SMALL + EDGE AIon-device classification, sensing, decision support, offline operation
A useful framework
The World Bank organizes AI readiness around four foundations.
C1CONNECTIVITYCan people and institutions reliably reach digital services?
C2COMPUTECan systems train, adapt, host, and run models at acceptable cost?
C3CONTEXTDo data, language, and evaluation represent the setting?
C4COMPETENCYCan people build, procure, use, audit, and govern AI?
Six opportunity pathways
AI can expand capability through more than automation.
01REACHextend scarce expertise across distance and time
02UNDERSTANDtranslate language, data, images, and signals into usable information
03ANTICIPATEforecast risks early enough to change an outcome
04COORDINATEallocate resources, route services, and connect institutions
05PERSONALIZEadapt support to a person, place, crop, classroom, or clinic
06CREATE CAPABILITYhelp communities build knowledge, services, businesses, and public infrastructure
01
The foundations
Connectivity, compute, context, and competency determine what an AI system can become in practice.
Connectivity is a chain
Being online is not the same as having usable access.
01 · COVERAGEDoes a broadband network reach the place?physical infrastructure
02 · SERVICECan the connection deliver enough speed and reliability?quality and congestion
03 · AFFORDABILITYCan a household or institution pay repeatedly?data and device cost
04 · USABILITYDoes the device, language, and interface fit the user?accessibility and skill
05 · VALUEIs there a trusted service worth using?relevance and safety
The usage gap
Most people who are offline already live within mobile-broadband coverage.
4%of the global population is outside mobile-broadband coverage
38%is covered but does not use mobile internet
12%women in low- and middle-income countries are less likely than men to use mobile internet
Infrastructure matters. So do affordability, literacy, safety, relevance, and control.
Affordability changes the architecture
A 5 GB mobile package can consume more than 15% of an average low-income budget.
DESIGN THAT ASSUMES ABUNDANCE
large downloads and frequent updates
continuous cloud connection
video-first interfaces
high-end smartphones
unmetered data and reliable power
DESIGN FOR CONSTRAINT
small models and compressed assets
offline-first interaction and queued sync
speech, text, and low-bandwidth modes
shared or low-cost devices
graceful recovery after interruption
Compute is a portfolio decision
Training a frontier model is only one way to build AI capability.
BUY AN APIFast access to strong models, but recurring cost, network dependence, and limited control.LOW STARTUP · HIGH DEPENDENCE
HOST A MODELMore control over data, latency, and customization, but higher operational burden.MORE CONTROL · MORE CAPACITY
ADAPT OPEN MODELSFine-tune, retrieve, compress, or specialize an existing model for local needs.LOCAL FIT · TECHNICAL SKILL
SHARE INFRASTRUCTURERegional clouds, public compute, university clusters, and cooperatives spread fixed costs.COLLECTIVE CAPABILITY
Compute concentration
The geography of servers shapes the geography of experimentation and control.
20,000×more servers per person in the United States than in low-income countriesCAPACITY
200×more servers per person in the United States than in middle-income countriesCAPACITY
77%of data-center capacity located in high-income countriesJUNE 2025
<0.1%located in low-income countriesJUNE 2025
Small AI and offline-first systems
A capable system can keep working when the network disappears.
CAPTURE LOCALLYvoice, image, sensor reading, form, or text enters on the device
RUN A SMALL MODELclassify, translate, retrieve, or suggest without a cloud call
QUEUE THE STATEstore actions and uncertainty until connectivity returns
ESCALATE SELECTIVELYsend difficult cases to a server or specialist when needed
LEARN + UPDATEsync approved data, corrections, and model improvements
The architecture treats connectivity as intermittent, not exceptional.
Data scarcity is not just fewer rows
A setting can be data-rich and still be AI-poor.
NOT DIGITIZEDknowledge lives in speech, paper, practice, or memory
NOT LABELEDrecords exist but lack task-relevant annotations
NOT LINKABLEsystems use incompatible formats and identifiers
NOT REPRESENTATIVEthe visible data excludes key groups, seasons, places, or cases
NOT GOVERNABLEconsent, rights, stewardship, and access are unclear
NOT USABLEquality, documentation, language, provenance, or tooling is missing
Low-resource describes an AI ecosystem, not the value of a language or community.
A local data lifecycle
Development value depends on who defines, collects, validates, and governs the data.
01DEFINEWhat outcome matters, and to whom?
02COLLECTWhich people, places, languages, and conditions appear?
03LABELWhose expertise decides the ground truth?
04GOVERNWho can consent, access, revoke, and benefit?
05EVALUATEWhere does performance fail, and at what cost?
06MAINTAINWho corrects drift, updates knowledge, and responds to harm?
How multilingual models work
Shared representations can transfer learning across languages, but transfer is uneven.
CORPUSexamples across languages, scripts, domains, and modalities
REPRESENTATIONshared tokens, speech units, or embeddings
OBJECTIVEpredict, translate, align, reconstruct, or follow instructions
TRANSFERpatterns learned in one language support related tasks elsewhere
ADAPTATIONlocal data and human feedback specialize performance
Transfer is strongest when the representation and training data expose useful shared structure.
Script, domain, related languages, tokenization, data quality, task, and safety examples all affect the result.
The coverage frontier
Language technology is expanding quickly, but a language list is not an evaluation.
200NLLBmachine translation across 200 languages, including 150 described as low-resource
1,107MMSspeech recognition and synthesis coverage announced across more than one thousand languages
1,600Omnilingual MTmachine translation research announced in 2026 across 1,600 languages
517Cheetahtext generation research across African languages and language varieties
Coverage ≠ quality ≠ safety ≠ usefulness.
The adaptation toolbox
Low-resource AI is built through several forms of transfer and local evidence.
MULTILINGUAL PRETRAININGshare representations across languages and tasks
RELATED-LANGUAGE TRANSFERborrow structure and examples from linguistic neighbors
SYNTHETIC + BACK-TRANSLATED DATAcreate candidate training pairs, then validate quality
ADAPTERS + FINE-TUNINGspecialize a model with a smaller trainable component
LOCAL RETRIEVALground responses in trusted local documents and databases
SPEECH-FIRST + ACTIVE LEARNINGserve oral use and prioritize the examples humans should label next
Evaluation must match the setting
One average score can hide the exact people a system fails.
EVALUATE BY
LANGUAGE
PLACE
TASK
CONDITION
MODEL QUALITY
dialect and code-switching
region and institution
diagnose, translate, advise
device, noise, network
HUMAN OUTCOME
understanding and trust
access and workflow
decision quality
error cost and recovery
GOVERNANCE
who validates?
who is represented?
who is accountable?
who can contest?
What the evidence shows
Multilingual coverage still contains systematic performance gaps.
83 languagesMEGAVERSE assembled 22 datasets to compare model performance beyond English and warned about benchmark contamination.
Low-resource gapA study of Indigenous Brazilian languages found weaker LLM performance than in high-resource language settings.
Open local modelsEthioLLM built models and benchmarks across five Ethiopian languages plus English for locally relevant tasks.
Safety is multilingual too
A model can be safer in English than in the language where it is deployed.
HIGH-RESOURCE TESTINGSafety tuning and red-teaming are concentrated.
More examples, evaluators, benchmarks, and policy attention can produce stronger safeguards.
≠
LOW-RESOURCE DEPLOYMENTThe same model may refuse less, misunderstand more, or generate less relevant help.
Translated policies do not guarantee translated behavior.
Safety evidence must be collected in the actual language, task, and setting.
Competency is institutional
AI readiness is not the number of people who can prompt a chatbot.
BUILDdata engineering, modeling, product, security, and infrastructure
ADAPTlanguage, domain knowledge, evaluation, and human-centered design
PROCURErequirements, contracts, portability, audit access, and exit plans
USEworkflow redesign, professional judgment, and escalation
GOVERNlaw, policy, rights, risk management, and public participation
MAINTAINmonitoring, updates, incident response, budgets, and local ownership
Bottlenecks compound
The weakest foundation can dominate the whole system.
1NO RELIABLE ACCESSthe service cannot reach people consistently
2NO LOCAL CONTEXTthe model does not understand the language, institution, or goal
3NO CAPACITY TO ACTa prediction arrives, but people or agencies cannot respond
4NO ACCOUNTABILITYfailure cannot be contested, repaired, or governed
A useful prediction without a pathway to action is only information.
02
The applications
AI can extend scarce expertise, anticipate risk, and coordinate services across health, education, agriculture, public systems, and livelihoods.
Health
AI can extend clinical and public-health capacity without pretending expertise is automatic.
Make scarce expertise more reachable.
Support health workers, detect patterns earlier, translate guidance, coordinate supplies, and connect patients to the right level of care.
SCREEN + TRIAGEsymptoms, images, vital signs, and risk factors help prioritize attention
DIAGNOSTIC SUPPORTimaging, pathology, and clinical decision support highlight possible findings
PUBLIC HEALTHsurveillance, outbreak forecasting, and resource planning
ACCESS + CONTINUITYlocal-language navigation, reminders, follow-up, and supply forecasting
Education
The largest opportunity may be giving every learner and teacher more usable support.
7 in 10
children in low- and middle-income countries cannot read and understand a simple paragraph by age 10. AI matters only if it improves learning, not screen time.
TEACHER COPILOTSlesson planning, differentiation, formative feedback, and translation
PERSONALIZED PRACTICEpace, reading level, language, misconception, and accessibility
EARLY SUPPORTidentify attendance or learning patterns for human follow-up
OPEN KNOWLEDGElocal curriculum, offline content, and access across language and disability
Agriculture
AI can turn weather, soil, markets, and local knowledge into timely farm decisions.
Advice for this field, this week, in this language.
The opportunity is not generic answers. It is contextual guidance that a farmer can understand, trust, afford, and act on.
CROP + PEST VISIONphone images and field sensors support earlier detection
FORECAST + TIMINGweather, water, planting, harvest, and disease-risk windows
VOICE ADVISORYlocal-language interaction for oral and low-literacy settings
MARKET + LOGISTICSprices, storage, routes, demand, credit, and input planning
Disaster response
AI can shorten the time between a satellite image and a life-saving decision.
OBSERVEsatellites, weather stations, phones, drones, and local reports
DETECTflood extent, damaged buildings, blocked roads, and exposed settlements
VERIFYanalysts and local responders check uncertain or high-impact findings
PRIORITIZEroute rescue, shelter, supplies, cash, and repair
LEARNcompare maps with outcomes and update future response
Faster mapping matters only when it enters an operating response system.
Food security and humanitarian operations
AI can help institutions see a crisis earlier and move resources faster.
90+countries covered by WFP machine-learning food-security early-warning workSCALE
60 daysforward-looking warning window described by WFPANTICIPATION
48 hoursopen-source satellite assessment example described by WFPAFTER
The value is not the prediction. It is better anticipation, targeting, logistics, and action.
Public services and digital public infrastructure
AI becomes more useful when it can connect to trusted public rails.
From answer to completed service.
A public-service assistant can explain eligibility, help fill a form, verify status, schedule an appointment, or route an appeal only when secure systems exist behind the interface.
DIGITAL IDprove eligibility without creating a universal surveillance system
PAYMENTSdeliver benefits, refunds, fees, and emergency support
DATA EXCHANGEshare authorized records across institutions with safeguards
SERVICE LAYERlanguage access, navigation, case support, and human escalation
Finance, small firms, and trade
AI can widen access to capital and markets, or scale exclusion more efficiently.
Make smaller actors legible to larger systems.
Alternative data, translation, forecasting, and workflow tools can help people and firms participate where conventional records or expertise are scarce.
ALTERNATIVE CREDITuse transaction and business data where formal credit files are thin
FRAUD + COMPLIANCEdetect unusual patterns and reduce review burden
SME COPILOTSbookkeeping, pricing, contracts, customer support, and planning
TRADE ACCESStranslate listings, standards, documents, and buyer communication
Climate, energy, and nature
AI can connect planetary observation to local adaptation.
EARLY WARNINGflood, heat, wildfire, drought, storm, and crop-risk signals
GRID + ENERGYdemand forecasts, renewable integration, loss detection, and maintenance
WATERleak detection, allocation, quality monitoring, and watershed planning
NATUREspecies monitoring, land-use change, forest loss, and restoration
RESILIENT CITIESheat maps, mobility, drainage, infrastructure risk, and response
LOCAL KNOWLEDGEcombine community observation with sensors and remote sensing
The opportunity is a tighter loop from observation to collective action.
An evidence ladder
A promising demo is the first gate, not the development outcome.
1TECHNICAL VALIDITYDoes the model work on representative data?
2CONTEXTUAL FITDoes it work by language, group, device, and setting?
3WORKFLOW FITCan people understand, trust, and act on the output?
4SERVICE OUTCOMEDoes care, learning, response time, income, or access improve?
5DISTRIBUTIONWho benefits, who is excluded, and who bears errors?
6DURABILITYCan the system be maintained, governed, financed, and contested?
The economic architecture
Build, buy, adapt, or share is a development decision.
BUILDhighest potential control and learning, highest capital and talent requirementASK: IS THE CAPABILITY STRATEGIC?
BUYfastest path, but creates recurring fees, vendor exposure, and switching costsASK: CAN WE EXIT?
ADAPTcombine an existing model with local data, retrieval, interface, and evaluationASK: WHAT MUST BE LOCAL?
SHAREpool compute, data infrastructure, standards, talent, and governance regionallyASK: WHO GOVERNS THE COMMON ASSET?
03
Power, participation, and dependency
The development question is not only what AI can do. It is who defines the problem, owns the stack, performs the labor, and governs the outcome.
The AI value chain
Value and decision power accumulate across the entire stack.
01DATAlanguage, records, images, behavior, and local knowledge
02LABORcollect, label, moderate, evaluate, and maintain
03COMPUTEchips, data centers, energy, networks, and cloud
04MODELweights, training methods, licenses, and safety systems
05INTERFACEproduct, workflow, language, and human escalation
06DISTRIBUTIONplatform, public service, market access, and institution
Ask where local actors contribute value and where they can retain it.
The invisible workforce
Automation is often built on hidden human judgment.
DATA WORKtranscription, translation, collection, labeling, verification, and local knowledge make models possible
SAFETY WORKcontent moderation, red-teaming, evaluation, and incident review absorb difficult and sometimes harmful material
MAINTENANCE WORKcorrecting records, handling exceptions, supporting users, and repairing failures keeps systems functioning
If human labor is necessary for quality, it belongs in the system diagram and the budget.
The dependency stack
A useful product can create strategic dependence one layer at a time.
INTERFACEWorkflows become organized around one product.habit and process lock-in
MODELPrompts, evaluations, and controls depend on one provider.behavioral lock-in
API + CLOUDPrice, access, latency, and terms can change externally.commercial lock-in
DATAHistory, feedback, and local knowledge may not be portable.asset lock-in
INSTITUTIONInternal capacity weakens as outside capability becomes normal.capability lock-in
Participation has levels
Listening is not the same as sharing decision power.
INFORMpeople are told what the system will do
CONSULTpeople provide feedback after key choices are framed
CO-DESIGNpeople shape goals, data, interface, and evaluation
DECIDEaffected groups hold authority over consequential choices
OWN + GOVERNcommunities or public institutions control assets, rules, benefits, and exit
Participation is meaningful when it can change the design, deployment, or distribution of value.
Translation is not localization
Fluent words can still deliver the wrong system.
TRANSLATION
convert words across languages
preserve basic semantic meaning
support wider reading and conversation
reduce one access barrier
LOCALIZATION
adapt dialect, code-switching, examples, units, and interface
match local law, institutions, services, and professional practice
reflect social roles, goals, risk, and cultural meaning
change the workflow when the original design does not fit
Sometimes the correct localization is a different product.
Data sovereignty and local knowledge
Data can represent a community without belonging to the people represented.
AUTHORITYWho decides whether data can be collected, combined, modeled, or shared?
STEWARDSHIPWho maintains context, access rules, provenance, quality, and security?
BENEFITWho receives services, revenue, recognition, capacity, and control?
Consent is not a one-time checkbox when data becomes a durable AI asset.
Public-interest governance
The strongest response to concentrated AI is not only regulation. It is institution building.
BUILD PUBLIC CAPABILITY
digital public infrastructure with safeguards
regional or public compute access
local-language datasets, benchmarks, and models
universities, public-interest labs, and professional capacity
procurement, audit, and incident-response expertise
SET BINDING RULES
rights, privacy, labor, safety, and nondiscrimination
participation and meaningful contestability
portability, interoperability, and exit requirements
transparent evidence for high-stakes claims
benefit sharing and cross-border cooperation
Human-AI collaboration
Joint systems work when each partner covers the other's limits.
AI SYSTEMscan large data streamstranslate and retrieve quicklygenerate options and forecastsoperate repeatedly at low marginal cost
LOCAL PROFESSIONAL + COMMUNITYbring context, goals, and lived knowledgeinterpret uncertainty and exceptionsjudge social meaning and practical feasibilitytake responsibility and build trust
INSTITUTIONconnect outputs to resources and authorityset rules, evidence thresholds, and escalationmaintain systems over timeprovide appeal, repair, and public accountability
THE UNIT OF PERFORMANCE IS THE JOINT SYSTEM
A design audit
Before asking whether the AI works, ask whether the whole system works here.
01REAL PROBLEMWhat capability or choice should expand, and who defines success?
02OPERATING REALITYWhat are the device, network, energy, language, and workflow constraints?
03LOCAL EVIDENCEDoes performance hold by place, group, task, and consequential failure?
04PATH TO ACTIONCan people and institutions act on the output?
05POWER + VALUEWho owns, decides, works, pays, benefits, and can exit?
06DURABILITYWho maintains, finances, audits, contests, and repairs the system?
Near-term horizon
The next wave can make AI more local, resilient, and publicly useful.
OFFLINE SPEECH ASSISTANTShealth, agriculture, education, and public services on low-cost devices
LOCAL-LANGUAGE MODEL LAYERSspeech, translation, retrieval, and safety governed with communities
PUBLIC-SERVICE AGENTSnavigate benefits and forms with identity, payment, and appeal safeguards
REGIONAL COMPUTE POOLSshared infrastructure for universities, governments, and local firms
FEDERATED LEARNINGlearn across clinics or farms without centralizing every raw record
EDGE VISIONcrop, health, infrastructure, and environmental analysis on phones and sensors
ANTICIPATORY ASSISTANCElink early warning to pre-arranged cash, supplies, and response
CONSENT-BASED DATA TRUSTScommunities govern access, reuse, revenue, and benefit sharing
Farther horizon
Some future systems could redistribute expertise and productive capacity at global scale.
UNIVERSAL SPEECH + SIGN ACCESSreal-time communication across spoken, signed, and oral languages
COMMUNITY-OWNED MODELSlanguage and cultural systems governed as shared infrastructure
NEGLECTED SCIENCEdiscover drugs, crops, materials, and diagnostics for underfunded needs
LOCAL MANUFACTURINGAI design and robotics support distributed production and repair
PLANETARY EARLY WARNINGcombine sensors, simulation, and local knowledge into anticipatory action
ADAPTIVE CIVIC SIMULATIONcommunities explore tradeoffs in housing, water, transport, and climate plans
REGIONAL COMPUTE COMMONSclean-energy infrastructure supports public, scientific, and entrepreneurial use
AI DIVIDEND INSTITUTIONSpublic ownership or taxation returns part of AI-generated value to capability building
Speculative, not inevitable. The future is partly a technical design and partly an institutional choice.
A closing design test
When you encounter an AI-for-development idea, ask:
01 What human capability or choice should expand?
02 What infrastructure, language, data, and institution make it work here?
03 What evidence connects model performance to a real outcome?
04 Who defines the problem, does the labor, owns the assets, and captures value?
05 Can people contest failure, change the system, and maintain capability locally?
CAPABILITY IS TECHNICALDEVELOPMENT IS A JOINT SYSTEMTHE FUTURE IS A CHOICE