AI for charities: how small teams use AI agents to do more (2026 guide)

Charity teams are small, the workload is not, and a surprising amount of it is paperwork that sits next to the mission rather than inside it. Funder research, board packs, monitoring spreadsheets, supporter emails and unpaid invoices all need doing, and they tend to land on the same few people.

AI agents are good at exactly this kind of work. This guide explains what they are, where they help in a charity, where they cause harm, and how to start in a week. It is written for chief executives, operations and fundraising leads, communications staff and trustees.

James McAulay teaches teams to build agents at the Agent Accelerator, and registered charities get 30% off every training format. Everything in this guide is useful whether or not you ever take a course.

An agent does the task, where a chatbot only talks about it

A chatbot answers a question. You paste in some text, it replies, and you carry the result to wherever it needs to go. It forgets you between conversations unless you set it up otherwise.

An agent has three things a chatbot lacks. It has context, meaning your policies, past applications and templates. It has tools, meaning it can read a folder, search the web, query a spreadsheet or draft an email in your inbox. And it has instructions to follow a process, step by step, and to stop and ask when it is unsure.

In charity terms, a chatbot is a well-read volunteer who knows nothing about your organisation and starts every shift cold. An agent is a colleague who has read your last five funder reports, knows which trustees want which figures, and can go and fetch the monitoring data before drafting anything.

The practical difference is where the effort goes. With a chatbot you spend your time moving text around. With an agent you spend it deciding what good looks like, then checking the result.

Grant applications are now AI talking to AI, so quality beats volume

Agents are well suited to funder research. They can read a funder's published priorities, past awards and application guidance, compare them with your work, and tell you whether a fund is worth the hours. That filter alone saves time, because most bids that fail were never a good fit.

Drafting is the second use. An agent that has your previous successful applications, your theory of change and your latest outcomes can produce a first draft in your voice, matched to the funder's questions and word limits.

There is a catch, and it is worth being plain about it. Applicants are using AI to write, and funders are starting to use AI to read. When everyone can produce a competent-sounding 1,500 words in minutes, competent-sounding stops being a differentiator. A funder skimming a stack of near-identical applications notices the ones with something real in them.

That changes how to use the tools:

1. Use agents to apply to fewer, better-fitting funders. Let the agent do the research, then decide with your own judgement where to spend effort.
2. Add what only your team knows. A named local example, a quote from a service user (with consent), a number from your own monitoring data and a plain account of what went wrong last year are all things a model cannot invent for you.
3. Never let an agent invent evidence. Ask it to flag any claim it cannot source from your documents. A made-up statistic in a bid is a reputational risk and, with some funders, a compliance one.
4. Check each funder's stance on AI. Some may ask applicants to say whether AI was used, so follow the guidance in front of you.

The aim is speed on the plumbing and more attention on the substance.

Trustee and board papers can be drafted from the sources you already hold

Board papers are repetitive in a useful way. The structure repeats every quarter, the inputs live in known places, and the person writing them is usually also running a service.

An agent can assemble a first draft from the finance report, the risk register, the previous minutes and your reserves policy. It can flag what has changed since last time, list decisions the board is being asked to take, and draft the executive summary.

Three habits make this reliable:

• Keep a template per paper type (finance, safeguarding, strategy) so the agent follows your format every time.
• Make the agent cite where each figure came from, so the person signing off can check it in a minute rather than an hour.
• Keep the judgement human. An agent can tell a trustee what the numbers say. The recommendation, and the accountability for it, stays with the chief executive and the board.

Trustees can use agents too. A trustee who receives a 90-page pack can ask for a briefing on the three items that need a decision, with the relevant figures, before the meeting. That tends to improve the quality of the discussion.

Impact reports get faster when the agent reads your monitoring data first

Most charities collect more outcome data than they have time to analyse. It sits in spreadsheets, survey exports and case management systems, and the annual report becomes a scramble.

An agent can read those sources, calculate the figures you define, draft the narrative around them and produce a version for each audience: a funder report, a trustee summary and a supporter update, all from the same checked numbers.

The rule that matters most here is to agree the numbers before the prose. Decide which outcomes you report, how each is calculated, and where the source data lives. Then the agent's job is to apply your definitions consistently, not to choose them. If it cannot find a figure, it should say so rather than estimate.

Anonymise or aggregate data before it reaches any agent where you can. For reporting, you rarely need names.

Supporter, donor and volunteer admin is the safest place to begin

If you want a low-risk first project, this is it. The work is high volume, the stakes per item are low, and a human can check a sample easily.

Examples that work well:

• Sorting the shared inbox and drafting replies to common questions for a person to approve.
• Writing personalised thank-you messages from a donation record and your tone guidance.
• Summarising volunteer feedback forms into themes.
• Turning a long email thread into a short action list.
• Preparing event briefings from sign-up lists.

Keep the fundraising rules in view. Anything that sends mail to supporters needs to respect consent and preferences, so route sends through the system that already holds those records, and keep a person in the loop on anything that goes out in your name.

Chasing overdue invoices is the quiet win in finance

Charities that trade, through training, hire, contracts or consultancy, often have a finance person who is also doing six other jobs. Chasing overdue invoices slips, because it is awkward and never urgent until it is urgent.

An agent is patient at exactly this. It can read your accounting system, list what is overdue, draft a polite and specific reminder for each customer using the invoice details, and prepare a short summary for a person to approve before anything is sent.

One example from the Agent Accelerator is a debate-education organisation whose team built an invoice-chasing agent and recovered money that was over a year old. Chasing consistently is the part that tends to slip, and it is the part an agent does without getting tired of it.

The same pattern helps with reconciliation. An agent can match bank lines to invoices, flag the ones that do not match and explain why, leaving the finance lead to resolve the exceptions.

Start read-only. Let the agent report what is overdue for a few weeks before you let it draft anything, and keep a human approving every message until you trust the output.

Responsible AI for charities starts with data protection and a human reviewer

Charities hold some of the most sensitive data there is, about people at difficult moments. The guardrails below are the ones to settle before you build anything.

Beneficiary data stays out of consumer tools. Do not paste case notes, health information or anything identifying a service user into a free chatbot account. Use a work account on a business or nonprofit plan, and check the provider's terms on data retention and training.

Run a data protection impact assessment before any agent touches personal data. The Information Commissioner's Office publishes guidance on AI and data protection for UK organisations. Name who is responsible for it in your organisation.

Give agents the least access they need. Think of onboarding a new starter. They begin with read-only access to the folders they need, not the whole drive and not the finance system. Widen access once the agent has earned it.

Keep a person reviewing anything that leaves the building. Grant applications, supporter emails and public statements all need a human read. For anything affecting a beneficiary directly, such as eligibility or safeguarding, an agent should only ever prepare information for a person to decide on.

Write a one-page AI policy that covers approved tools, what data may be used, who reviews what, and how staff report a mistake. Share it with trustees. A short policy that people follow beats a long one that nobody reads.

Be honest about it. If AI helped draft something that matters, such as a funder report or a policy, say so where the recipient would expect to know.

Claude and Microsoft 365 Copilot are both credible, and many charities will use both

Most charities do not need to pick a winner. They need to know where each fits.

Microsoft 365 Copilot is where many charities already are. According to Microsoft's nonprofit offers page, Microsoft Copilot Chat is included in nonprofit Microsoft 365 Business Basic and Business Premium at no extra charge, and Microsoft 365 Copilot is available to nonprofits as an add-on at $25.50 per user per month on a yearly plan. It works inside the tools your team already opens, such as Outlook, Word, Excel and Teams. On 15 October 2026, Microsoft announced that Copilot Studio agents can now have their own accounts, process memory and triggers, which makes it practical to build agents that run recurring processes inside Microsoft 365 rather than waiting to be asked.

Claude suits teams who want agents that work across documents, spreadsheets, the web and connected tools, and who want more control over how the work is set up. Anthropic offers a nonprofit programme with discounted Claude Team and Enterprise plans. Its nonprofit page lists the Team plan at $8 per user per month for the Standard seat, with a minimum of two seats, and eligibility is verified through a partner called Goodstack.

How to decide:

• If your data, email and documents already live in Microsoft 365, start with Copilot Chat and see how far it gets you. The integration is hard to beat for everyday drafting and summarising.
• If you want agents that handle multi-step work across several systems, or your team is comfortable with a more hands-on setup, try Claude on a small pilot.
• If you are unsure, run the same small task on both for a week and compare the output, the review time and what your team prefers.

The skills carry over either way. Writing a clear brief, supplying the right context, setting limits on access and reviewing the result are the same on every platform. That is why the Agent Accelerator teaches the method first, uses Claude as its default, and runs Microsoft 365 Copilot sessions for charities who ask.

Prices and programme terms change. Check the Microsoft nonprofit offers page and Anthropic's nonprofit page before you budget.

Start in a week with one task, one safe workspace and one reviewer

You do not need a strategy document to begin. You need one task, one owner and a week.

1. Day 1: pick the task. Choose something repeatable that costs a few hours a week and is low risk if the draft is wrong. Chasing invoices, drafting the monthly board finance summary and triaging the shared inbox are all good candidates.
2. Day 2: set the boundaries. Decide what data the agent may see, which account it runs under, and who reviews its output. Write it down in half a page.
3. Day 3: gather the context. Collect three or four good examples of the finished task, plus any policy or template that governs it. Quality here decides quality later.
4. Day 4: build the first version. Write the instructions as you would brief a new colleague: the goal, the steps, what to check, and when to stop and ask. Run it on a real example.
5. Day 5: review with a person. Compare the output with what a person would have produced. Note what it got wrong and tighten the instructions.
6. The following week: run it alongside the old way. Keep the human process running in parallel until you trust the agent, then switch over with a reviewer still in place.

Build one agent, learn from it, then pick the next. If you want a structured route with other teams building alongside you, the Agent Accelerator curriculum sets out what the programme covers, and the live cohort and self-paced course, like team bootcamps, are 30% off for registered charities. The charities page explains how to apply for the rate.

Frequently asked questions about AI for charities

What can a charity use an AI agent for?

Charities use AI agents for funder research and first drafts of grant applications, trustee and board papers, impact reports built from monitoring data, supporter and volunteer admin, and finance jobs such as chasing overdue invoices. The best first project is a repeatable task that currently eats a few hours a week and carries little risk if the draft needs fixing.

Is it safe to put charity data into AI tools?

It depends on the data and the tool. Beneficiary and case data should never go into a consumer chatbot. Use a work account with a business or nonprofit plan, check what the provider does with your data, and run a data protection impact assessment before anything involving personal data goes live. A person should review every output before it leaves the building.

Should a charity use Claude or Microsoft 365 Copilot?

Both are credible. Many charities already have Copilot Chat included in their Microsoft 365 nonprofit plan, and it works well inside Outlook, Word and Teams. Claude suits teams who want agents that work across files and tools, and Anthropic has a nonprofit programme for Team and Enterprise plans. The skills of writing a good brief, giving an agent the right context and reviewing its work transfer between the two.

Will funders reject applications written with AI?

Funders vary, so check each funder's guidance. The bigger risk is a generic application. Funders can receive many near-identical AI drafts, so applications that show genuine fit, real local evidence and a specific voice stand out. Use AI for research and drafting speed, then add the detail only your team knows.

How much does AI cost for a small charity?

Prices change, so check the vendor pages. At the time of writing, Microsoft 365 Copilot for nonprofits is $25.50 per user per month on a yearly plan, and Anthropic lists Claude Team Standard for verified nonprofits at $8 per user per month with a two-seat minimum. Copilot Chat is included in nonprofit Microsoft 365 Business Basic and Premium.

Do you offer training for charities?

Yes. Registered charities get 30% off every format: the live Agent Accelerator cohort, the self-paced course and team bootcamps (10 people minimum). James McAulay teaches the cohort, and the course covers building agents on Claude, with Microsoft 365 Copilot sessions available on request.

How long does it take to get started?

One focused week is enough to pick a task, set up a safe workspace, build a first agent and review its output with your team. Expect the second and third agents to be faster, because the context and review habits carry over.