We've proven value, but we're stuck operationalising it
Embedding AI into everyday workflows through people-centred process redesign, with the governance, security, data and legacy foundations that make it sustainable.
The pilot worked. Now the hard work begins.
Moving from “it works in the pilot” to “it’s how we work” is where many organisations stall. The technology has earned its place. What’s now under pressure is everything around it: workflows, roles, governance approvals and security models all designed before AI existed. The organisations that break through redesign work around AI, rather than bolting AI onto work that was never designed to use it.
Cafcass Family Court Advisors writing 80,000 letters per month
Scribe (Version 1) embeds AI into the existing ChildFirst case management system to free time for direct family support
NHS chest X-rays now AI-assisted (2.4 million scans)
Action Plan One Year On, 2026
ATRS records published by central gov departments by end-25;
Mandatory across all in-scope departments and ALBs. Responsible AI reporting is no longer optional. It is a compliance obligation.
"Even when early pilots show clear promise, integration into day-to-day operations is challenging. Senior sponsorship can dramatically accelerate the move from experimentation to scalable success."
Version 1 ‘From Playbook to Progress: Delivering AI in the Public Sector’
Understanding the challenge
You have proven the technology works. The barriers now are systemic, and they cluster in five areas:
- People. Roles, capability and confidence have not yet caught up with the new way of working.
- Process. Workflows still optimised for pre-AI patterns, with manual steps AI now makes redundant.
- Governance and risk. Assurance, ATRS reporting and the AI Playbook's ten principles need to be embedded into delivery, not bolted on as a gate.
- Data and security. Data quality, access controls, lineage and classification issues that were manageable at pilot stage become blockers at scale. Addressing these foundations properly and using AI tooling to accelerate that work where possible, pays back across every future deployment, not just this one.
- Legacy IT. Most UK Public Sector organisations operate complex estates spanning mainframes, packaged applications and cloud-native services. AI on top of legacy is its own discipline.
All five have to move together. If one is left behind, operationalisation stalls.
Three things that will move this forward this quarter:
1. Run a problem-and-user-first review on every active deployment. Confirm the user need, the workflow it changes, and whether AI is actually the right answer. Be willing to stop deployments that fail this test.
2. Embed governance into the delivery rhythm. ATRS records, responsible AI assessments and security reviews as standing items, not pre-launch gates. Aligned to the AI Playbook’s ten principles by default.
3. Stand up role-specific capability building with named manager owners. Measured by confidence and behaviour change, not training completion. The middle layer is where adoption actually happens.
AI on top of legacy is its own challenge
Most UK Public Sector organisations are not building AI into a clean estate. You are building it on top of decades of accumulated systems, with limited APIs, inconsistent data structures and fragile change windows.
Operationalising AI in that context is making honest calls about what to integrate, what to insulate, and what to modernise alongside the AI itself. The Action Plan One Year On acknowledges this directly. The AI Growth Zones and the National Data Library exist partly so departments have somewhere modern to land AI workloads. The decision is rarely all-or-nothing; it is a sequence.
REACH
At this stage: Acquire and build Capacity
Acquire / Architect
Make the new way of working visible. What does the redesigned workflow look like? Where does AI act, where does a human validate, where does a human decide?
Capability / Cultivate
Build real competence in the flow of work, not in a one-off training session. Role-based learning, peer support, protected practice time and managers who actively enable.
Designing the work, not just deploying the tool
The starting point is the business problem and the people doing the work today, not the AI. The sequence we use:
1. Define the problem. Be specific about the user (caseworker, citizen, analyst, decision-maker) and the moment in their work that changes.
2. Understand the users. Where are the friction points and the rework loops? Field work, not desktop.
3. Decide whether AI is the right answer. Sometimes it's automation. Sometimes process redesign. Sometimes better data. Principle 6 of the Playbook is explicit: use the right tool for the job.
4. Define the human-AI interaction. Where AI acts, where humans validate, where humans decide. Aligns to Principle 4 (meaningful human control).
5. Rebuild with purpose. Design the workflow so AI accelerates the people doing the work, not so it removes parts of their role without addressing what comes next.
What success looks like
Workflows redesigned around the user need
End-to-end processes rebuilt to address the actual problem, with AI powering the parts where it adds value and humans focused on judgement and exception.
Clear human-AI roles
Defined responsibilities for those who train, validate, oversee and decide alongside AI. Boundaries written down, understood and reviewed.
Governance built in, not bolted on
The ten principles, ATRS records, security assurance and data protection considered at design time and reviewed continuously.
Capability that lasts
Role-specific learning embedded in delivery. Confidence measured alongside completion rates.
A culture of visible success
Celebrate working AI systems as they deliver. Teams that share wins openly through internal communications, peer showcases and leadership recognition.
Practical actions for your teams
- Build governance into the delivery rhythm. ATRS records, responsible AI assessments, security reviews and data protection sign-offs as continuous activities, not pre-launch gates.
- Decide your legacy strategy deliberately. Per workflow, decide whether to integrate, insulate or modernise. AI rarely fixes a legacy issue; sometimes it amplifies one.
- Build capability in the flow of work. Replace generic training with role-specific learning, coaching, peer support and protected practice time. Manager involvement is the multiplier.
- Create human-digital pods. Small cross-functional teams where AI handles the routine and humans focus on judgement, creativity and the relationships that matter.
VERSION 1 IN ACTION
Cafcass: redesigning the work, not just deploying the tool
Cafcass Family Court Advisors (FCAs) write around 80,000 letters per month to children and families across 100+ templates. Each letter takes 30–40 minutes of manual editing. Version 1 worked with Cafcass on Scribe, an AI-assisted letter-drafting tool embedded directly into the existing ChildFirst case management system.
The starting point was the user, not the AI: who writes the letters, how the data flows, where the friction sits. Built on Microsoft Azure AI services with security, governance and human-in-the-loop oversight from day one — “AI providing essential time-saving assistance to the FCAs writing those letters”, as Cafcass CIO Rob Langley put it. Workflow redesigned around the user need; AI does what AI is good at; humans stay in control of the decisions that matter.
