Eight considerations before launching a public-facing AI tool

1. Define what success looks like

"Improve efficiency" and "modernise our service" aren't specific enough to guide good decisions. The most useful goals describe cause and effect for each person using the service. For example, “as a result of residents liable for council tax engaging with an LLM on a local authority website, we should see fewer ineligible people attempting to submit a discount application”. From goals like this, you can build meaningful OKRs and track the right things, like user adoption, conversation completion rates, reduction in contact centre queries, satisfaction scores and response accuracy.

2. Design for real user journeys, not controlled pilots

Internal pilots with staff differ significantly from public-facing services. Think carefully about where users encounter the AI and whether it supports or interrupts the broader journey. What happens when the LLM can't resolve a query? Is there a clear route to a human, or is it a dead end?

A tool that sounds robotic or dismissive can damage trust and cause real harm for vulnerable users, so prompts and training data need to be specific and accurate. What works in a pilot may not translate across the full diversity of your user base, so testing with real users in real scenarios, including edge cases, is important before you commit to launch.

3. Build infrastructure that can scale

A pilot handling a few hundred queries is fundamentally different from a live service facing tens of thousands of users. Public sector websites experience significant traffic spikes when policy changes, deadlines approach or crises emerge. Think carefully about your hosting setup and whether it can scale quickly, your fallback options if your AI provider goes down, how you're managing API costs if usage exceeds projections, and how the tool integrates with existing systems.

Slow responses will drive users away, particularly on mobile devices, which is where a significant proportion of public sector traffic comes from, and where digital exclusion often correlates with the populations most in need.

4. Manage security, privacy and compliance

Every query entered into an AI tool may reveal sensitive information about health, finances, housing or more. Beyond GDPR and the Data Protection Act 2018, it's worth thinking about protections against prompt injection attacks, where malicious users attempt to manipulate the AI into generating harmful outputs or exposing restricted information. Safeguarding also requires consideration. If someone uses your AI to signal immediate risk of harm, do you have mechanisms to detect that and escalate appropriately?

Consider whether the tool works for users using screen readers, those with cognitive impairments, low literacy or limited English, and those on older or lower-powered devices. Poorly designed AI tools risk unlawfully excluding the very people who most need access to public services.

Conduct a Data Protection Impact Assessment before launch to ensure security, safeguarding and incident response processes are robust, tested and defensible under public scrutiny.

5. Be transparent about what the tool is and isn't

Public trust in government technology has been shaped by past failures. Users approach new AI tools with justified caution. Users should know they're interacting with AI and understand what it can and can't do. Make clear whether it's offering guidance rather than definitive decisions, and make sure disclaimers are visible, not buried in terms of service.

Citing sources also helps. In our work with RNIB, we built a knowledge base for call centre teams that provides clear citations so advisers can validate responses and explore reference material when needed. If your AI makes claims about legislation or policy, people need to be able to see where that information comes from.

6. Prepare your team, not just your technology

Frontline advisers and contact centre teams need clarity on the tool's capabilities and limitations, as well as clear processes for handling escalations, complaints and feedback.

Clear governance is equally important. Who monitors performance? Who responds when accuracy issues are raised? Who responds to safeguarding issues? How are legislative or policy updates reflected in the system?

Senior leadership should understand the risks and the mitigations in place, and be able to explain their decisions if they're subject to Freedom of Information requests or parliamentary scrutiny.

7. Roll out incrementally

With public-facing AI, starting small is almost always the right call. Consider launching with one topic area, one user cohort, or one section of your website.

We took a similar approach with the London Museum, launching conversational search to help users explore collections more intuitively with a small subset of webpages before broader rollout. Have clear criteria for pausing or pulling back. If accuracy drops, safeguarding concerns emerge, or user satisfaction falls, you need the authority and processes in place to stop quickly.

Start small, test rigorously with real users, and build in the feedback loops that let you iterate and improve as you go, expanding when the evidence supports it.

8. Consider whether AI is actually the right answer

The Government Digital Service principle of "digital by default, not digital only" is worth keeping in mind. Not everyone can or wants to interact with AI, and alternative routes should remain available. These considerations aren't about slowing innovation. They're about making informed decisions before committing public resources, organisational reputation, and public trust to a service that may not be ready. Public sector organisations that do this well involve legal, security, content, safeguarding and frontline teams early. They pilot carefully and scale when the evidence supports it, prioritising transparency, accessibility and public safety alongside efficiency.

If you are considering launching a public-facing LLM, treat these considerations as part of maximising the likelihood of success. When implemented well, AI can reduce barriers to access, provide support outside working hours and allow teams to focus on complex cases requiring human judgment. But that depends on readiness, governance and a sustained commitment to earning and maintaining public trust.