Can your AI be trusted? A practical guide for UK SMEs and charities

AI is now part of everyday work for many businesses and charities in Wales and the North West. Teams use it to draft letters, answer enquiries, summarise case notes and find information faster. The useful question is no longer "should we use AI?" but "how do we know it's doing what we think it's doing?"

What "trustworthy AI" actually means for a small organisation

For an SME or charity, trust comes down to three things:

• It handles information safely. Personal and sensitive data goes only where you've agreed it can go, and stays under UK GDPR.
• It gives answers people can rely on. Output is accurate enough for the job, and someone checks it where mistakes would matter.
• It behaves as intended. It doesn't leak internal information, follow malicious instructions hidden in documents or emails, or drift into topics it shouldn't touch.

If you can't say how your tools meet each of these, you don't yet know whether to trust them.

Where AI risk usually hides

The risks we see most often aren't dramatic. They're everyday gaps:

• Staff pasting sensitive data into public tools. Client records, HR issues or safeguarding notes go into a free chatbot, and you have no idea where they're stored or whether they're used for training.
• Confident wrong answers. Large language models can produce plausible text that's simply untrue. A customer-facing assistant that invents a policy or price causes real problems.
• Prompt injection. When an AI reads emails, web pages or uploaded files, those can contain hidden instructions that make it reveal information or act in ways you didn't intend.
• No owner. Nobody is responsible for the tool, so nobody notices when it starts going wrong.
• Shadow AI. People use tools the organisation hasn't approved because they're quicker than asking.

Test it before you trust it

You wouldn't launch a new website without clicking every link. AI deserves the same treatment. Before putting an AI tool in front of staff or customers:

• Write a set of realistic questions, including awkward ones, and check the answers against what you know is true.
• Try to make it misbehave: ask for information it shouldn't share, feed it a document with odd instructions, ask questions outside its remit.
• Record what went wrong, fix the setup (instructions, data sources, permissions) and test again.
• Repeat on a schedule. Models and data change, so a tool that passed in spring may not pass in autumn.

Larger organisations do this formally as "LLM security testing". The idea scales down well: a structured afternoon of testing is far better than none.

Put some simple governance around it

There is now an international standard for managing AI responsibly, ISO/IEC 42001, which sets out how an organisation should run an AI management system. Few small businesses or charities need certification, but its thinking is useful at any size: know which AI you use, assess the risks, assign owners, and review regularly.

In practice, for a small team that looks like a one-page AI policy, a short list of approved tools, a named person accountable for each one, and a review date in the diary.

Keep sensitive data under your control

For charities handling client records and SMEs holding customer or staff data, the biggest single step is choosing where the AI runs. Public consumer tools are convenient, but you often can't control where data is processed.

A private AI setup keeps your data in an environment you control. At AIFor.Wales, our Private AI work is UK-hosted, and we don't train models on client data. That makes it far easier to answer your trustees, funders or customers when they ask where their information goes.