The Agentic State: A Platform for AI in the Public Sector
On October 28, the Budget cut departmental administration budgets by 20% by 2029/30. It named AI and digital transformation as part of how departments will get there. It also put £35m into GovAI to give civil servants frontier tools and set up a £50m Civil Service AI transformation fund (Tussell, 2026). The savings are now in the numbers. Whether your department can deliver them depends on how it uses AI from here.Two days before the Budget, the Alternative People Survey 2026 from Re:State, Public First and Civil Service World showed where things stand. 82% of civil servants now use AI at work, up from 30% a year earlier. Most of that use is drafting, at 74%. Only 16% use AI to automate a process (Re:State, 2026). That last figure is the one to watch. A faster draft helps one person. A redesigned process changes what a service costs to run.
Why pilots stall
The UK public sector has laid good foundations. The Defence AI Centre is up and running. The Incubator for AI has made tools such as Minute and Extract open source. The cross government Microsoft 365 Copilot experiment published its findings in June 2025. Government knows how to test AI quickly and in the open.
The trouble starts when a pilot tries to become a service. Each pilot builds its own link to the data, its own version of the rules and its own screen. When it ends, the next pilot starts again from nothing. Off the shelf tools make this worse. They arrive built for someone else’s workflow and sit beside your systems instead of inside them.
The result is AI that cannot see your case notes, your policies or your records. So it answers in general terms when your people need specifics. Your pilots do not scale because nothing underneath them was built to scale.
The missing layer is a platform
Elsewhen calls the answer the Agentic State. It is an institutional model where agents become a core part of how a department is built, instead of assistants added on the side. The models themselves are now widely available. The advantage sits in the layer between the models and the service. Take health and social care as the example.Business context. This is a structured model of how health and social care actually works. It holds the legislation, the eligibility rules, the care pathways and the data in your case management and finance systems. Every agent starts from the same understanding of how the service runs.
Headless agents. These agents have no screen of their own. They run a workflow, or a chain of workflows, wherever they are called from. One gathers the evidence for a referral. Another checks it against the thresholds and prepares a summary. Caseworkers spend their day on the decisions only a caseworker can make.
Orchestrator. This is the coordination layer. Worker agents retrieve policy and trace where data came from. A supervisor agent reviews the work for compliance. A person signs off. Every action an agent takes is logged and attributable, so anyone can review it later.
Skills library. The rules live in plain language files that the people who know them can write and change. An eligibility skill for Care Act 2014 assessments can state that it supports the assessment and that a qualified social worker makes the final decision. When guidance changes, the team updates the skill. The people who know the rules are the ones who write them down.Generative UI. Interfaces are created on demand for the case in front of you. There is no new system to learn.
The department owns this core. New models plug straight in, so you are never tied to one vendor. Each new agent ships faster and costs less than the last, because the context, the rules and the controls already exist.
Humans stay where the judgement is
In public services, governance cannot be added at the end. An auditor, a select committee or an ombudsman will want to know who decided what and on what basis. A platform answers that by design. Agents prepare and check the work. People make the calls that carry legal or human weight, and the record shows both.
The Department for Work and Pensions shows what this looks like at scale. A machine learning model in the Child Maintenance Service predicts which payment arrangements are likely to break down, so caseworkers can step in early. DWP estimates it keeps an extra 715 cases a month out of arrears and saves the time of 49 full time staff (National Audit Office, 2026). The AI changes the process. The caseworker still makes the decision.That is the shape of saving the Budget is counting on. It comes from redesigned work, measured against a real baseline.
Four weeks to evidence you can act on
A platform is built once, but the first use case has to earn it. Elsewhen runs that first step as a four week sprint.In week one, you map the context, the data and the risks, and agree a scoped use case with a human baseline. In week two, you build a working prototype behind human gates and put it in front of a real user. In week three, you test it with users against that baseline, so you measure performance instead of collecting anecdotes. In week four, you turn the findings into a costed route to beta.At the end you have a clear decision to proceed, pivot or stop, backed by evidence from real users. If the answer is proceed, the context and controls you built become the start of your platform.Elsewhen is a Crown Commercial Service supplier on the Artificial Intelligence DPS, G-Cloud 15 and Digital Outcomes and Specialists 7, and holds Cyber Essentials Plus. Watch the full session above to see the health and social care case in detail.
