“TechFest 2025 has been a fantastic opportunity to bring together the whole directorate, its given us the opportunity to celebrate our successes and achievements, hear from Matt our Exec Director about our future direction, and gain insight and knowledge into topics, from AI to prevent burnout, managing experience management to understanding how to make the most out of Copilot. We are incredibly grateful to our partners and suppliers who delivered some incredible sessions! Thank you – we couldn’t have done it with out you”
Sally Bogg, Head of IT, NHS England
NHS England’s Technology Directorate recently hosted its first internal TechFest, a week that brought people together to look at how technology and data can support better services.
Hippo presented three talks across TechFest. Together they explored what artificial intelligence is and how it works, how to get more from language models through better prompts and how to move data projects forward in health when live datasets are difficult to use. Each session approached that theme from a different angle, so people could move between understanding, interaction and applied data work in ways that suited their role.
Making artificial intelligence understandable
When we talk about AI, what are we really describing?
Senior Data Science Engineer, Magdalena Nowakowska, began with human intelligence rather than technical detail. She asked people to think about how they gather information through their senses, process the information and give it meaning, learn to recognise patterns and based on previous experiences make educated guesses about the furniture. Our brains take in a lot of information, pick out what matters, connect it to past experience and decide how serious the request is. Over time we learn from what happens and adjust that judgement.
Magdalena used this pattern to show how many AI systems work. They also start with data, transform it, identify useful features and learn relationships during training. When they receive something new they use those learned patterns to make a prediction. The underlying ideas have existed for a long time. What has changed is the scale of data and the computing power that make large models possible.
Trust sat at the centre of the conversation: Because these systems learn from existing data they can repeat and strengthen patterns that are already present, including unfair ones. In health that carries obvious risk. The message was to treat AI as a tool that helps people work with information, not as an independent decision maker. Professional judgement, context and governance remain essential. Framing intelligence in this way made it easier to move past hype and gave teams language they could share across their work.
Designing better prompts with language models
Our ‘Become a prompt engineer’ talk, focused on how people in the NHS can use language models more deliberately. Principal Content Designer, Josh Healey, drew on his experience working where content design and AI meet.
Josh began with how these models behave. They do not recall information in the way a reference book does. They predict the next likely word or token based on patterns learned from large collections of text. The clarity of the response is therefore closely tied to the clarity of the request.
From there, he treated prompts as something that deserves design attention. A prompt sets the task, frames expectations and signals tone. To make this practical he shared a straightforward structure people could apply. Be clear about why you are asking for help. Describe the task in concrete terms. Offer enough context so the model knows who and what the content is for. Set instructions so it understands what to do with that context. Add sensible limits on length, format or style. Make the audience visible so the output is written with the right reader in mind.
Josh used examples from contexts such as policy summaries, internal guidance and service communications to show how the structure works in practice. In each case, small changes like adding a line of background or tightening the task helped make the resulting outputs more precise and easier to review. Prompting was treated less as trial and error and more as a designed part of content and service work, an area where the herd already partners with teams.
Finding options when real data is not available
There is strong interest in AI and advanced analytics, yet in practice the data needed to explore those ideas can often be unavailable or inaccessible. This is especially the case with personally identifiable data, with healthcare a prime example of this. Furthermore, processes may be new, data may sit in separate systems and there are important questions around privacy and governance.
Technology Director Ed Hiley and Lead Data Science Engineer Suze Hawkins shared three ways to keep moving within those constraints.
The first was using pre-trained models, where appropriate. In some areas existing models can address common problems without the cost of building something new. They still need careful testing but they can help teams run early experiments and make more informed decisions about when to buy and when to build.
The second approach was knowledge transfer. Here the starting point is the insight that already exists in clinical and operational teams. That expertise can be turned into rules or signals that stand in for large labelled datasets. Those signals then feed analytic tools, with people reviewing and refining the results as more evidence is gathered.
The final strand was synthetic data. Synthetic data is designed to behave like real data and follow realistic patterns. It can be critical when teams need something that looks and feels like a live dataset for design and testing but cannot yet work with actual data.
To show how this can work, Ed described the Virtual Ward OMOP Generator project. A virtual ward team wanted to understand how risk might change over time for people receiving care at home and how that information could support clinical decisions..
Working closely with clinicians, the team instead defined realistic patient journeys and created synthetic records that followed those journeys. They used this to build code, explore different approaches to risk scoring and create visualisations that clinicians could respond to. When live data becomes available, access to tools, pipelines and questions will be ready to test and refine. Throughout, there was clear oversight and a focus on improving care rather than experimenting for its own sake. Synthetic data did not replace real analysis. It created space to learn earlier and more safely.
Carrying the work forward
Across TechFest a consistent pattern emerged. People were not looking for grand statements about AI. They wanted language that made intelligence understandable, ways to use models more carefully and routes for progressing data work within the realities of health services.
Hippo’s sessions supported that pattern from different directions. Together they moved from explanation, to practical use, to the realities of working with data in clinical contexts. The discussions that followed were grounded in responsibility, value and the needs of patients and staff.
That is the space the herd will continue to work in with NHS colleagues. Helping people understand AI and data clearly, explore ideas in low-risk ways and design services that stay human at their core. TechFest showed the appetite for this kind of learning and highlighted how much can be achieved when capability, care and delivery sit side by side.
