Skip to content
29Jun 2026

AI for charities: practical uses, risks and first steps

AI for charities is most useful for reducing routine administration: sorting enquiries, drafting correspondence, summarising documents and finding approved information. Start with a narrow task that staff can check, rather than decisions about beneficiaries, safeguarding or access to support.

This guide is for UK charity leaders, operations managers and trustees choosing a practical first project. It covers data protection, integration and costs, with examples relevant to charities and membership organisations.

TL;DR

  • Start with repetitive administration where mistakes are easy to spot and reverse.
  • Keep safeguarding, eligibility decisions and sensitive casework under accountable human control.
  • Check data protection, supplier terms and access permissions before sharing information.
  • Use approved sources and existing systems rather than creating another disconnected inbox.
  • Measure checking time, errors and service quality before expanding.

Where AI can help charities in practice

The strongest starting points usually sit behind service delivery, not inside decisions about people. Look for work that follows a recognisable pattern and involves copying, searching or drafting from maintained information.

  • Enquiry handling: suggest categories and draft replies about opening hours, donation receipts or volunteering.
  • Document work: summarise consultation papers, meeting notes and funder guidance, with references to the originals.
  • Staff knowledge: find answers in approved expense policies, volunteer handbooks and operating procedures.
  • Fundraising administration: draft acknowledgements and adapt approved campaign material for review.
  • Reporting: turn checked activity records into a narrative draft without inventing outcomes.

A membership charity might start with renewal enquiries. An advice charity might choose internal policy search, keeping beneficiary records outside the initial project. A fundraising team might prepare correspondence without allowing the tool to send messages.

Choose work with enough repetition to justify setup, training and maintenance. Our guide to document summarisation with AI explains source references and omission checks. Treat summaries as reading aids, not substitutes for checking requirements before a funding submission.

Protect beneficiary, donor and volunteer information

A charity enquiry can contain health information, financial hardship, immigration details or allegations of abuse. An ordinary-looking inbox is not necessarily a low-risk dataset. Removing names may still leave enough detail to identify someone.

The ICO’s AI guidance explains how data protection principles apply, including fairness, transparency, accuracy and security. Use it with your data protection lead or adviser to assess purpose, lawful basis and any additional conditions for special category data.

Before choosing a tool, map what information enters it and whether the task could use less. Ask the supplier:

  • Who can access prompts, uploaded files, outputs and logs?
  • Is your information used to improve models?
  • Where is information processed, how long is it retained and how does deletion work?
  • What contractual terms and international transfer arrangements apply?
  • Can access and retention settings be controlled centrally?

The ICO’s DPIA guidance explains when a data protection impact assessment is required, including processing likely to result in high risk to individuals. Assess the proposed workflow rather than assuming all AI projects have identical requirements.

Pro Tip: Test with invented enquiries first. Obtain internal approval before introducing real personal information.

An approved public handbook is generally a more manageable starting source than beneficiary case notes. These are practical risk considerations, not legal advice or confirmation of compliance.

Set boundaries before staff start experimenting

A useful AI policy tells people what they may do on a busy working day. Name approved tools, permitted information, review requirements and the person to contact when something goes wrong. “Use AI responsibly” is not enough.

Assign an operational owner who understands the workflow and identify responsibility for information governance and supplier management. Arrange trustee oversight appropriate to the risks, without making trustees approve every draft email.

Our UK AI policy guide helps turn those decisions into working rules. Give volunteers and temporary staff the same relevant instructions as employees.

Keep these activities outside the first project

  • Assessing safeguarding disclosures or deciding how urgently someone needs protection.
  • Accepting or rejecting applications for support without accountable human assessment.
  • Approving payments or changing bank details from an AI recommendation.
  • Making employment decisions from unverified summaries or inferred characteristics.

The Charity Commission’s internal financial controls guidance covers financial controls for charities in England and Wales. Use it when reviewing finance workflows; charities elsewhere in the UK should consult their own regulator’s guidance.

Keep payment authorisation and separation of duties outside the model’s discretion. Agree a stop condition: exposure of restricted information or repeated mishandling of sensitive enquiries should trigger a pause and investigation.

Connect AI to the systems you already use

The useful part of an AI project is often the integration, not the chat window. If staff must copy every answer into a CRM, find the original document and update a spreadsheet, much of the apparent benefit can disappear.

Map the complete journey for a membership enquiry: receiving the email, locating the record, drafting the response and recording the approved reply. Specify which steps use AI, which use ordinary software rules and which require a person.

Use structured rules for exact facts such as membership status, subscription balances and payment dates. AI can explain a verified result, but it should not guess missing account information or assume that similar names identify the same person.

An internal AI assistant should retrieve only information the user is authorised to see. Start with maintained policies, assign source owners and remove superseded versions. Test access restrictions using accounts with different permissions.

Our guide to AI email triage covers inbox workflows. Keep a monitored route for messages the system cannot classify, rather than leaving them in an unattended queue.

Pro Tip: Give the first integration read-only access. Require staff approval before it changes records or sends messages.

Record actions and source references so colleagues can understand what happened. Limit logs to what is needed for that purpose, with appropriate access and retention controls.

Choose tools around the job, not the demonstration

Begin with capabilities already available in your organisation, but do not assume an existing subscription makes every use appropriate. Personal accounts, business services and connected assistants can have different data handling and administrative controls.

If your charity uses Microsoft 365, check the precise product and configuration under consideration. Microsoft’s privacy and security documentation describes Microsoft 365 Copilot. Review document permissions before connecting search: an assistant can make existing access problems easier to discover.

Our Microsoft Copilot for business guide discusses chat access and governed implementation. Check charity eligibility and current terms directly with suppliers rather than assuming discounts cover the proposed service.

A public chatbot is not automatically the best answer. Clear service pages, an accessible contact form or reliable search may solve the problem with less maintenance. If a chatbot is justified, restrict its sources and provide a clear handover to a person.

Our comparison of website chatbot options sets out the choices. Test whether people can find opening hours, understand service boundaries and reach staff without being trapped in a conversation.

Budget for setup, integrations, training, checking and support alongside licences. Ask who maintains the system when policies change, the supplier updates its model or the responsible employee leaves.

Run a pilot, then manage it as a service

Choose one workflow, one owner and a defined user group. Record the current process, including awkward cases and correction time. Our AI readiness assessment guide helps identify missing information, permissions or ownership before implementation.

A pilot brief should state the task, permitted information, prohibited actions, review process and success criteria. For example, draft routine volunteering replies from approved guidance, with every response checked before sending.

Measure the whole job, not just generation speed. Compare the original and proposed workflows using a representative mix of tasks, including difficult cases:

  • Time spent preparing, checking, correcting and recording responses.
  • Incorrect answers, omissions and unsupported statements.
  • Messages escalated correctly, missed or escalated unnecessarily.
  • Whether staff can find and verify source information.
  • Extra work created for other teams.

Include an ambiguous request, an outdated document, a complaint and a message containing sensitive information. Test what happens when no reliable answer exists. If checking takes longer than writing, improve the sources, simplify the task or stop.

Plan for ordinary operation

Someone must update sources, review errors, manage access and respond to supplier changes. The NCSC’s secure AI development guidelines cover security across design, development, deployment and operation. They provide useful questions for suppliers about maintenance and incident handling.

For assistants reading incoming messages or external documents, test attempts to make them ignore instructions or reveal information. Treat that content as untrusted: it must not grant permission to access records or trigger actions. Keep a fallback process for outages and remove access when staff or volunteers leave.

Expand only after the original workflow proves useful in routine operation. Assess each new use separately: drafting volunteer replies and summarising confidential casework carry different risks.

My starting point for a charity AI project would be the process, not the model: who owns it, which information is reliable and what staff must check. I would rather pursue a modest improvement to routine enquiries than build an impressive demonstration around vulnerable people’s records. Where ordinary software rules or a human decision are more appropriate, use them. AI does not need to be in every step.

- Rob, Cloud 9

About Cloud 9

Cloud 9 is a UK AI consultancy helping businesses and membership organisations assess and implement practical AI workflows. For charities, relevant services include readiness assessment, administrative automation and assistants connected to approved information.

The starting point is the process, including where ordinary automation or human judgement is more suitable. To discuss a specific workflow, book a discovery call covering its potential value, risks and implementation requirements.

Frequently asked questions

What is the best first use of AI for a small charity?

Choose a frequent, low-risk administrative task using approved information, such as drafting volunteer replies or searching a handbook. Establish the current workload and choose outputs staff can check easily. Keep sensitive beneficiary information outside the initial test wherever possible.

Can charities put beneficiary information into ChatGPT?

Availability alone is not a reason to upload it. Review the specific service, purpose, contractual terms, access controls and retention arrangements with your data protection lead or adviser. Use the ICO guidance cited above to assess the proposed processing, especially where sensitive information is involved.

How much does AI cost a charity?

Costs depend on the workflow, supplier, integrations and oversight required. Include licences, implementation, training, review time and maintenance. Compare total operating effort with the existing process: inexpensive drafts may still create substantial correction work.

Can an AI chatbot answer questions from people seeking help?

It can help with bounded questions about service locations, opening hours and published application steps. Use approved sources, explain its limitations and offer a human contact route. Do not position a general chatbot as a safeguarding service or let it determine entitlement to support.

Sources

  1. Artificial intelligence, ICO.
  2. Data protection impact assessments, ICO.
  3. Internal financial controls for charities (CC8), Charity Commission for England and Wales.
  4. Data, privacy and security for Microsoft 365 Copilot, Microsoft Learn.
  5. Guidelines for secure AI system development, NCSC.