The Public Services AI Playbook: From Pilots to Lasting Value
Public-sector leaders do not need another presentation about AI’s potential. They need a clear way to decide where it can help, what responsible implementation looks like, and how to turn an initial experiment into a service improvement that lasts.
That is the focus of our Public Services AI Playbook video. It brings together the work from Datnexa’s summer 2026 AI Summer School: five sessions, 108 slides and feedback from 170 alumni across the UK and beyond.
The central message is simple: start with a real service problem, choose a proportionate intervention, and measure the impact.
For senior leaders, this means approaching AI as a service and organisational decision - not simply a technology purchase.
Start with the problem, not the product
The first question should not be “Which AI tool should we buy?” It should be “What are we trying to improve?”
In the video, I suggest a useful test: if AI did not exist, how would you solve the problem?
That question directs attention to the service itself: the people using it, the workflow, the evidence and the workarounds staff have developed. It also helps distinguish a genuine opportunity for AI from a process problem that needs a different intervention.
Before considering suppliers, leaders should be able to explain:
What is not working, and for whom?
What evidence shows that this is a problem?
Where does the current process create delays, duplication or unnecessary effort?
What would a better outcome look like?
The aim is to find technology that fits a clearly understood need—not reshape a service around a product before establishing whether it helps.
Bring disconnected experiments into a coherent approach
The playbook describes a familiar starting point: people experimenting with their own tools, using AI for drafting and summarising, and running pilots that may not connect with one another.
The leadership task is not necessarily to replace all of this with a large transformation programme. It is to give useful experimentation a clearer structure.
That means providing safe tools, sharing learning across teams and selecting a small number of problems worth solving. Individual projects should contribute to a wider direction without becoming an unfocused collection of demonstrations.
“Start small, dream big” is the principle here. A focused intervention can be useful in its own right while also helping an organisation learn what it will need for broader adoption.
Choose the right level of AI
Not every problem needs an autonomous agent.
The video distinguishes four broad ways AI can support work:
Assistance: helping someone complete a task, such as drafting an email.
Automation: carrying out a defined task or process.
Augmentation: supporting professional judgement.
Acting: taking steps on someone’s behalf through an agentic approach.
The important decision is which level the problem requires.
An elaborate system is not automatically a better system. In the playbook, I emphasise that additional complexity brings more to maintain, more cost and more opportunities for things to go wrong.
For a senior leader, the question is therefore not whether a proposal uses the most advanced technology. It is whether the proposed approach is proportionate to the task and delivers the intended benefit.
Improve the process underneath the technology
The playbook’s position is that AI should not be expected to remove problems that have not been addressed in the service design.
Before introducing it, bring together the people who understand the work. Look at the user journey, the information required, the hand-offs between teams and the points where work gets repeated or delayed.
This is also where success measures should take shape.
Rather than focusing only on licences purchased, usage levels or whether a system technically works, ask what difference the intervention should make. Depending on the service, the intended measures might include less avoidable work, more timely responses or a clearer route for residents to get help.
These are measures to define and test, not benefits to assume before delivery.
Understand your data without waiting for perfection
The video makes an important distinction: understanding data readiness is non-negotiable; having perfect data is not.
Leaders need a clear account of the information a proposed tool will use:
How complete and reliable is it?
Who owns it and maintains it?
Why was it originally collected?
Who can access it?
What legal and governance considerations apply?
Could its proposed use create new risks?
The point is not to wait indefinitely for an ideal dataset. It is to understand the limitations, decide whether they are acceptable for the intended use, and establish what needs to improve.
The playbook also highlights an organisational opportunity: examining how data is used can reveal gaps, overlaps and barriers between teams. That learning belongs in the service design, not just in the technical specification.
Treat governance as part of delivery
Governance is presented in the video as a way to protect people and improve design—not simply a final approval hurdle.
That requires a joined-up review. Passing a proposal sequentially between teams can leave gaps where each assumes that someone else has considered an issue.
The playbook calls for a broader view that brings together technical, service, commercial and governance considerations, with a named person accountable for the overall project.
That accountability must extend beyond launch. Leaders should be clear about:
Who owns the service and the AI tool?
Who monitors its performance and risks?
Who manages changes and ongoing costs?
Who responds when something goes wrong?
What happens when a pilot becomes an organisation-wide service?
A successful pilot is a decision point, not the end of the work.
Decide whether to buy, build or reuse
The video sets out three delivery routes.
Buy an off-the-shelf product
This may suit a common need where an existing product is a good fit. The key question is whether the organisation understands the process changes required—and will carry them through rather than leave staff relying on new workarounds.
Build a bespoke solution
This may suit a distinctive requirement where a tailored approach is justified. The playbook highlights the need to account for delivery, maintenance and long-term ownership, rather than considering only the initial build.
Assemble proven components
Between those options is an approach based on reusable components configured around a particular service.
In the video, I discuss Datnexa’s work with the AWS stack and Oscar AI for resident-facing communications, triage and signposting as an example of this approach.
The leadership decision is about fit, cost, risk and ownership. There is no need to choose a delivery model before understanding the problem.
Look beyond lawfulness to public trust
The playbook argues that being lawful is necessary, but not sufficient to earn trust.
Leaders should also be able to explain what a system does, who is accountable, how it is monitored and how an affected person can challenge what happens.
One question in the video brings this into sharp focus:
Could you defend this decision publicly, today, to the people it affects?
Another is more personal: would you be comfortable with the same tool being used on you?
These questions are not substitutes for formal assurance. They help expose whether the organisation can give a clear, credible account of its choices.
Use the five-point AI readiness check
The video brings these principles together into five questions:
Is the problem clear? Can you describe it in a sentence that someone outside the project can understand?
Is the data ready enough? Do you understand its condition, limitations and the work needed to make it suitable?
Is governance joined up? Have the relevant teams considered the proposal together, including its relationship to organisational strategy?
Have procurement and risk been considered? Is the investment proportionate to the intended outcome, with risks identified and addressed?
Are trust and human accountability in place? Is there an accountable owner, a human endpoint and a clear challenge route?
This is not intended as a one-off checklist or a reason to stop all experimentation. It is a set of questions to revisit as a project develops—particularly before scaling.
Turn the playbook into a 90-day test
The video closes with a practical challenge: identify a low-risk, high-value AI-assisted improvement that can be thought through in the next 30 days and tested within the next 90.
Start by describing a specific service pain point. Map the user journey, the data involved and the current workarounds. Bring together the people who understand the problem, then consider the simplest useful intervention.
Before choosing a tool, define the outcome you will measure. Use the readiness check to test whether the proposal is ready to proceed, and make ownership beyond the pilot explicit.
The immediate goal is not an organisation-wide AI transformation. It is a well-designed test that can produce evidence for the next decision.
Watch the Public Services AI Playbook for the full briefing. To discuss a service challenge, AI readiness or a practical pilot, contact Adam at adam@datnexa.com.
