Small tools, real impact: what ‘small AI’ means for social protection
Professor Bhaskar Chakravorti is often credited with coining the term ‘small AI’, with the World Bank then popularising it. The framing is modest. These are not systems that transform entire bureaucracies. They are tools that photograph a paper certificate, identify it and extract information from it, or send a hyper-local weather forecast by SMS so a farmer knows when to plant. The questions they answer are specific. The data they need is manageable. And the infrastructure they require is, often, a smartphone and a data connection. That framing translates remarkably well to social protection (SP).
Discussions on AI in international development tend to fixate on big headline deployments. These generally require server farms, specialist teams, multi-year implementation programmes and of course large budgets. Instead, this blog explores smaller, limited-scope tools that run on the devices governments and communities already have, are trained on data that already exists, and solve problems that are local and well-defined.
What is meant by ‘small AI’?
Let’s start with AI itself. The AI Hub’s Global Evidence Review defines AI as computational systems that use statistical learning, pattern recognition, or generative models to perform tasks that would otherwise require human cognitive effort. That definition draws a clear line; systems that apply fixed eligibility rules, however automated, are not AI, because they do not learn from data.
‘Small AI’ is not a different technology. It is that same set of capabilities deployed at a more modest scale. What makes it small is its footprint, not its sophistication: a narrow, well-bounded task rather than a general-purpose system; a model trained on the modest data an institution already holds rather than a vast external corpus; computing demands light enough to run on existing or locally hosted infrastructure rather than rented cloud capacity; and a design that national staff can operate and maintain without permanent outside help.
These characteristics set small AI apart from the large, cloud-hosted models that dominate headlines. Those are general purpose, trained on vast datasets, run on vendor-owned infrastructure, and priced as a subscription. They are powerful, but they pull data outward, concentrate control with the provider, and create a dependency that is hard to reverse. A small tool inverts each of those properties. The trade-off is real, but for an institution with a defined problem, a tight budget, and a duty to protect citizens’ data, it often points firmly toward small. Two important benefits this may also bring are a lower carbon footprint and greater sovereignty. The latter will be discussed in more depth later. But first, it’s important to understand where small AI can play a role in the SP delivery chain.
Processing documents in the field
Offline document processing – a problem that sits at the intersection of enrolment efficiency and inclusion, is one application area where small AI can offer great value. Verifying identity and eligibility documents is one of the most labour-intensive tasks in SP administration. For beneficiaries in rural areas, it frequently requires a physical journey to a district office, documents in hand.
Mobile optical character recognition (OCR) tools change that calculation. A caseworker with a smartphone can photograph an identity card, a birth certificate, or a medical record in the field. Text-extraction software pulls the relevant information and pre-populates a registration form using machine learning (ML) technology. The caseworker checks and corrects; the system does the transcription.
This is no longer hypothetical. The South Africa Expanded Public Works Programme’s assistive OCR pipeline does this for identity documents. A lightweight chain of vision models (deep-learning computer vision, not a generative model) classifies the document, locates the relevant fields and transcribes them, attaching a confidence score to every field and showing the original image beside the extracted text. It was trained on a few hundred national ID cards and booklets, runs only as a first pass, and its model card requires a trained human to perform the verification in every case – it replaces manual data entry, not judgement [Model card].
Other examples include India’s CPGRAMS which uses classical ML to read, classify and route millions of citizen grievances to the right office, while officials decide how to resolve them [AI Hub case: IND-004]. Azerbaijan’s social-protection bot similarly answers routine pension and benefit queries using classical ML rather than a language model, which keeps the compute footprint small [AI Hub case: AZE-001].
When combined with offline capability (so that it works in areas without reliable connectivity) it becomes genuinely transformative for hard-to-reach populations.
How can small AI help reduce data sovereignty risk?
Put simply this comes down to two factors; the computational power needed to run an AI model and the sensitivity of the data that the model is trained on and/or processes. More computationally intensive models tend to require expensive graphical processing units (GPUs) that can be hard to procure. This pushes governments towards major AI compute providers, where the AI compute they offer tends to be in wealthier countries. When an externally hosted model is also trained on or processes sensitive national data (such as a household socio-economic data) this introduces a sovereignty risk.
A small AI model (trained on nationally held data and running on government/locally hosted infrastructure) does not create that risk. The model can be retrained as conditions change, with data and model weights safely stored. And the institutional knowledge required to maintain it is buildable – it does not require permanent external technical assistance.
A forthcoming report by the AI Hub considers these issues in depth.
Understanding the limits
None of this applies equally across everything SP systems do. The examples here show that smaller, advisory tools work best when they do not make final decisions.
Where decisions carry direct legal consequences (ex. whether someone receives a benefit, whether a payment is suspended, whether a fraud allegation proceeds) the governance requirements are categorically different. The systems that have generated the worst outcomes in SP are almost always ones where automated recommendations became, in practice, automated decisions.
Sovereignty on its own is no guarantee either. Colombia’s SISBEN, which uses a machine-learning model to score households to decide eligibility for dozens of programmes, is about as nationally owned as these systems come: built in-house, hosted domestically, the data never leaving the country. Yet its Constitutional Court found the system ‘arbitrary or unfair’, the algorithm is undisclosed, and manipulation has been documented [AI Hub case: COL-001]. A tool can be small, sovereign and still cause harm once it crosses from advising into deciding.
This boundary between advising and deciding is not the only limit. Tools that infer poverty from phone or satellite data are approximations that carry risks of their own. They can inherit the biases in their training data, misclassify households at the margins, and miss the people hardest to see: those with no phone, no signal, no digital footprint. A model that is 85 per cent accurate is also 15 per cent wrong, and in SP a false negative is a family that needed help and did not get it. The answer is not to avoid the tools but to treat their outputs as advisory, audit them for who they exclude, and keep a human route to redress.
Small also does not mean free or effortless. A small tool still needs data clean enough to learn from, workable connectivity, devices in caseworkers’ hands, and, most easily underestimated, the in-house skills to retrain a model over time and keep a person in the loop.
The promise of small, focused tools is real because they work within that boundary, not against it, and reduce administrative friction. That is not a modest ambition – for SP systems trying to serve the people hardest to reach, it is exactly the right one.
Where the AI Hub comes in
This is exactly the territory the AI Hub for Social Protection was set up to help partner countries navigate. Small AI sits squarely within the three principles that guide the Hub’s work. It is responsible, because it keeps people in the loop and risks contained. It is innovative, because it solves real problems with modest means. And it strengthens digital sovereignty, by keeping data and capability inside national institutions. For many partners, it may be the most realistic entry point into using AI.
Each of the Hub’s four action areas has a part to play. The knowledge base at spdci.org can document what works. The Hub can help turn small, reusable tools into digital public goods that many countries adapt rather than rebuild. The helpdesk and advisory work can help institutions judge when a small tool is the right answer and when a problem needs more. And capacity building can grow the skills – data, maintenance, governance – that small AI depends on.
Consider this an invitation. If you are working on, or weighing up, a small AI tool for SP – a targeting model, a chatbot, an offline document reader, an early-warning trigger – we want to hear from you! We welcome your ideas for small AI digital public goods that the Hub could help develop. Get in touch to tell us whether this is a topic worth pursuing, share what you are seeing, and explore the knowledge base at spdci.org.
The promise of small AI will be realised in practice, by practitioners, and the Hub is here to help connect that work.
Author: Robert Worthington
