review Finding: Heat, water stress and extreme weather threaten food production, with different levels of vulnerability across producers and regions. Key limitation: A broad assessment cannot establish the returns to a particular AI service. Study detailsStudy design: Assessment of climate impacts and adaptation evidence Findings: Assesses risks from heat, water stress and extremes to food production and differences in vulnerability. Limits: Assessment synthesis, not an AI evaluation or a universal ranking of constraints. Pages: Chapter 5; Table 5.3 |
Not applicable: Reviews many risks and responses; it does not assess one intervention for expansion. |
Stasiewicz et al. (2017) Sorting maize kernels to reduce mycotoxin contamination Technical validationPaper detailsFull title: Multi-spectral kernel sorting to reduce aflatoxins and fumonisins in Kenyan maize Review coverage: Publisher abstract and methods overview inspected. Read the paper or report ↗ |
Biophysical hazards and resource base Food contamination and safety A statistical classifier uses spectral measurements to sort maize kernels by contamination risk. AI usedIntervention detailsAI in this study: Statistical discriminant classifier; describe method rather than assume a modern deep-learning system. Other relevant problems: |
Eastern Kenya · maize Maize from markets and trials Sample: 46 market samples in subsequent sorting tests |
Laboratory validation Finding: Spectral classification and sorting reduced aflatoxin and fumonisin concentrations in the accepted maize streams tested. Key limitation: Farm income and human-health gains were not established; rejected grain and equipment costs also matter. Study detailsStudy design: Calibration and laboratory-scale sorting; 46 market samples in subsequent sorting Counterfactual: Toxin concentrations in maize before sorting and in accepted and rejected streams. Findings: Spectral discriminant classification and sorting reduced toxin concentrations in tested accepted maize streams. Limits: No farmer-income or human-health impact established; rejected mass and full equipment/operation costs matter. Pages: Abstract and Materials and methods; Food Control 78, 203–214 |
Sorting performance: Laboratory tests reduced contamination in accepted maize; the paper also describes hardware and throughput. Main gap: Sustained mill operation and full costs need review. Scalability details- Sustained use Not yet reviewed
Repeated use by mills or traders has not yet been reviewed. - Costs and funding Not yet reviewed
Equipment, operation and rejected-grain costs have not yet been reviewed in detail. - Delivery Reported
The study describes sorting hardware and throughput; operating requirements need fuller review. Pages: Materials and methods - Use in other settings Measured
Calibration and laboratory sorting used Kenyan maize, including 46 market samples. Pages: Abstract; Materials and methods - Who benefits Not yet reviewed
Whether smaller mills and traders could access the equipment has not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Longer-term operation and effects on farmer income and consumer health have not yet been reviewed.
|
Delgado et al. (2021) Why food is lost along agricultural supply chains Background evidencePaper detailsFull title: On the origins of food loss Review coverage: Publisher discussion and existing rationale inspected. Read the paper or report ↗ |
Biophysical hazards and resource base Postharvest deterioration and physical loss Uses producer surveys to examine factors associated with food loss. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help predict spoilage risk and improve decisions about storage, transport and sales timing. |
Seven countries · selected crops Producers in selected crop–location samples |
Observational study Finding: Food losses are associated with climate, pests, knowledge, equipment, credit and markets; these relationships differ across commodities. Key limitation: Observed associations do not show how much loss an intervention would prevent. Study detailsStudy design: Producer surveys and associations Findings: Losses relate to climate, pests, knowledge, equipment, credit and markets, with relationships differing across commodities. Limits: Observational relationships are not causal recovery estimates; physical, quality and value loss differ. Pages: Discussion; Applied Economic Perspectives and Policy 43, 750–780 |
Not applicable: Examines causes of losses; it does not evaluate a service for expansion. |
Jensen (2007) Mobile phones and fish-market performance Impact evaluationPaper detailsFull title: The Digital Provide: Information (Technology), Market Performance, and Welfare in the South Indian Fisheries Sector Review coverage: Primary metadata and abstract inspected; full extraction pending. Read the paper or report ↗ |
Information and knowledge Price and market information Mobile phones allow fishers and traders to obtain price information and coordinate sales. No AI in interventionIntervention detailsAI in this study: Non-AI phone communication. Possible AI contribution: While no AI was used, AI could help summarize market information and suggest buyers or landing locations based on prices and transport costs. |
Kerala, India · fisheries, 1997–2001 Fishers, traders and fish markets |
Quasi-experimental study Finding: Phone adoption was associated with smaller price differences across markets, less waste, and improved producer and consumer welfare. Key limitation: Findings concern particular perishable-product markets; transfer to other settings requires testing. Study detailsStudy design: Nonrandomized analysis of mobile-phone rollout Counterfactual: Market outcomes before and after mobile coverage expanded, including locations not yet covered. Findings: Phone adoption was associated with less price dispersion and waste and improved producer and consumer welfare. Limits: Specific perishable-product markets; no isolated AI component or automatic transfer to crop markets. Pages: Abstract; Quarterly Journal of Economics 122, 879–924 |
Market coverage: The study follows expanding mobile coverage and changes in market outcomes. Main gap: Sustained use, affordability and transfer to other markets need fuller review. Scalability details- Sustained use Measured
The study examines phone adoption during network expansion; detailed use and retention measures still need review. Pages: Abstract - Costs and funding Not yet reviewed
Phone, airtime and business operating costs have not yet been reviewed. - Delivery Reported
The mechanism relies on mobile coverage and communication between fishers and traders. Pages: Abstract - Use in other settings Not yet reviewed
Evidence beyond Kerala fisheries has not yet been reviewed. - Who benefits Not yet reviewed
Access and benefits by producer size or other characteristics have not yet been reviewed. - Lasting benefits and wider effects Measured
The analysis covers market outcomes during 1997–2001; outcomes after that period need review. Pages: Study setting
|
Bergquist et al. (2024) Digital matching and trade in agricultural markets Impact evaluationPaper detailsFull title: Search Costs, Intermediation, and Trade: Experimental Evidence from Ugandan Agricultural Markets Review coverage: NBER indexed primary abstract inspected; direct page request returned 403. Read the paper or report ↗ |
Information and knowledge Price and market information A digital marketplace helps buyers and sellers find trading opportunities. AI role not establishedIntervention detailsAI in this study: Digital search/matching; ML additionality not isolated. Possible AI contribution: Potential AI use: improve buyer–seller recommendations using product, location and demand data. The study does not identify the added effect of AI. Other relevant problems: |
Uganda · agricultural markets Farmers and trading intermediaries |
Randomized trial and economic model Finding: The platform improved trade and market integration. Larger intermediaries drove engagement, while smaller farmers rarely used it. Key limitation: Estimated economy-wide effects differ from directly measured trial effects; the added contribution of AI is not isolated. Study detailsStudy design: Randomized marketplace rollout plus equilibrium trade model Counterfactual: Markets not offered the marketplace during the randomized rollout; wider effects are also analysed using an economic model. Findings: A matching platform improved trade and market integration; larger intermediaries drove engagement and smaller farmers rarely used the platform. Limits: Direct effects and model-based equilibrium effects differ. No separate AI-versus-non-AI comparison. Pages: Working Paper 33221, December 2024; abstract |
Participation: Larger intermediaries used the platform more than smaller farmers; the study also examines wider market responses. Main gap: Smallholder participation and the cost of ongoing provision need closer review. Scalability details- Sustained use Measured
Platform participation was concentrated among larger intermediaries; smaller farmers rarely used it. Pages: Abstract - Costs and funding Not yet reviewed
Operating costs, pricing and financing of continued service have not yet been reviewed. - Delivery Reported
A digital marketplace was introduced through a randomized rollout. Delivery arrangements need fuller review. Pages: Abstract - Use in other settings Not yet reviewed
Performance in other regions and product markets has not yet been reviewed. - Who benefits Measured
The findings identify unequal participation by larger intermediaries and smaller farmers. Pages: Abstract - Lasting benefits and wider effects Projected
An economic model examines wider market effects. Persistence after trial support needs separate review. Pages: Abstract
|
Kramer et al. (2025) Digital credit assessment and insurance for marginal farmers Review pendingPaper detailsFull title: KhetScore: Impacts of using digital technologies to unlock credit and insurance for marginal farmers in Odisha Review coverage: 3ie report metadata verified; download link looped to landing page. Read the paper or report ↗ |
Finance Seasonal liquidity and credit KhetScore uses remote-sensing information to support credit assessment alongside crop insurance. AI role not establishedIntervention detailsAI in this study: Remote-sensing-based credit assessment; precise model and component comparison to verify. Possible AI contribution: Potential AI use: estimate plot productivity from satellite data to inform credit assessment. The precise model and its separate contribution need verification in this report. Other relevant problems: |
Odisha, India · farming Small and marginal farmers |
Impact evaluation report Finding: The report examines digital credit and insurance provision for farmers who lack conventional collateral; outcome estimates have not yet been reviewed. Key limitation: The full methods and results need review before drawing conclusions about effects or the contribution of AI. Study detailsStudy design: Impact evaluation report; detailed methods pending Findings: Identifies a digital credit and insurance intervention relevant to farmers without conventional collateral. Limits: No outcome estimate or AI-specific attribution imported before the full report is extracted. Pages: 3ie Impact Evaluation Report 142 |
Review pending: The report concerns a digital lending and insurance service; its scaling results have not yet been reviewed. Main gap: Reach, lending terms, costs and continued service provision need review. Scalability details- Sustained use Not yet reviewed
Borrower participation, repeat borrowing and insurance renewal have not yet been reviewed. - Costs and funding Not yet reviewed
Credit terms, insurance pricing, service costs and financing have not yet been reviewed. - Delivery Not yet reviewed
Lender capacity, field support and data requirements have not yet been reviewed. - Use in other settings Not yet reviewed
Validation in other locations or farming systems has not yet been reviewed. - Who benefits Not yet reviewed
Access and outcomes by farm size, tenure and gender have not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Continued operation, loan repayment and resilience to shocks have not yet been reviewed.
|
Brune et al. (2016) Savings accounts and agricultural investment Impact evaluationPaper detailsFull title: Facilitating Savings for Agriculture: Field Experimental Evidence from Malawi Review coverage: NBER primary abstract and published-version listing inspected. Read the paper or report ↗ |
Finance Savings and intertemporal cash management Farmers received help opening savings accounts and arranging direct deposits, including commitment-account options. No AI in interventionIntervention detailsAI in this study: Non-AI savings access. Possible AI contribution: While no AI was used, AI could help tailor savings reminders and cash-flow guidance to crop calendars and expected expenses. |
Malawi · cash crops Cash-crop farmers |
Randomized trial Finding: Facilitating formal savings increased input use, subsequent crop-sale proceeds and household expenditure. Key limitation: Effects may reflect more than increased savings; the study does not test AI-based financial advice. Study detailsStudy design: Randomized offers of account-opening and direct-deposit facilitation (ordinary and commitment variants), versus cash payments. Counterfactual: Farmers receiving cash payments without the account-opening and direct-deposit offer. Findings: Formal savings facilitation increased input use, subsequent crop-sale proceeds and household expenditure. Limits: Mechanisms extend beyond the measured increase in savings; not evidence of AI additionality. Pages: Working paper 2015; published in Economic Development and Cultural Change 64(2), 187–220 (2016) |
Bank delivery: The intervention uses account-opening support and direct deposits through formal banking services. Main gap: Continued account use and the full cost of delivery need review. Scalability details- Sustained use Not yet reviewed
Continued account use and retention after facilitation have not yet been reviewed. - Costs and funding Not yet reviewed
Bank fees, farmer travel costs and delivery costs have not yet been reviewed. - Delivery Reported
Delivery involved account-opening assistance and direct deposits, with ordinary and commitment-account options. Pages: Abstract - Use in other settings Not yet reviewed
Performance beyond the studied cash-crop setting has not yet been reviewed. - Who benefits Not yet reviewed
Access and results across farmer groups have not yet been reviewed. - Lasting benefits and wider effects Measured
The study reports subsequent crop-sale proceeds and household expenditure; longer-run effects need review. Pages: Abstract
|
Casaburi and Macchiavello (2019) Deferred payments as a savings device Impact evaluationPaper detailsFull title: Demand and Supply of Infrequent Payments as a Commitment Device: Evidence from Kenya Review coverage: Primary metadata and abstract inspected; full extraction pending. Read the paper or report ↗ |
Finance Savings and intertemporal cash management Examines farmers’ demand for less frequent payments and buyers’ ability to commit to paying them. No AI in interventionIntervention detailsAI in this study: Non-AI payment and contract arrangements. Possible AI contribution: While no AI was used, AI could help flag missed payments and forecast buyers’ cash needs; payment enforcement would still depend on institutions. Other relevant problems: |
Kenya · dairy Dairy farmers and buyers |
Experiments and economic model Finding: Farmers valued deferred payments as a way to save, but weak enforcement constrained buyers’ ability to offer these arrangements. Key limitation: The findings concern payment institutions; they do not establish benefits from automated contracts or AI. Study detailsStudy design: Multiple experiments and a model of payment commitments Counterfactual: Alternative payment schedules and commitment arrangements across the paper’s experiments. Findings: Farmers valued deferred payments as a saving device; weak enforcement constrained buyers’ ability to offer them. Limits: Local mechanisms do not prove effectiveness of digital contracts or AI. Pages: Publisher abstract; American Economic Review 109(2), 523–555 |
Payment enforcement: Credible payment commitments are a condition for expanding deferred-payment arrangements. Main gap: Whether other buyers can honour these commitments needs further assessment. Scalability details- Sustained use Measured
Experiments examine farmers’ demand for deferred payments; longer-term participation needs review. Pages: Abstract - Costs and funding Not yet reviewed
Buyers’ financing costs and farmers’ costs of delayed payment have not yet been reviewed in detail. - Delivery Measured
Weak enforcement constrained buyers’ ability to offer credible deferred-payment commitments. Pages: Abstract - Use in other settings Not yet reviewed
Transfer to other products, buyers or institutional settings has not yet been reviewed. - Who benefits Not yet reviewed
Differences in access and benefits across farmer groups have not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Payment reliability over longer periods and during financial shocks has not yet been reviewed.
|
Bold et al. (2017) Input quality and returns to agricultural technologies Background evidencePaper detailsFull title: Lemon Technologies and Adoption: Measurement, Theory, and Evidence from Agricultural Markets in Uganda Review coverage: Author-hosted working paper and published-version metadata inspected. Read the paper or report ↗ |
Markets and transaction costs Quality verification and grading Tests input quality and examines its implications for returns and technology adoption. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help screen inputs using validated sensor data and target laboratory inspections. |
Uganda · fertilizer and hybrid seed Agricultural inputs and input markets |
Input testing and economic analysis Finding: Poor input quality can reduce returns and make it harder for farmers to learn whether agricultural technologies are worthwhile. Key limitation: The study does not evaluate an AI screening service or establish current national input-quality rates. Study detailsStudy design: Input testing, agronomic return analysis and market model Findings: Poor input quality can lower returns and impede learning about agricultural technologies. Limits: Selected input markets; no evaluated AI screening or current national prevalence estimate. Pages: Earlier title: Low Quality, Low Returns, Low Adoption; published QJE 132(3), 1055–1100 |
Not applicable: Diagnoses an input-market problem; it does not evaluate expansion of a quality-screening service. |
Bergquist and Dinerstein (2020) Competition and trader entry in maize markets Impact evaluationPaper detailsFull title: Competition and Entry in Agricultural Markets: Experimental Evidence from Kenya Review coverage: Primary metadata and abstract inspected; full extraction pending. Read the paper or report ↗ |
Markets and transaction costs Competition and bargaining Uses cost and demand shocks and offers to support trader entry to study competition. No AI in interventionIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI was used, AI could help analyse trading patterns and improve buyer–seller matching; effects on competition would need testing. |
Kenya · maize markets Maize traders |
Randomized experiments and economic model Finding: The study finds substantial trader market power; encouraging entry had negligible effects on prices. Key limitation: These estimates do not establish the effect of an AI marketplace on competition or farmer returns. Study detailsStudy design: Randomized cost/demand shocks and entry offers; structural competition model Counterfactual: Markets or traders not receiving the randomized shocks or entry offers. Findings: Finds substantial market power; induced entry had negligible price effects. Limits: Local structural estimates do not measure all farmer losses or establish the impact of AI matching. Pages: Publisher abstract; American Economic Review 110(12), 3705–3747 |
Market response: Encouraging trader entry did not produce meaningful price changes in the study setting. Main gap: Lower entry costs alone may not generate stronger price competition. Scalability details- Sustained use Not yet reviewed
Entry and continued operation by new traders need fuller review. - Costs and funding Not yet reviewed
Costs of supporting entry and operating additional trading businesses have not yet been reviewed. - Delivery Reported
The experiments included offers designed to encourage trader entry. Pages: Abstract - Use in other settings Not yet reviewed
Results in other products or market structures have not yet been reviewed. - Who benefits Not yet reviewed
How gains and losses were distributed across producer groups has not yet been reviewed. - Lasting benefits and wider effects Measured
Induced entry had negligible price effects during the study; persistence beyond follow-up needs review. Pages: Abstract
|
Goldstein et al. (2018) Land demarcation and agricultural investment Impact evaluationPaper detailsFull title: Formalization without certification? Experimental evidence on property rights and investment Review coverage: Author manuscript abstract on World Bank repository inspected. Read the paper or report ↗ |
Institutions and public delivery Land rights and tenure security A participatory programme demarcates land boundaries without immediately issuing formal certificates. No AI in interventionIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI was used, AI could help map visible parcel boundaries and organize land records; recognition of rights would still require legitimate local procedures. |
Benin · rural land Rural landholders |
Randomized trial Finding: Demarcation increased long-term investment and changed how female landholders allocated investment. Key limitation: Mapping boundaries alone does not establish or protect land rights. Study detailsStudy design: Randomized land demarcation program Counterfactual: Communities not assigned land demarcation during the study. Findings: Demarcation increased long-term investment and changed investment allocation by female landholders. Limits: Participatory recognition of rights is not equivalent to remote parcel mapping; no AI evaluated. Pages: Journal of Development Economics 132, 57–74 |
Institutional delivery: The intervention relies on participatory recognition of boundaries and land rights. Main gap: Costs and conditions for legitimate implementation elsewhere need review. Scalability details- Sustained use Not yet reviewed
Continued reliance on demarcated boundaries and records has not yet been reviewed. - Costs and funding Not yet reviewed
Surveying, administration, dispute resolution and maintenance costs have not yet been reviewed. - Delivery Reported
The intervention demarcated land through a programme recognising local rights; implementation details need fuller review. Pages: Abstract - Use in other settings Not yet reviewed
Transfer to different tenure systems or legal settings has not yet been reviewed. - Who benefits Measured
The study examines changes in investment allocation by female landholders. Pages: Abstract - Lasting benefits and wider effects Not yet reviewed
Persistence of land rights, investment and returns beyond study follow-up has not yet been reviewed.
|
Brooks and Donovan (2020) Rural bridges, market access and farm profits Impact evaluationPaper detailsFull title: Eliminating Uncertainty in Market Access: The Impact of New Bridges in Rural Nicaragua Review coverage: Existing local rationale extraction retained. Read the paper or report ↗ |
Institutions and public delivery Physical and communications infrastructure New bridges improve access to markets and employment during floods. No AI in interventionIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI was used, AI could help forecast access disruptions and prioritize infrastructure maintenance; it would complement physical investment. |
Nicaragua · rural villages Households facing flood-related isolation Sample: 15 villages |
Quasi-experimental study Finding: More reliable access reduced flood-related disruption to labour earnings and affected farm profits. Key limitation: The study covers a small setting and mainly labour-market access; it does not evaluate a digital marketplace. Study detailsStudy design: Quasi-experimental bridge evaluation Findings: Reliable access reduced flood-related disruption to labor earnings and affected farm profits. Limits: Small setting and principally labor-market access; do not infer effects of a digital market platform. Pages: Econometrica 88(5), 1965–1997 |
Reliable access: Bridges reduced flood-related disruption in the studied villages, illustrating the importance of physical access. Main gap: Construction costs, maintenance and transfer to other locations need review. Scalability details- Sustained use Not yet reviewed
Bridge use and continued use by different household groups have not yet been reviewed. - Costs and funding Not yet reviewed
Construction, maintenance and life-cycle costs have not yet been reviewed. - Delivery Reported
Physical bridges address flood-related isolation. Their construction and maintenance requirements need fuller review. Pages: Econometrica 88(5), pp. 1965–1997 - Use in other settings Not yet reviewed
Transfer to other terrain, hydrological conditions and market settings has not yet been reviewed. - Who benefits Not yet reviewed
Differences in access and benefits across household groups have not yet been reviewed. - Lasting benefits and wider effects Measured
The study examines outcomes during flood-related disruptions; longer-term maintenance and continued benefits need review. Pages: Econometrica 88(5), pp. 1965–1997
|
World Bank (2021) Digital transformation of the agrifood system ReviewPaper detailsFull title: What’s Cooking: Digital Transformation of the Agrifood System Review coverage: Primary institutional report passages inspected. Read the paper or report ↗ |
Institutions and public delivery Physical and communications infrastructure Reviews digital applications and the infrastructure, data and policies needed to support them. AI discussedIntervention detailsAI in this study: Digital and AI applications discussed, not one isolated intervention. Other relevant problems: |
Global · agrifood systems Farmers, service providers and public institutions |
Policy review Finding: Connectivity and maintained public data, including weather and soil information, are important foundations for agricultural digital services. Key limitation: The report does not estimate the causal returns to a particular public AI platform. Study detailsStudy design: Institutional policy synthesis Findings: Identifies connectivity and maintained public data, including weather and soil information, as enablers of digital services. Limits: Policy synthesis does not identify causal returns to a particular public AI platform. Pages: Policies for Tier 2 Enablers: Enabling access to data in agriculture |
Not applicable: Reviews conditions for digital services; it does not evaluate one intervention’s expansion. |
GRFC partners / FAO (2025) Food crises and their main drivers Background evidencePaper detailsFull title: Global Report on Food Crises 2025: joint release Review coverage: Official joint release read. Read the paper or report ↗ |
Institutions and public delivery Conflict and continuity of public delivery Summarizes food-insecurity conditions and their main drivers. AI role not establishedIntervention detailsAI in this study: None evaluated. Possible AI contribution: Potential AI use: combine weather, price and conflict information to support earlier warnings and help target agricultural assistance. |
Countries assessed for food crises · 2024 People experiencing acute food insecurity |
Multi-agency assessment Finding: Conflict, economic shocks, extreme weather and displacement are major drivers of acute food insecurity in the assessed countries. Key limitation: Acute food insecurity is not a measure of farm income or agricultural productivity. Study detailsStudy design: Multi-agency assessment summarized by FAO Findings: Conflict, economic shocks, extreme weather and displacement are major drivers of acute food insecurity. Limits: Selected crisis countries; acute food insecurity is not farm income or productivity. Pages: 16 May 2025 release, drivers of food crises |
Not applicable: Documents crisis conditions; it does not evaluate an early-warning or agricultural-service intervention. |
FAO (2025) Land degradation and crop production Background evidencePaper detailsFull title: The State of Food and Agriculture 2025: Addressing Land Degradation across Landholding Scales Review coverage: Existing local rationale retained; DOI fetch blocked during this search. Read the paper or report ↗ |
Biophysical hazards and resource base Soil fertility and degradation Estimates crop-production shortfalls associated with land degradation across landholding scales. AI role not establishedIntervention detailsAI in this study: None evaluated. Possible AI contribution: Potential AI use: identify areas at risk of soil degradation and support locally tailored soil-management advice. |
Global · cropland Landholdings of different sizes |
Model-based assessment Finding: The assessment links land degradation to modelled crop-yield shortfalls and examines differences across landholding sizes. Key limitation: Estimated shortfalls do not establish the gains achievable through AI or land restoration. Study detailsStudy design: Model-based land-degradation assessment Findings: Links land degradation to modeled crop-yield shortfalls and examines differences across landholding scales. Limits: Modeled shortfalls are not measured losses recoverable by AI. Pages: Boxes 7–8, as extracted in original rationale |
Not applicable: Assesses land degradation; it does not evaluate a specific restoration or advisory service for expansion. |
Vietnam statistical office (2021) Pig-sector recovery after African swine fever Background evidencePaper detailsFull title: Livestock of pig has recovered Review coverage: Official full page inspected. Read the paper or report ↗ |
Biophysical hazards and resource base Livestock health, stress and welfare Describes livestock losses and recovery following an African swine fever outbreak. No AI intervention evaluatedIntervention detailsAI in this study: None evaluated. Possible AI contribution: While no AI intervention was evaluated, AI could help identify unusual mortality patterns and prioritize veterinary investigation. |
Vietnam · pigs, 2019–2020 National pig sector |
Official statistics Finding: Nearly six million pigs were destroyed during the 2019 outbreak; recovery also faced financing constraints. Key limitation: A national outbreak does not establish recurring losses across countries or the effectiveness of an AI monitoring system. Study detailsStudy design: Official descriptive statistics Findings: Nearly six million pigs were destroyed during the 2019 outbreak; recovery also faced financing constraints. Limits: Acute national shock, not recurring LMIC-wide losses or evidence for poultry acoustic monitoring. Pages: Paragraphs on December 2019 losses and 2020 recovery |
Not applicable: Documents an outbreak and recovery; it does not evaluate a service for expansion. |
Majowicz et al. (2026) The health burden of foodborne diarrhoeal hazards Background evidencePaper detailsFull title: WHO estimates of the global, regional, and national burden of 14 foodborne diarrhoeal enteric hazards, 2000–21: an updated data synthesis Review coverage: PubMed primary metadata/abstract and existing rationale inspected. Read the paper or report ↗ |
Biophysical hazards and resource base Food contamination and safety Estimates illness and health losses from selected foodborne hazards. AI role not establishedIntervention detailsAI in this study: None evaluated. Possible AI contribution: Potential AI use: help target food-safety sampling and inspections using validated predictions of contamination risk. |
Global and regional · 2000–2021 Populations exposed to foodborne disease Sample: 14 foodborne diarrhoeal hazards |
Model-based evidence synthesis Finding: The synthesis estimates substantial illness and healthy-life loss from the hazards studied, with marked regional differences. Key limitation: It does not cover all foodborne hazards, quantify farm losses or estimate benefits from a particular detection tool. Study detailsStudy design: Updated modeled burden synthesis Findings: Documents illness and healthy-life loss from the studied foodborne diarrheal hazards, with substantial regional differences. Limits: Does not cover all foodborne hazards, quantify farm losses or estimate gains from a particular detection tool. Pages: Abstract; Lancet Global Health 14(8), 103997 |
Not applicable: Estimates health burdens; it does not evaluate an intervention’s expansion. |
Ceballos et al. (2021) Picture-based advice and insurance for farmers Impact evaluationPaper detailsFull title: Assessing feasibility and effects of personalized remote advisories based on smartphone pictures: A formative evaluation in India Review coverage: CGIAR primary abstract inspected. Read the paper or report ↗ |
Finance Insurance and other risk sharing Farmers receive generic advice, advice informed by smartphone pictures, or picture-based advice plus insurance. AI role not establishedIntervention detailsAI in this study: Picture-based services; automated AI diagnosis or claims assessment not isolated. Possible AI contribution: Potential AI use: screen crop photos and help advisers tailor responses more quickly. Automated diagnosis and its added contribution are not established here. Other relevant problems: |
Punjab and Haryana, India · farming Farmers receiving remote advisory services Sample: 203 villages |
Cluster-randomized trial Finding: Picture-based advice improved knowledge but did not establish increased short-run adoption of recommended practices. Key limitation: Participation and perceived risk reduction do not establish farm-profit gains. Study detailsStudy design: Cluster randomized trial across 203 villages: generic advice, picture-based advice, and advice plus insurance Counterfactual: Generic advice, compared with picture-based advice and picture-based advice plus insurance. Findings: Picture-based advisory improved knowledge but did not establish increased short-run adoption of recommended practices. Limits: High participation and perceived risk reduction are not measured farm-profit effects. Pages: CCAFS Working Paper 406, abstract |
Participation: The study reports participation in picture-based services; wider delivery requirements still need review. Main gap: Support costs, repeated use and farmer benefits over time need review. Scalability details- Sustained use Reported
The abstract reports high participation; unique-user counts, repeated use and retention need review. Pages: Abstract - Costs and funding Not yet reviewed
Picture submission, expert support and continuing delivery costs have not yet been reviewed. - Delivery Reported
Service delivery uses smartphone pictures and remote advice, with insurance in one treatment group. Pages: Abstract - Use in other settings Not yet reviewed
Transfer beyond the two studied Indian states has not yet been reviewed. - Who benefits Not yet reviewed
Access by smartphone ownership, gender and other farmer characteristics has not yet been reviewed. - Lasting benefits and wider effects Not yet reviewed
Longer-term adoption, farm profits and continued operation have not yet been reviewed.
|