How charities are using Microsoft Copilot to solve real problems
24 August 2026 | Estimated reading time: 9 minutes
In 2024, the UK Department for Business and Trade gave 1,000 staff access to Microsoft 365 Copilot and ran one of the more substantial independent evaluations of the tool. The headline result looked promising: 72% of respondents were satisfied or very satisfied, and most reported some time savings. But the evaluation found no evidence that those savings had translated into improved productivity across the department. In a small set of observed tasks, people using Copilot completed Excel analysis more slowly and produced less accurate work. And 22% of respondents said they had encountered Copilot presenting false or misleading information as fact.
That is a useful place to begin, because it captures the gap between the way most AI case studies are written and the reality of using these tools inside an organisation. People can like a tool. It can save them time on individual tasks. That does not automatically mean the organisation delivers more, makes better decisions or creates more impact.
The same gap appears in the charity sector. The 2026 Charity Digital Skills Report found that 79% of UK charities were using AI in some form, but only 38% described that use as active or strategic. Copilot is especially interesting because so many charities already use Microsoft 365. It sits inside Outlook, Teams, Word, Excel and SharePoint, close to the emails, meetings and documents that make up much of everyday charity work.
But Copilot is not really one tool. Some organisations use the chat interface, some use it inside Teams or Word, and others connect it to organisational files or use it for data analysis. The risks and results are very different in each case.
So I went looking for third sector organisations using Copilot to tackle a real problem, with enough detail to understand what happened next. I deliberately looked beyond Microsoft customer stories. The most useful examples I found were not all success stories. Some showed where Copilot saved time. Others showed where it lost nuance, created more work or exposed problems in an organisation’s existing information governance.
One honest note before I get into them: this still is not an audit. The evidence ranges from an organisation’s published methodology to conference presentations and interviews with staff. Two of the cases reflect Copilot as it existed in 2024, and the product has changed since then. I have linked to the underlying source for each entry so you can see what was reported and apply your own judgement.
What I was looking for was the same as in my earlier post on charities using Claude: a real problem, a specific use of the tool and a human still responsible for the decision at the end of it. I also wanted to know what did not work. That is often more useful than another estimate of hours saved.
Here is what I found.
1. Mercy Corps – global humanitarian organisation
Mercy Corps’ monitoring, evaluation and learning teams work with exactly the kind of material people assume generative AI will be brilliant at: large programme documents, narrative reports, spreadsheets and piles of qualitative data. Staff wanted to know whether Copilot could turn spreadsheet data into draft reports, create presentations from existing reports, code qualitative material and check documents against good practice.
Rather than buying licences for everyone, Mercy Corps formed a working group, identified specific use cases and bought a small number of licences to test them against the features Microsoft said the product could provide.
The results were mixed. Copilot’s automated notes in Teams were a significant time saver. Email summaries were useful, and it was good at finding material already stored in SharePoint. But it struggled with large programme documents. Producing usable summaries took several attempts and still required an experienced analyst. The results were described as rudimentary rather than the holistic analysis the team needed. It also performed poorly at coding qualitative data and interpreting the messy spreadsheets common in real monitoring and evaluation work.
Mercy Corps concluded that the tool required more time and effort than expected and that its usefulness was uneven. It was finding more direct value from custom chatbots built for narrower uses. Summarising a meeting and analysing a complex programme may sound similar, but one produced a clear saving while the other still needed sustained expert judgement. The lesson is not that Copilot will always be bad at analysis. It is that charities should test the work they actually do, using the documents and data they actually have, rather than buying on the strength of a feature list.
Source:Â INGO Experiences Adopting Microsoft Copilot, MERL Tech. This account comes from a July 2024 presentation by Mercy Corps staff and reflects the product available at that time.
2. MĂ©decins Sans Frontières – international humanitarian organisation
At Médecins Sans Frontières, the immediate barrier was not whether Copilot could write a good email. It was whether connecting a powerful search tool to the organisation’s existing information could surface material people did not realise was available to them.
MSF is decentralised, with separate entities using cloud infrastructure hosted through a common Microsoft tenant. Copilot can retrieve information a user already has permission to access, but large organisations often have years of imperfect SharePoint and OneDrive permissions. A document being technically accessible does not mean its author expected it to surface instantly in response to a natural-language question.
At the time of MSF’s presentation, uncertainty about who controlled access, who administered Copilot and how permissions would be governed had stalled wider adoption. MSF saw potential in office tasks, fundraising, communications and translation, but the organisational risk came first. It developed a stateless internal service where prompts and data could be handled locally, used the US National Institute of Standards and Technology’s AI Risk Management Framework, and began a six-month test with selected staff.
This is a useful corrective to the idea that adopting Copilot is mainly a prompting or training challenge. If your SharePoint is full of old folders, inconsistent access controls and documents whose owners have moved on, Copilot does not tidy that up. It makes the consequences easier to discover. The human responsibility here sits with leaders, information owners and IT teams deciding what staff should be able to access before the tool makes that access frictionless.
Source: INGO Experiences Adopting Microsoft Copilot, MERL Tech. This account comes from a July 2024 presentation by MSF’s Director of Systems.
3. Scottish Council for Voluntary Organisations – Scotland
SCVO uses a much more contained application: Microsoft Teams transcription and Copilot-generated meeting summaries. Capturing a useful record takes time, and the person writing notes is never fully present in the conversation. The organisation chose this approach specifically to keep meeting data inside its existing Office 365 environment.
The interesting part is not the summary itself. It is the boundary SCVO has drawn around it. Participants are told what is happening and asked for consent. There is a retention policy for the raw transcript. Staff review the draft summary to make sure it matches their recollection of the conversation.
Most importantly, SCVO says it would probably not use the process for a high-stakes call where an accurate record was critical. That is proportionate governance in a sentence. Copilot can help someone remember the actions from a routine project meeting. It should not quietly become the definitive record of a safeguarding discussion, disciplinary process or another conversation where the exact words matter.
The strongest AI policies do not simply divide tools into approved and banned. They help staff recognise when a normally acceptable use becomes too risky. SCVO’s approach leaves the final judgement with the person organising the meeting and the people checking the record.
Source:Â AI Playbook for Charities 2026.
4. London Funders – UK
London Funders had more than 130 responses from grantees to analyse. Free-text responses are valuable because people can describe their work and experience in their own language, but they are also slow to analyse consistently. It is easy for a small number of memorable comments to shape an organisation’s view while patterns across the full dataset remain hidden.
The team used Copilot to group the responses into broad themes. It completed the initial analysis much faster than a manual process, and the themes broadly matched London Funders’ programme areas and what staff expected to find.
But it also grouped similar-sounding ideas too loosely and missed nuance in the way organisations described their work. That matters. Two charities can use similar language while describing very different needs, communities or power relationships. A thematic summary can make a dataset look tidier by smoothing away the distinctions that made the original responses useful.
London Funders’ advice was to be specific about what you ask and to understand the data yourself before asking AI to interpret it. In other words, Copilot can provide a fast first pass, but the person reviewing it needs enough subject knowledge to notice what has disappeared. The human is not there to rubber-stamp the themes. They are there to challenge the model’s framing and return to the original responses when something has been flattened.
Source:Â AI Playbook for Charities 2026.
5. RNIB – UK
RNIB’s Eye Health Information team sometimes uses Microsoft Copilot to suggest simpler wording and help content meet a target reading age of 9 to 11. It is a modest use case, but one where the details matter. Eye-condition information needs to be understandable to a wide audience without becoming medically misleading.
RNIB is unusually clear about the division of labour. It does not use AI to generate its eye-condition content. Copilot can suggest an alternative phrase, but a person checks whether the new wording preserves the intended meaning. Every piece remains written and approved by people. Accuracy checks involve people living with sight conditions, medical and healthcare professionals and RNIB’s internal teams.
The organisation also states that private, sensitive or confidential information is not uploaded to AI tools. This keeps Copilot away from individual medical advice and personal cases. Its job is to help improve the accessibility of general information that has already been developed through a documented, human-led process.
This is a strong example because nobody is pretending the tool has expertise it does not have. Copilot helps propose words. People with medical knowledge, editorial responsibility and lived experience decide whether those words are accurate and useful.
Source:Â RNIB, How we produce our eye condition information.
6. American Institutes for Research – US
The American Institutes for Research used Microsoft 365 Copilot in a published qualitative study of Chicago’s out-of-school-time system. The work involved six focus groups, eight interviews, documentary research and discussions with people across the system.
Copilot was used to transcribe recordings, help code the interview and focus-group transcripts, and organise and shorten the findings. The research team prompted it to look across themes including funding, coordination, data systems, policy, the needs of young people and families, and gaps not captured by the initial categories.
Researchers who had conducted the interviews and focus groups reviewed and adjusted every output to check the coding. The findings were then taken into two sessions with 15 participants, followed by an in-person gathering involving 31 leaders, where they could be tested against people’s lived knowledge of the system.
Copilot did not replace qualitative research. It helped the team organise a large body of material, while researchers remained responsible for interpretation and participants had an opportunity to challenge whether the findings reflected reality. AIR also disclosed its use of Copilot in the final report, something far more organisations should do when AI has played a meaningful role in producing research.
Source:Â Chicago Out-of-School-Time Landscape Study Findings, American Institutes for Research.
7. Big Brothers Big Sisters of Puget Sound – US
Big Brothers Big Sisters of Puget Sound had identified a fundraising problem. Donors wanted to hear more stories about the organisation’s programmes, but producing a steady flow of useful blog posts, emails and updates takes time from an already busy development team.
Staff used content-generation features in Microsoft Copilot to turn that insight into a more regular stream of programme stories and donor communications. The ability to work with organisational information inside its Microsoft environment was part of the attraction, rather than staff moving data into a collection of free consumer tools.
There is an important limit to what has been published. We know the organisation produced more content, but there is no reported evidence yet that the content improved donor retention, increased income or changed the way supporters understood its work. More output is not the same as better stewardship.
An independent Chronicle of Philanthropy article also describes the organisation using AI to reduce a list of hundreds of possible volunteer and child matches to a shortlist for staff. But it does not establish that Copilot powered that system, so it would be misleading to present the matching work as another Copilot use case. The human-led principle still applies to the communications work: Copilot can help produce material, but people remain responsible for consent, accuracy, dignity and whether a person’s story should be told at all.
Source:Â Nonprofits Are Turning to AI to Save Time and Money, Chronicle of Philanthropy.
What these seven examples tell us
The clearest Copilot successes here are not dramatic. They are bounded tasks close to work that already happens in Microsoft 365: capturing the actions from a meeting, finding a document, producing a first draft or proposing simpler language. The user understands the task, can see the output and can usually spot when it is wrong.
The weaker results came when organisations asked Copilot to move from organising information to interpreting it. Mercy Corps needed more than a basic summary of a complex programme. London Funders found that broad themes could hide meaningful differences. The government evaluation found that confidence and speed did not guarantee accurate Excel analysis.
MSF’s experience points to another lesson. Copilot’s ability to search across organisational information is only as safe as the permissions and data practices underneath it. AI readiness is partly data governance, records management and information security under a new name. If those foundations are weak, connecting a conversational interface to everything can make the weakness much more consequential.
There is also a measurement problem. Several organisations can show that Copilot helped them create a summary or produce more content. Far fewer can show what happened because of the time saved. Did services improve? Did staff workload fall? Did more people receive support? Did fundraising income increase? Or did the expected volume of emails, reports and meeting notes simply rise?
That does not make the smaller gains worthless. Saving someone from an hour of meeting administration can be a perfectly good outcome. Helping a member of staff write more accessible information can matter enormously. But we should name the outcome we actually have rather than automatically translating time saved into greater impact.
The strongest examples share the same pattern as the best uses of Claude, ChatGPT or any other generative AI tool. The organisation begins with a problem it understands. The tool is given a narrow job. People check the source material and remain responsible for the judgement that matters. And someone is prepared to say where the tool should not be used.
SCVO will not rely on it for a high-stakes record. RNIB will not let it generate medical information. AIR’s researchers reviewed the coding and then took the findings back to participants. Mercy Corps was prepared to conclude that some advertised uses were not good enough.
That is what a human in the lead looks like. Not a person added at the end to approve whatever the system produced, but people deciding what problem to tackle, what information the tool can touch, how its work will be checked, and when it is safer or simply better not to use it at all.
Source for the opening evaluation:Â Department for Business and Trade, Evaluation of the M365 Copilot Pilot.