Back to Insights

Insights
Hub team |
Nov 17, 2026

How is the social sector building AI tools?

Part 1 of 2. Part 2 looks at how organizations evaluate these tools.

Key takeaways

Drawn from 571 AI use cases on the Social Good AI Hub and interviews with seven organizations.

This piece examines three choices organizations make when building AI tools: what to build, where to start from, and what to release openly. For each, we also note where funders can help.

  1. AI tools take on a wide range of problems, led by diagnosis and screening, tutoring, and forecasting (15% to 17% each). Each involves making sense of large amounts of material, such as scans, students’ answers, or weather data, that would otherwise take scarce expertise or time.
  2. Chatbots and other conversational tools are the most common format, seen in roughly 34% of use cases. Nine in ten chatbots run on a large language model.
  3. Nearly 70% of use cases build on a foundation model, and many combine it with additional AI approaches (such as bespoke deep learning or classical ML) or customize it to fit their specific local or programmatic context.
  4. The majority of use cases on the Hub are not open (open source or source available), but some sectors like climate have a large proportion of open use cases that others can build on. Staff bandwidth and funding incentives keep many tools closed.
  5. Shared benchmarks and context layers, more usable administrative data, and a higher quality and quantity of training data for underserved languages would improve many AI tools at once, which makes them strong areas for funders to invest in as they look to support impactful AI deployments in the social sector.

Introduction

The social sector is investing significant time, money and attention in AI, and organizations are building tools to tackle problems in education, health, livelihoods, climate and beyond. Behind each tool is a series of choices that shape what gets built and how useful it can become.

To understand how the social sector is building AI tools, we analyzed the 571 use cases documented on the Social Good AI Hub and interviewed seven organizations that have deployed them.

In this post, we examine three key choices that organizations face: what to build, where to start from, and what to release openly. In a follow-up post in a few weeks, we will cover a fourth critical choice that organizations deploying AI tools make: how to evaluate whether these tools deliver the outcomes they were built for.

1

What to build

The kinds of tools organizations build and the platforms they run on

2

Where to start from, and what to add

How much organizations rely on models others built, and what they add themselves

3

What to release openly

Whether organizations publish their code or models, and what holds them back

1

What to build

Two questions come up frequently when choosing what to build: what problem AI should take on, and what kind of tool to build. Both shape what a tool costs to run.

1What problem the AI tool takes on

Among the use cases found on our Hub, the tasks AI performs vary widely. The most common are diagnosis and screening, tutoring, and forecasting, each found in 15 to 17% of use cases. This is likely because AI tools are effective at taking in a large amount of material that would require significant time and/or scarce expertise to interpret (such as photos and scans, students’ answers, or weather data, disease monitoring data at population scale, etc.) and turn it into an assessment a human can act on.

Another question of interest is where organizations choose to point AI: at internal operations such as finance, HR, and reporting, to run their organizations more efficiently, or towards improving the services or programs they deliver. However, since the Hub selects for tools that are program-facing, this is not something we are able to see in the data.

Figure 1

What problems AI tools take on

Share of the 571 use cases carrying each of the main functional tags.

Diagnosis and
screening
17% (99)
Tutoring and
learning
17% (95)
Forecasting and
prediction
15% (83)
Monitoring and
early warning
13% (76)
Data
analytics
13% (72)
Recommendation and
matching
6% (32)
Search and
knowledge retrieval
5% (29)
Content
generation
5% (28)
Accessibility and
assistive technology
4% (25)
Scientific
research
4% (24)
Translation and
transcription
4% (21)
Benefits and
entitlements access
3% (15)
Note: A use case can carry more than one function, so shares add up to more than 100%. Monitoring and early warning covers tracking disease outbreaks, ecosystems, pollution, and climate hazards. Content generation covers lesson plans and learning materials, as well as images, video, and professional text.

2What kind of tool to build

Beyond the problem a tool tackles, organizations also need to choose the form it takes, meaning how people actually interact with it. The most common format among use cases on our Hub is conversational.

34%
of use cases involve a chatbot, voice assistant, SMS service or helpline, and 90% of these run on a large language model.

That makes chatbots the most common single type of tool. Other formats found across use cases include mobile apps that read photos or scans to spot disease, dashboards that help officials plan, devices like sensors and handheld scanners that collect data in the field, or datasets and models that other organizations can build on.

Chatbots tend to cluster in a few different problem areas – 54% of chatbots give information or advice, including helping people find services, and about 25% support tutoring and learning (Figure 2).

Figure 2

What problems chatbots take on

Share of the 195 chatbots on the Hub, by main function.

Information, advice, and helping people find services
54% (106)
Tutoring and learning
24% (46)
Diagnosis, screening, and triage
8% (16)
Other
14% (27)
Note: Covers chatbots, voice assistants, SMS services, and helplines (34% of use cases). Each is counted once, by its main function. “Other” covers analytics, forecasting, monitoring, content generation, translation, assistive technology, and research tools, each 3% of chatbots or less.

Most organizations that build chatbots deploy them on platforms people already use, rather than on new apps. Of the 80 chatbot use cases that name a channel, 80% run on at least one existing platform such as WhatsApp, SMS or a phone line. Only one in five reaches users solely through a specialized app or website. WhatsApp is the platform of choice for nearly two-thirds of chatbots (64%).1

Three organizations we interviewed attributed this to users’ familiarity with the platform and the fact that nothing new needs to be installed. However, using an existing platform like WhatsApp also means being subject to its pricing, limits and available features. As messaging costs have risen, one organization we spoke to is raising funds to move to its own app.

Another noteworthy trend among chatbots is that they rely on the same few providers – among the 72 chatbot use cases that name their model, 81% use one of three frontier model families: GPT, Gemini, or Claude.

Our interviews highlighted that builders are thinking carefully about how changes in model pricing and capabilities could affect their deployments. Building on outside platforms and proprietary models means that others set much of what a tool costs to run. Digital tools have traditionally been expensive to build but cheap to run, so serving more people adds little to the cost. A chatbot built on a hosted model is different: organizations pay for every message and model call, at rates set by the messaging platform and the model provider, so costs rise with each new user and continue after grants and donated credits run out. Running costs came up as a growing concern in our interviews, and some organizations now treat cost as part of how they evaluate their tools.

Nearly seven in ten use cases on the Hub build on a foundation model, and among those that name their model, about two-thirds use GPT, Gemini, or Claude. Given how widely these models and platforms are used across the sector, these types of changes could have implications for many tools across the ecosystem – far beyond chatbots alone.

What funders can do

Fund running costs, and the work that brings them down

AI tools scale differently from the digital tools funders are used to.

Those were expensive to build but cheap to run, so reaching more people added little cost. With AI, every message and model call costs money, so costs rise with each new user. Those costs are also largely set by others: platforms and model providers decide the pricing, limits, and features a tool relies on, and a change on their side can affect many tools at once. Organizations are already responding – Rising Academies is reducing how often it needs to call a large language model and Edufarmers is raising funds to move away from WhatsApp. Funders can support this by covering running costs that grow with use for tools that have shown they work, and by funding work that reduces exposure to a single provider, such as routing simple requests to smaller models or building tools in ways that allow a model or platform to be swapped.

2

Where to start from, and what to add

Once builders know what problem to tackle with AI, and what form that tool will take, the next decision is which AI approach to start from, and what else they need to add to make it work for their use case. Our analysis finds that most organizations start from a foundation model, and then build around it.

1Which approach to start from

Across all use cases, four underlying AI approaches2,3 are most common:

  1. Foundation modelsLarge general-purpose models pretrained on vast amounts of text, images, audio or other data, which others can use for many tasks without additional training. Examples include GPT, Gemini, Claude, Llama and the speech model Whisper.
  2. Bespoke deep learningNeural networks trained for one specific task, either from scratch or by fine-tuning an existing model, such as spotting crop disease in photos, mapping settlements from satellite imagery or transcribing a low-resource language.
  3. Classical machine learning and statistical methodsEstablished techniques such as regression, decision trees and clustering, that are applied to structured data like survey responses or administrative records to predict outcomes, or group cases.
  4. Symbolic, optimization, and decision methodsApproaches built on explicit rules, objectives and constraints, or that learn the best action through trial and feedback, such as eligibility rules, route and resource-allocation optimization, knowledge graphs and adaptive experiments.

Nearly 70% of use cases build on top of a foundation model, making this the most widely used approach among use cases on the Hub (Figure 3). Foundation models are deployed in at least 60% of cases in every sector except climate (where that share is 44%). In sectors like education, livelihoods, governance, justice and social protection, they are used in more than 80% of use cases. As noted in Choice 1, these models come mostly from a small number of providers.

Figure 3

Which AI approaches tools use

Share of the 571 use cases using each approach.

Foundation
models
69% (392)
Classical machine learning
and statistical methods
49% (282)
Bespoke deep
learning
33% (191)
Symbolic, optimization,
and decision methods
8% (48)
Note: A use case can combine more than one approach, so shares add up to more than 100%; 49% combine two or more. 25 use cases do not specify their approach. Most classical machine learning entries describe the method only as “machine learning” or “predictive analytics.”

2Going beyond foundation models

As Figure 4 shows, across sectors and use cases, foundation models are often combined with other approaches. After foundation models, the most common approach is classical machine learning and statistical methods, listed in about half of use cases (49%). Bespoke deep learning appears in a third of use cases (33%).

Looking at how AI approaches split by use case function reveals some distinct trends. AI tools that work with language, such as those that answer questions, tutor or generate content, almost always build on a foundation model, and between 40% and 60% use it on its own. Tools that make predictions or spot patterns in images, sensor readings or other types of records use foundation models far less frequently: for example, more than half of forecasting and monitoring use cases do not deploy foundation models at all. This also explains most of why climate, where use cases like forecasting and specialized monitoring are more common, relies more heavily on bespoke deep learning and classical machine learning than other sectors do. Tools that work with satellite images, sensor readings, or photos, such as those that map settlements or spot crop disease, tend to require organizations to train their own models for that kind of data. One coalition we spoke to built its own models because no existing source provided the data it needed.

Figure 4

How use cases combine AI approaches, by sectorHow use cases combine AI approaches, by function

Each use case falls in one group. The three red segments include a foundation model, so together they show the share that builds on one.

  • Foundation model only
  • Foundation model + bespoke deep learning
  • Foundation model + classical ML (or both)
  • Bespoke or classical ML, no foundation model
  • Other
Note: Sectors overlap, so a use case can count in more than one; the “All use cases” row counts each once. Classical ML includes use cases described only as using “machine learning” or “predictive analytics.” Other is symbolic or optimization methods only, or no approach recorded.Note: A use case can have more than one function, so it can count in more than one row; the “All use cases” row counts each once. Benefits and entitlements access (15 use cases) is not shown. Classical ML includes use cases described only as using “machine learning” or “predictive analytics.” Other is symbolic or optimization methods only, or no approach recorded.

3Other types of customizations

Another key component of making a use case work beyond the capabilities that foundation models can offer is context-specific customization. Some organizations write instructions for the model or give it their own documents and guidelines to draw on. Others add parts of their own: fixed rules or an additional smaller model trained for one part of the task. A foundation model might understand a user’s question and draft a reply, for example, while the organization’s own rules and models check the answer against guidelines, adapt advice to local conditions, or decide what a learner sees next. The balance between off-the-shelf and in-house parts varies from tool to tool, and the examples below show how organizations make those choices.

In general, the more a tool depends on local knowledge, such as which local crops, laws, or flood patterns are relevant in a given place, the more custom components will need to be built.

Where shared infrastructure could help

The need for some components that are frequently built in-house, such as making administrative data usable, local-language resources, and model benchmarks, are widely shared. Building this infrastructure once and making it available broadly could limit the need for each organization to start on the same problems from scratch, ultimately directing limited resources towards tool improvements, evaluations, and scaling.

Government data commons

For organizations working with government stakeholders in the Global South, the data they need, such as school, health or disaster records is often already available, but in formats that are hard to use. CivicDataLab, for example, spends much of its build effort creating pipelines that make datasets from different departments work together. If departments that hold frequently accessed data adopted a common standard, the cleaning and harmonizing could be done once at the source rather than repeated downstream by each organization that builds a tool. This would also have benefits beyond AI tools, including to government departments themselves – efforts to make government administrative data more interoperable have been underway for a long time but have a mixed record. Where data exists but is not publicly available, such as records of disaster losses, agreed rules on who can use it and what for could make sharing less risky for the government entities that hold it.

Language models for underserved languages

Most major AI models are trained mainly on English and other widely used languages, so they perform worse in languages for which digitized text is less available or not representative of how the language is used in practice – many of these languages are spoken by tens of millions of people in the global south. Closing the gap depends on more and more representative language data, such as text and speech that reflect real dialects, accents and mixing of languages, and on testing models on the tasks users need – something that researchers at Stanford HAI, The Asia Foundation and the University of Pretoria highlight in a 2025 white paper.4 Several organizations we spoke to named better models for local languages as a significant gap and described how they are getting around this in the meantime. Edufarmers’ tech partner, for example, has built a database of local dialects. Foundation models have since improved enough to make that specific database unnecessary, but progress is uneven, and organizations working in underserved languages still lack the models they need.

Independent benchmarks for common tasks

The organizations we interviewed found that general rankings of AI models told them little about how a model would perform in their own tool. A benchmark, a shared test with agreed correct answers, can fill that gap, but only if it tests a specific task, such as reading policy documents or marking a student’s maths answer. Without one, each organization has to build its own tests, which takes time many cannot spare. While this may be necessary in highly specialized contexts, there are many common use-case domains (such as clinical decision support or lesson planning) where shared benchmarks would still go a long way towards enabling better evaluation practices and easier comparison of different tools that address similar problems. We return to benchmarks in Part 2, since they also shape how organizations judge their own model results.

What funders can do

Invest in language models for underserved languages

Today’s AI models handle many underserved languages poorly, so organizations build their own language resources or limit their tools to widely spoken languages.

Groups such as Masakhane, AI4Bharat, Karya and Bhashini have built datasets and models for African and Indian languages, and in September 2026, 60 organizations committed to making AI usable for 3.4 billion people who speak underrepresented languages. Even so, most of these languages still lack enough data, whether raw text and speech recordings or examples that people have annotated to train and test models. Funders can support the collection of language data as well as models, prioritizing data that reflects how people actually speak and the sectors where tools will be used, such as health and education. They can also involve native speakers in designing these datasets, pay contributors fairly, and fund evaluations of how well models perform in these languages, since standard benchmarks often miss local context.5

Invest in shared knowledge for AI tools

Most AI tools that give advice rely on documents their builders collect and check themselves, so the same guidelines are gathered, cleaned and integrated into the tool many times over, at varying quality.

A shared context layer built from official sources such as national curricula, health guidelines and benefits eligibility rules would give every tool the same trusted starting point and let each answer be traced back to a named source. Learning Commons and Oak National Academy already do this for school curricula in the United States and England, and IDinsight is building similar layers for education in several low- and middle-income countries. Funders can support the development of these public goods in other sectors and countries.

Help unlock government data

Efforts to standardize and share government data have a long and mixed record, from uncoordinated mobile health pilots6 to disaster loss data that countries report unevenly.7

Newer platforms, such as the UN System Data Commons launched in September 2026, bring published national statistics together, but data held by sub-national departments remains harder to reach. Where standards already exist, such as UNDRR’s DELTA Resilience for disaster loss data, funders can underwrite the work of getting departments to adopt shared standards, which is often much harder and more time-consuming than designing the standard itself.

3

What to release openly

Once a tool is built, organizations decide whether to release its code or models openly, so that others can check, reuse, or build on them.

We count a use case as open when it is publicly described as either open source or source available, where the code can be viewed but reuse may be restricted. Of the 315 use cases on the Hub that disclose whether this is the case for their tool, 41% are open.8 There is significant variation in the degree of openness across sectors: from 73% for climate-focused use cases to 21% for livelihood-focused use cases. Climate, humanitarian response, governance, and agriculture stand out as sectors where open use cases appear most frequently.

Figure 5

How open use cases are, by sector

Among the 315 use cases that disclose their status. The four sectors where most tools are open are highlighted.

Climate
73% (40 of 55)
Humanitarian
71% (15 of 21)
Governance
63% (19 of 30)
Agriculture
61% (11 of 18)
Social protection
37% (17 of 46)
Education
34% (28 of 82)
Health
27% (25 of 92)
Justice
24% (4 of 17)
Livelihoods
21% (6 of 28)
Note: Our definition of open includes source-available tools. A use case can belong to more than one sector, so sector totals add up to more than 315. Several sectors rest on small numbers of use cases, so treat differences between sectors as indicative.

Why some sectors may be more open than others

Our limited interviews point to two possible explanations, though more work is needed to understand the dynamics of open-sourcing in the context of AI tools deployed in the social sector.

Types of data leveraged and research norms in climate

Much of the AI used in climate is built by teams training their own models, often drawing on public data that carries little privacy risk, from satellite imagery to published laws and policies. Many of these tools are built and used by researchers who expect to see the data and methods behind a result. Together, these factors may help explain why AI tools in the climate sector tend to be open more often than in other sectors. The two climate organizations we spoke to treat openness as part of their work from the start: Climate TRACE publishes its data and methods, and Climate Policy Radar, which is funded by philanthropy, aims to put its work in the public domain so others can check and improve it.

Government customers

Having a government entity as the end user or customer can also push developers towards greater openness. CivicDataLab, a private company, publishes all its code openly. Its government partners want to own the tools they use and avoid being tied to one supplier, so for CivicDataLab, open code is an advantage that allows them to build for governments. Noora Health’s government partners ask for two things: open code, which their own staff can run without a lengthy tender, and open-weight models, which reduce dependence on models built abroad and keep data in the country. This may explain why governance is among the more open sectors, though governments also buy plenty of proprietary software.

What keeps organizations from opening up their tools

Most use cases on the Hub that disclose their status (59%) are not open. It is also likely that use cases that do not disclose are closed, on average, since teams with open tools often have a strong incentive to say so. The organizations we interviewed pointed to two related reasons:

Funding incentives

Releasing a tool’s code or models can give away what sets the organization apart, or weaken the case for further funding, since once a good-enough open version exists, funders may stop paying for a tool’s development or maintenance. This points to a deeper gap that predates generative AI: the sector has no established, sustainable way to fund the maintenance and growth of shared digital tools once initial grants run out.

Bandwidth

Putting code online is not the same as making it usable. Releasing a tool properly means documenting the solution and its limitations, separating it from internal systems, removing personal information, and maintaining it and supporting its users after release. Rising Academies, for example, has to check any dataset it releases for students’ names and locations, which students sometimes type alongside their answers. Organizations with small teams often cannot spare the time, and two that we spoke to said that code released without this work is of little use to anyone else.

What funders can do

Pay for the work of releasing code and models openly

Releasing a tool’s code or models openly takes time: documenting them, separating them from internal systems and personal information, maintaining them, and responding to the questions and requests that come once others start using them.

Funders have backed open source for years. The Principles for Digital Development, endorsed by more than 300 organizations including donors, ask organizations to share and reuse their work, and the Digital Public Goods Alliance’s standard requires documentation good enough for someone unfamiliar with a tool to launch and run it themselves. Our interviews suggest the barrier is funding: few grants pay for the documentation, cleanup and user support that open release requires. Funders can close this gap by adding a dedicated open-release line to AI grants, and by paying for maintenance and user support for a set period after launch, so the team can fix problems and answer questions as others start using the tool.

What these choices have in common

Across the three choices, organizations tend to start from general-purpose models and platforms and adapt them to their own context. Most start from a foundation model, and most chatbots run on platforms people already use. What they build themselves is the part their context requires, such as local knowledge or rules that check a model’s answers. Even so, some of what organizations build for their own contexts is needed by many others. Building these once as shared resources is a tide that would lift many boats.

For organizations building AI tools, two lessons stand out from our interviews: weigh running costs from the start, and if a tool may be released later, document it and keep it separate from internal systems as it is built. Funders can take on the costs that organizations struggle to cover alone by paying for running costs as tools grow, for shared resources such as language data and benchmarks that many teams now build separately, and for the work of releasing tools openly.

These investments also make it important to understand what we know about the tools being built, and where the evidence is still limited. Part 2 looks at how organizations assess that.

Next in this series

Part 2 of this series looks at how organizations assess whether their AI tools work. It follows the evidence organizations collect and publish, from how well a model performs to whether people’s lives improve, and the reasons that evidence often stops short of providing clarity on whether the AI tools are enabling the outcomes organizations ultimately care most about.


Explore the Social Good AI Hub

Search by sector, geography and organization, and see what each use case is built on and whether it is open.

The Social Good AI Hub

Notes

  1. These figures are based on the 80 chatbot use cases (out of 195) that name a channel. For the remaining 115 chatbots, the platform is unspecified. ↩
  2. We follow the OECD’s broad definition of an AI system, which includes machine learning, knowledge-based and rule-based systems. OECD (2024), ‘Explanatory memorandum on the updated OECD definition of an AI system’, OECD Artificial Intelligence Papers, No. 8. ↩
  3. The OECD defines what counts as AI but does not group it into types. The four categories used here are our own. ↩
  4. Pava, J. N., Meinhardt, C., Badi Uz Zaman, H., Friedman, T., Truong, S. T., Zhang, D., Cryst, E., Marivate, V. and Koyejo, S. (2025), ‘Mind the (Language) Gap: Mapping the Challenges of LLM Development in Low-Resource Language Contexts’, white paper, Stanford Institute for Human-Centered Artificial Intelligence, The Asia Foundation and University of Pretoria. ↩
  5. Pava et al. (2025), ‘Mind the (Language) Gap’, see note 4. ↩
  6. Alunyu Egwar, A., Amiyo, M. R. and Nabukenya, J. (2024), ‘Framework for standardizing digital health in resource-constrained settings: a case study of Uganda’s digital health communication infrastructure’, Oxford Open Digital Health, 2, oqae018. ↩
  7. International Science Council (2019), Disaster loss data in monitoring the implementation of the Sendai Framework, policy brief; UNDRR (2026), ‘Sendai Framework data shows both progress and persistent gaps in reducing disaster risk’. ↩
  8. Just over half of use cases in the Hub disclose their status. Openness is recorded as stated publicly; we did not verify whether each tool’s code or models were readily available. ↩

Methodology

The descriptive findings come from the Hub’s repository of 571 use cases across nine sectors, compiled from public sources and reviewed by our team. Shares are of all 571 use cases unless stated; open-source shares use the 315 use cases that disclose their status.

The Social Good AI Hub is a living repository. New use cases are added regularly, and organizations can add and edit their own entries, so the Hub changes in ways we do not always see or control. The figures in this piece are based on the 571 use cases on the Hub as of 9 October 2026, so numbers on the Hub after this date may differ from those cited in this piece.

The Hub is not a random sample, and there is no complete count of AI use cases in the social sector to compare it against. Because it is built from funder, accelerator and directory lists, it likely over-represents organizations within these networks and those that document their work online in English, and under-represents early-stage, local and government-built tools. Findings therefore describe documented use cases rather than the sector as a whole. Some sector breakdowns rest on fewer than 20 use cases and should be read as indicative.

To put these findings in context, we also spoke with seven organizations in September and October 2026 about how they built their AI tools, how they evaluate them, and what has helped or held them back. We purposely selected organizations that took different approaches to building and evaluating their tools and used their experiences to illustrate the patterns in the data. Seven conversations cannot represent the sector, so these insights are not generalizable. They are intended to provide practical examples of how organizations are approaching these questions and the challenges they encounter.

Acknowledgements

We are grateful to the organizations that shared their time, experience and insights with us: CivicDataLab, Climate Policy Radar, Climate TRACE (through WattTime), Edufarmers International Foundation, EIDU, Noora Health and Rising Academies.

The Social Good AI Hub is built by IDinsight with support from Google.org.

Stay in the loop

Get new insights in your inbox

New thought pieces and data updates as they publish — no more than one note a month.