18th November 2026:
Most of the public sector leaders I speak to are not asking whether AI works. They have already seen it work. A pilot summarised case files in seconds, or drafted replies to routine queries, or triaged a backlog that had been growing for months. The team liked it. The demo went well.
And then it stopped.
Six months later, the pilot is still a pilot that never made it through the steps between a promising test and a service the organisation will stand over.
Getting from a successful pilot to a live service is the challenge I hear about most, and it's very rarely a technology problem. Minister James Lawless made the same point in October: "the transition from AI experimentation to trusted, real-world deployment will be critical".
The OECD lists the usual reasons public sector pilots stall: skills shortages, problems accessing and sharing data, cost, outdated regulation and legacy IT. That matches what I hear in public sector bodies, where those pressures usually show up as four practical walls.
Governance and risk review. A pilot built to prove value is rarely built to pass review. When the risk committee or the Data Protection Officer asks who can see the data, how decisions are checked and what happens if the model gets it wrong, the answer is often "we haven't worked that out yet".
No safe place to run it. A pilot usually runs in a test environment, with sample data and a small group of users - all kept apart from the organisation’s existing and legacy systems. A live service is different. It has to connect to those systems, handle real personal data, control who can see what, keep a record of what it does, and have someone watching for problems every day. Most organisations don’t yet have a platform that provides this, so each pilot has to build it from scratch, and that's usually where and why it stops.
Nobody owns it once it's live. Once a pilot ends, the people who built it usually move on to the next project, and staff with the skills to run AI day to day are hard to find. IT is reluctant to take on a system it didn’t build and may not fully understand. The service area wants the benefits but not the accountability if something goes wrong. So when it's time to go-live, there's no one in a position to say “this is mine, and I'll stand over it”. Without that person, the pilot waits.
The AI Act is getting closer. High-risk obligations apply from 2 December 2027. Organisations are right to be careful, but careful without a plan becomes stuck. More on that below.
From 2 December 2027, the EU AI Act’s high-risk rules apply to any AI that helps decide whether someone is eligible for public benefits or services. The same rules cover AI used in recruitment, education and several other areas where public bodies are already running pilots.
The detail that catches people out is this: a pilot is not yet in service.
Many organisations assume a pilot that started before the deadline will be treated as an existing system. It won't. A pilot is not yet in service, so if it goes live after 2 December 2027, it must meet the rules from its first day. The only extra time is for AI a public body already has in live use before that date: those systems must be brought into line by 2 August 2030. But if one of them is significantly changed before then, for example redesigned to do something new, the rules apply from the moment of that change. So holding a pilot back until after the deadline does not get it out of the rules. It only leaves less time to get it ready, and every month it waits is a month the service goes without the benefit.
Legally, these obligations sit with the public body using the AI. In practice, they land on the desks of specific people:
Under Ireland's model, AI is supervised by existing regulators, including some that public bodies already deal with, such as the Data Protection Commission. Anyone can complain to them about a suspected breach, including a member of the public affected by an AI-supported decision.
I've seen plenty of organisations reach for a policy document at this point. It won't be enough. What a regulator or auditor will want to see is a working system - monitoring that runs every day, logs that are kept, oversight built into each agent, and evidence you can hand over when asked. Anyone who starts on this in late 2027 will be racing against the clock. Start now, and you can have your first agents live and learning from real use well before the deadline arrives.
When a pilot stalls, the instinct is often to go back to the drawing board: a new use case, a new supplier, a new pilot. That throws away the most valuable thing the organisation has, which is evidence that AI can help with a real problem in its own service.
Most stalled pilots don't need replacing. They need assessing. What can be reused? What needs to change before it can go live safely? Where does a person need to review or approve what the AI produces? Those questions have answers, and working through them is far quicker than starting over.
When I look at the organisations that have moved pilots into live service, they have all generally paid attention to the same 4 key things;.
They sorted out where the AI would run before they needed it. They set up one secure, monitored place to run AI before the first agent went live. It keeps a record of what each agent does and alerts their own staff if something starts to go wrong. It is work you only want to do once, and after that the second and third agents can use the same setup instead of starting from scratch.
They built in a human check from the start. No platform can fix an agent that was designed badly. If an agent approves or rejects applications with nobody reviewing the decision, it is non-compliant wherever it runs. The organisations that get through decide early where a person needs to review or approve what the AI produces, and build those checks in.
They gave the reviewers what they needed. In my experience, risk committees and DPOs are not trying to block anything. They need to see what the agent does, what data it uses, who checks its work and how problems will be caught. When they get that in a form they can sign, the conversation moves quickly.
They brought their staff with them. A tool nobody trusts or uses doesn't deliver anything. People need to know what the AI is doing, where its limits are and when to step in. Under the AI Act, supporting staff to build that understanding has been a requirement since February 2025.
This is the same care that any public body already takes with any critical service - and applied to AI. The Guidelines for the Responsible Use of AI in the Public Service sets out the principles. The hard part is showing that you actually meet them for each agent.
If you have a pilot that worked but hasn't gone live, the most useful thing you can do is find out exactly what stands between it and live service. In most cases the gap is smaller than it looks, and a standard-risk agent can be live within weeks, not months.
That is why we built the Managed AI Framework: to give public bodies a governed route from pilot to live service, with the platform, the design discipline and the ongoing oversight in one place.
If you'd like to talk through where your pilot is stuck, I'm always happy to have that conversation. tclancy@openskydata.com
Tom Clancy is Head of Public Sector Business Development at OpenSky.