How Charities Are Using Claude to Solve Real Problems

A 2026 study of 346 nonprofits by Virtuous and Fundraising.AI found that 92% had adopted AI in some form, but only 7% said it had meaningfully expanded what their teams could accomplish. Closer to home, the 2026 Charity Digital Skills Report – based on 807 UK organisations – found much the same shape: 79% of charities now use AI in some way, but only 38% describe that use as active or strategic, and 56% say limited skills are their single biggest barrier to progress, up from 43% the year before.

Same shape on both sides of the Atlantic: adoption is everywhere, impact is rare. Closing that gap is most of what I do for a living, and the fastest way to close it is to look at organisations who’ve actually done the work, not talk about it in the abstract.

So I went looking for ten organisations doing something real with Claude, Anthropic’s AI model, right now. I’ve focused on Claude because it’s the tool I’ve spent the last few months digging into, not because Third Sector Lab thinks it’s the only one worth using. We’re tool-agnostic: swap in Copilot, ChatGPT, Gemini or whatever your organisation already runs, and the lessons below hold up just as well. What matters is how you use the tool, not which logo is on it.

One honest note before I get into them: most of what’s published about nonprofits using AI comes from the organisations themselves or from Anthropic, because they’re the ones writing the case studies. I’ve linked directly to the original source for each entry so you can read the primary account yourself, and where a figure comes straight from Anthropic, I’ve said so. I’m not treating this as an audit. I’m treating it as what it is: first-hand accounts of practical use, worth reading and worth applying your own judgement to.

What I was looking for in each one was the same thing: a real problem, a specific use of the tool, and, the bit that matters most, a human still responsible for the decision at the end of it. That last part is the difference between a human in the lead and a human merely in the loop, and it’s the thread running through everything below.

Here’s what I found.

1. MyFriendBen – Colorado, US

MyFriendBen is a benefits-screening tool, incubated at Gary Community Ventures, built to solve a specific problem: people eligible for public benefits and tax credits often don’t know it, and the process of finding out is normally a maze. MyFriendBen takes someone through a six-minute questionnaire and tells them which of 40+ benefits and tax credits they might qualify for.

Anthropic reports that Claude now powers agents tracking 40+ programmes per state behind the tool, identifying over a billion dollars in potential value for families. The organisation itself was already doing this work and getting covered by the Colorado Sun years before Claude entered the picture, which is part of why it’s a good place to start this list: the problem and the impact both predate the AI.

For the human side of the story: New online tool helps Coloradans quickly determine which public benefits they might be eligible for (Colorado Sun). For the origin story: MyFriendBen – Gary Community Ventures.

2. London Museum – UK

London Museum’s collection is enormous, and most visitors don’t know the catalogue terms or object names that would let them search it well. In July 2025 the museum launched Clio 1.0, a conversational search agent sitting on collection pages, London stories and blog posts, combining OpenAI’s embedding model with Claude Haiku 3.5, answering only from the museum’s own data.

What makes this one of the strongest examples I found anywhere, not just in the UK, is how openly the museum documents what the tool can’t do. It warns that Clio may simplify dates, titles or materials, that some catalogue records are incomplete, that older records may contain outdated or offensive terminology, and that it shouldn’t be trusted for policy questions, which should go to the original documents instead. It even explains choosing a smaller Claude model partly to limit energy use.

That’s a museum being honest about the boundaries of its own tool rather than a vendor describing a success. The catalogue’s accuracy and the museum’s own historical judgement stay entirely with its curators. Clio just helps visitors find the right door to knock on.

Source: London Museum, Introducing our conversational search agent.

3. The Patrick J. McGovern Foundation – global philanthropy

PJMF is a global philanthropy focused on AI and data for social good, and their own internal tool, “Grant Guardian,” points Claude at a genuinely tedious problem: standardising financial due diligence across the hundreds of grantee organisations they evaluate every year. That used to mean inconsistent manual spreadsheet work, with every reviewer doing it slightly differently.

Anthropic reports that Grant Guardian has processed hundreds of grants since launch, replacing that inconsistency with a standard process. The people who actually decide whether to fund an organisation are still PJMF’s grant-making staff. What’s changed is how much manual document-wrangling they have to do before they get to make that call.

Source: The Patrick J. McGovern Foundation standardizes grant evaluation with Claude.

4. Arcola Theatre – UK

Arcola is a small east London theatre charity, the kind of organisation where a handful of staff cover grant writing, marketing, archive management and everything in between. Its privacy policy discloses using Claude.ai and Claude Code across all of that: drafting grant applications, preparing marketing materials, summarising documents, and managing the theatre’s archive.

What earns Arcola its place here is the governance sitting alongside the use cases, written in plain language rather than legal boilerplate. The policy names what never goes into the tool: children’s or vulnerable adults’ data, safeguarding information, DBS or criminal record checks, health data, ethnicity data, bank and payment details. It states that commercial agreements stop that data training any model, that a staff member reviews every output, and that AI never makes automated decisions about individuals. A small cultural charity publishing this level of detail about what it won’t let AI touch is rarer, and more useful to other charities, than a polished success story.

Source: Arcola Theatre, Privacy Policy.

5. Brian Impact Foundation – South Korea

A 13-person philanthropic foundation in South Korea, tasked with reviewing far more fellowship applicants than 13 people can realistically read in depth. They built “BEES” – Benevolence Enhancing Expert System – which uses Claude to turn messy, unstructured fellowship applications into readable, structured summaries for human reviewers.

Project director Soomin Syd Kim’s framing of the problem is the best one-line explanation of AI’s use in philanthropy I’ve come across: most philanthropic data is unstructured, which is exactly the kind of problem this tool is suited to. BEES doesn’t decide who gets funded. It gives a small team enough structured information to make that decision properly, at a scale they couldn’t manage by hand.

Source: Brian Impact Foundation Claude Platform (API) case study.

6. Epilepsy Foundation – US

The Epilepsy Foundation built Sage, a Claude-powered assistant grounded in more than 25,000 pages of curated information from epilepsy.com, developed with AWS. The problem it’s aimed at: people living with epilepsy have questions at all hours, often about deeply personal and frightening situations, and a small charity’s staff can’t be on call around the clock.

Sage is designed to answer questions and recognise emotional nuance, which is a harder brief than most customer-service chatbots face. It’s a good example of a charity using AI to extend something it already does well (curated, trustworthy health information) into hours and volumes its staff physically can’t cover.

Source: Epilepsy Foundation Launches AI Assistant to Transform Epilepsy Community.

7. National Domestic Workers Alliance – US

My favourite example here. NDWA built Ask Aya to solve a specific, well-defined problem: domestic workers, nannies, house cleaners, home care workers, often don’t know their rights or how to raise a hard conversation with an employer, and don’t have an HR department to ask. Ask Aya helps them prepare: understanding what they’re entitled to, and practising what to say before they say it to someone who controls their income.

What makes this one stand out is who built it and who runs it. NDWA says around 1,000 domestic workers helped identify the problem the tool should solve, and elected worker councils continue to shape how it works, with privacy protections and a handoff to human support built in from the start. In a beta involving 35 English and Spanish-speaking domestic workers, Anthropic reports that 93% acted on Aya’s advice and 25% negotiated a pay increase.

This is the clearest example here of what “humans in the lead” actually looks like in practice. Not a vendor deciding what workers need and building it for them, but the workers themselves deciding whether the tool should exist, what it should do, and whether it worked.

Source: Ask Aya (National Domestic Workers Alliance), with the case study and a longer Q&A also listed on Claude’s customer stories page.

8. Robin Hood, Constellation Fund and Tipping Point Community – US

Three of the larger poverty-fighting philanthropies in the US, tackling a problem familiar to anyone who has ever applied for a grant: writing proposals and reports is a huge, recurring burden on the charities a funder supports, often disproportionate to their size. These three funders piloted Claude with more than 60 of their grantee organisations specifically to help those grantees draft proposals and reporting.

It’s a more interesting story than “a funder uses Claude.” It’s funders using AI to reduce the burden they themselves place on the charities they fund, which flips the usual direction of these stories: normally it’s the charity doing the adapting, not the funder.

Source: Introducing Claude for Nonprofits.

9. Pacific Community Ventures – US

Pacific Community Ventures is a nonprofit community investor supporting small businesses, working alongside two other Community Development Financial Institutions that Anthropic is backing with credits and technical support. PCV’s own credit-risk classifier, the model that actually assesses lending risk, is a separate, purpose-built machine-learning system they trained themselves, not Claude.

Where Claude comes in is the Radiant Data Hub: a shared resource that collects and synthesises voice-based feedback from small business clients and workers, at a scale PCV’s team couldn’t previously manage by hand, with an underwriting co-pilot reportedly still in development to support (not replace) human underwriters. The aim isn’t to hand lending decisions to a model. It’s to give the people who make those decisions better information.

Source: Introducing Claude for Small Business (Anthropic).

10. RAINN – US

RAINN is the largest anti-sexual-violence organisation in the US, and the problem they’re tackling is one of reach: survivors, especially younger ones, increasingly want to reach out through encrypted messaging platforms they already trust, rather than a phone line.

RAINN used Claude Code to help its engineers build integrations with encrypted platforms including Signal, so survivors could reach RAINN’s trained support specialists through those channels. The crisis support itself stays entirely human: RAINN says survivors’ messages are never sent to or stored by Claude. AI helped build the infrastructure and extend the reach. Every part of the actual crisis support is still delivered by a trained person. That division of labour is about as clear an example of humans retaining responsibility as you’ll find anywhere on this list.

Source: Claude’s customer stories page.

What this list actually tells you

None of this proves that AI is transforming the nonprofit sector, and I’d be wary of anyone who tells you it does from ten case studies. What it does show is a genuinely wide range of practical uses: benefits access, grant due diligence, museum access, theatre operations, fellowship review, health information, worker advocacy, funder-side grant support, fair lending, crisis response. It goes far beyond the generic chatbot use case most people picture when they hear “charity uses AI.”

These are first-hand case studies rather than independent evaluations, so the headline figures should be read as reported outcomes rather than settled evidence. A recent review of UK charity, housing and social enterprise disclosures found the same pattern close to home: six credible examples with genuine first-party evidence, but almost none publishing the hours saved, cost, error rates or user satisfaction that a mature case study would normally include. Even so, the range and specificity of the uses, in the UK and beyond, is revealing.

The most useful lesson isn’t the tool itself. It’s the combination that shows up in the strongest examples: a clear, specific operational problem, safeguards built in from the start, and a human who remains responsible for the decision that actually matters. Ask Aya’s worker councils, RAINN’s human crisis specialists, PJMF’s grant-making staff still deciding who gets funded: in every one of these, Claude did the drafting or the sorting, and a person decided what it meant. That’s the model worth taking into your own organisation, more than any specific tool: AI can help you draft, summarise and sort. It can’t decide what your mission means, what risks are acceptable, or what’s right for the people you support. That’s still your job.