Using AI in public services: where it helps and where it doesn't

AI tools have become capable enough that most public bodies are asking how to use them. The pressure is understandable. Teams are stretched, demand is high and residents expect services to be quicker and easier to use.

But public services are not like most businesses. Decisions affect people's lives, access to support and trust in institutions. Mistakes are harder to undo and harder to explain.

Our approach is simple. We use AI where it reduces effort and improves access, and we keep AI away from anything that has to be authoritative or auditable. We always leave a clear route to a person.

Where AI helps

Grounded assistants that cite their sources

A general-purpose chatbot that answers from its own training data is a risk in a public service. It can produce confident, plausible answers that are wrong.

A grounded assistant is different. It answers only from a defined set of approved content, such as the organisation's own web pages, policies and guidance, and it shows where each answer came from. If the content does not cover the question, it says so and points the user to a person.

Done well, this helps residents find information faster and reduces repetitive enquiries, while keeping the organisation's published content as the single source of truth.

Drafting with human review

AI is good at producing a first draft. Letters, summaries of long documents, meeting notes, internal reports and responses to routine enquiries can all start from an AI draft.

The key word is draft. A person reviews, edits and takes responsibility for what is sent. The tool saves time on the blank page. It does not replace professional judgement.

Plain-language search

Many public sector websites and internal knowledge bases are hard to search. People do not know the official name for the service they need.

AI can help match plain-language questions to the right content, even when the words differ. This improves findability without changing what the content says. It pairs well with good content design and a clear site structure, which is something we focus on in public service website launches.

Translation drafts reviewed by qualified translators

Machine translation can speed up making information available in other languages. It should not be published without review for anything important.

We use AI to produce a first draft and have it checked by a qualified translator before it goes live. That is especially important for health, safeguarding, legal and financial information, where a mistranslation could cause real harm.

Where AI does not belong

Authoritative decisions

We do not use AI to make or recommend decisions about:

Eligibility for services, benefits or support
School admissions or allocations
Funding, grants or financial awards
Safeguarding or risk assessments about individuals
Enforcement or sanctions

These decisions need to be explainable, consistent, open to challenge and accountable to a named person or body. Current AI models cannot reliably meet that standard. Even as an advisory input, they can introduce bias that is hard to detect and can anchor a human reviewer's judgement.

Where rules-based logic is needed, such as checking whether an application has all required documents, we use conventional, testable code rather than AI.

Anything without a route to a person

Every AI-assisted interaction should make it easy to reach a human. Residents should never feel trapped in a loop with a tool that cannot help them. This matters most for people who are vulnerable, in distress, or less confident with digital services.

Evaluation before go-live

An AI feature should not go live on the strength of a good demo. We evaluate it against the real task before launch:

Build a test set of realistic questions or inputs, including difficult and edge cases
Define what a good answer looks like, with input from the service team
Measure accuracy, whether sources are cited correctly, and how the tool handles questions it cannot answer
Test for harmful, biased or inappropriate outputs
Check accessibility and usability with real users where possible

Evaluation continues after launch. We monitor outputs, review feedback and re-test when content or models change.

Data protection and the DPIA

AI features often involve personal data, even when that is not the intention. Users type personal details into chat boxes. Documents being summarised contain names and case information.

A Data Protection Impact Assessment is usually appropriate. It should cover what data the AI processes, where the model is hosted, whether inputs are retained or used for training by the provider, how outputs are stored, and how individuals are informed. We build this assessment into every AI project delivered through our data and AI services.

Organisations should involve their Data Protection Officer early and check current ICO guidance on AI and data protection.

Transparency

People have a right to know when AI is being used in services that affect them. The Algorithmic Transparency Recording Standard provides a structured way for public bodies to publish information about the algorithmic tools they use, what they do and how they are governed.

Even where publication is not mandatory for an organisation, following the standard is good practice. It builds trust and forces useful clarity about what a tool actually does.

At service level, we recommend:

Telling users clearly when they are interacting with an AI tool
Explaining what it can and cannot help with
Showing sources for any information provided
Making it easy to give feedback or report a problem

A practical starting point

If you are considering AI for a public service, start with a narrow, low-risk use case where the benefit is clear and a person stays in control. Internal drafting tools and grounded search over published content are often good first steps. Learn from those before going further.

Policy and guidance on AI in the public sector is developing quickly, so check current government and regulator guidance before starting a project.

Talk to us

If you want to explore where AI could genuinely help your service, and where it should not be used, talk to us. You can also read about how we work.