Using an AI assistant at work

Many people now have an AI assistant at work. It can draft, summarise and search, and it can be wrong. UK public bodies have set out some habits for using one. Your employer's own policy on AI comes first, and each habit below sits inside it.

What to keep out of a public tool

The National Cyber Security Centre (NCSC) wrote about large language models on 14 March 2023. It explains that a question typed into a public tool is visible to the company that runs it. The NCSC says those questions are stored and will almost certainly be used to develop the service at some point.

The NCSC recommends "not to include sensitive information in queries to public LLMs". It also recommends "not to submit queries to public LLMs that would lead to issues were they made public".

Check the answer before you use it

The NCSC says these tools "can get things wrong and ‘hallucinate’ incorrect facts". The judiciary in England and Wales has its own guidance on AI, updated in October 2025. It is written for judges and the staff who support them. It warns of "AI hallucinations which generate incorrect or misleading information". It also reminds judicial office holders "not to enter private information into public AI tools".

The same habit works in any line of work. Read what the assistant wrote, check the facts and names in it, and only then send it on.

Stay in charge

The government's AI Playbook was published on 10 February 2025 for people working in government. Its first principle is "You know what AI is and what its limitations are". It adds that AI systems are not guaranteed to be accurate. Its fourth principle is "You have meaningful human control at the right stages". In everyday work, a person reads what the assistant produced and decides what happens to it.

In Great Britain, Acas gives employees and employers free, impartial advice on workplace rights. That can help if a question about AI turns into a disagreement at work.

Personal data is still personal data

Data protection law still applies when information about customers, colleagues or the public goes into an AI tool. The Information Commission (which took over from the ICO on 30 September 2026) keeps its guidance on AI in one place. It includes an overview of how to apply the principles of the UK GDPR to the use of information in AI systems. The AI Playbook suggests asking a lawyer or your data protection officer for advice.

Hidden instructions in a document

The Open Worldwide Application Security Project (OWASP) runs a Gen AI Security Project, which keeps a list of risks for applications built on large language models. Its 2025 list starts with prompt injection. It says indirect prompt injections happen when the model accepts input from external sources, such as websites or files.

The AI Playbook gives an example. An email attachment can carry a prompt, and the prompt runs when an assistant reads the attachment to summarise it. Take care with documents from outside before you hand them to an assistant.

Useful links