AMT in the Public Sector: A Sovereign Agentic Workforce Behind the Government Firewall

The problem the public sector cannot outsource

Government back offices run on caseloads, benefits assessments, procurement files, FOI requests, grant administration, and licensing. The work is high-volume, rules-bound, and saturated with personal and special-category data. It is exactly the kind of work agentic AI was built to carry, and exactly the kind of work that most agentic AI is legally forbidden to touch.

The reason is jurisdiction. A hosted agent runs on infrastructure owned by a US hyperscaler. Under the US CLOUD Act, that data sits within reach of foreign legal process regardless of where the server physically sits. For a UK local authority handling UK GDPR special-category records, an NHS body bound by the Data Security and Protection Toolkit, or a department holding material under the NIS Regulations, that exposure is not a risk to be managed. It is a line that cannot be crossed. The practical answer so far has been to do nothing, or to do it by hand.

Mickai removes the conflict at its root. It is a sovereign AI operating system: AI that a public body owns and runs inside its own walls, on-prem and air-gapped, with every action written to a tamper-evident, post-quantum-signed audit record. Within that operating system, AMT is the agentic workforce. The data never leaves the building. There is no hosted endpoint, no shared tenancy, and no foreign jurisdiction anywhere in the path.

What an agentic workforce does behind the firewall

AMT is not a chatbot bolted onto a portal. It is a coordinated set of agents that pick up structured and unstructured work, reason over it against published policy, take action across the systems a department already runs, and hand off to a named officer at the points where human judgement and human accountability are required.

In a benefits or grants context that means intake triage, eligibility checks against rules, document verification, exception flagging, and a drafted decision with its full reasoning attached. In procurement it means supplier due-diligence packs, compliance cross-checks, and contract-file assembly. In FOI and casework it means classification, redaction proposals, and routing. The work that used to queue for weeks moves at machine speed, and the officer spends their time on the cases that actually need a person.

The difference from a hosted assistant is that none of this is advisory text the officer then has to re-key. The agents operate inside the actual line-of-business systems, under the body's own access controls, with every read and every write recorded. The department is not sending its caseload to a model. It is running the model on its caseload, in its own building.

Why the audit record is the point, not a feature

Public-sector decisions are challengeable. A benefits refusal can go to tribunal. A procurement award can be judicially reviewed. A data-handling step can be examined by the regulator. An AI that cannot show its working is not usable in that world, however capable it is.

Every action AMT takes inside Mickai is written to the Operational Audit Record, the OAR: tamper-evident, post-quantum-signed, and append-only. That means the body can show exactly which policy version an agent applied, what data it read, what it concluded, and where a human stepped in. This is the artefact that turns automation from a liability into evidence. It is what lets a public body adopt agentic AI under PRA SS2/21-grade and EU AI Act high-risk expectations rather than in spite of them. Accountability is not retrofitted. It is the substrate.

Sovereignty as a procurement requirement, not a preference

The market is moving the same way the regulation is. Roughly 850,000 UK businesses, about 15 percent, and around 5 million across the EU are legally barred from running this class of workload in the public cloud. Public bodies sit at the strict end of that line. The sovereign AI market is sized at around USD 40 billion in 2025 on a path to roughly USD 148 billion by 2032, and the public sector is one of the clearest reasons why.

Mickai is built for that requirement rather than adapted to it. The capability ships as Greek-named Studios, each one a domain workforce: Nomos for compliance, Astraea for legal, Aletheia for audit, Nemesis for fraud and AML, Pythia for business intelligence, alongside underwriting, forecasting, clinical, finance, and customer-service Studios. A public body does not buy a platform and then hire integrators to make it lawful. It deploys a workforce that is already shaped to the regulated work and already sovereign by construction.

That posture is underwritten by IP. Mickai LTD holds 104 filed UK patent applications spanning roughly 2,340 claims, inventor Micky Irons. Filed, not granted, which gives a priority date and a prior-art position over the sovereign-AI architecture itself. The same estate maps to 196 companies and 311 patent-company pairs as potential licensees, including Microsoft, AWS, NVIDIA, Google, Adobe, and IBM, a potential-licensee sizing that shows where the architecture sits relative to the field.

Built, live, and building to scale

This is not a roadmap. The operating system is built and live, the Studios are running, and the OAR is the spine that records every action. As a third-party momentum signal, in June 2026 Micky Irons was independently ranked number four on Crunchbase, with the Mickai company profile in the top one to two percent globally. The company is UK-based with Birmingham manufacturing secured, building to scale.

The strategic shape is straightforward. There is a regulated buyer who cannot use hosted agents and now has a sovereign workforce that lets them. There is a second buyer in the IP itself. A Y5 revenue path to billions at high gross margin rests on the dual-buyer thesis and the patent estate beneath it. This is the kind of category a hyperscaler would rather own than compete with, and Mickai is positioned to lead it. We are an ally to the broader AI field, not a rival to it. We make the work that hosted models cannot legally do possible, inside the walls where it has to stay.

Contact: micky@mickai.co.uk

By Micky Irons, founder and CEO of Mickai.

Frequently asked questions

Why can't public bodies just use hosted AI agents?

Hosted agents run on infrastructure owned by US hyperscalers, which brings the data within reach of the US CLOUD Act and foreign legal process regardless of where the server physically sits. For bodies handling UK GDPR special-category data, NHS records under the DSP Toolkit, or material under the NIS Regulations, that exposure is disqualifying. Mickai runs entirely on-prem and air-gapped, so the data never leaves the building and no foreign jurisdiction sits in the path.

What is AMT inside Mickai?

AMT is the agentic workforce that runs inside the Mickai sovereign AI operating system. It is a coordinated set of agents that pick up structured and unstructured back-office work, reason over it against published policy, act inside the department's own line-of-business systems under its own access controls, and hand off to a named officer where human judgement and accountability are required.

How does Mickai satisfy auditability and challenge requirements?

Every action AMT takes is written to the Operational Audit Record (OAR): tamper-evident, post-quantum-signed, and append-only. The body can show which policy version an agent applied, what data it read, what it concluded, and where a human intervened. That record is what makes outputs defensible at tribunal, judicial review, or regulator examination, in line with PRA SS2/21-grade and EU AI Act high-risk expectations.

Is Mickai built or still in development?

It is built and live. The operating system runs on-prem and air-gapped, the Greek-named Studios are running, and the OAR records every action. The company is UK-based with Birmingham manufacturing secured and is building to scale.

What underpins Mickai's IP position?

Mickai LTD holds 104 filed UK patent applications spanning roughly 2,340 claims, inventor Micky Irons. They are filed rather than granted, which establishes a priority date and a prior-art position over the sovereign-AI architecture. The estate maps to 196 companies and 311 patent-company pairs as potential licensees, including Microsoft, AWS, NVIDIA, Google, Adobe, and IBM.