The General Counsel Case for Owning Your AI: Legal Privilege, Discovery and Data Residency

By Micky Irons, founder and CEO of Mickai

The question no cloud AI vendor wants the GC to ask

Where does your privileged material physically live, and who else can be compelled to produce it?

For a General Counsel, that is not a procurement footnote. It is the whole case. The moment legal reasoning, draft advice, matter strategy or witness preparation passes through a public-cloud model, three things become true at once. A copy exists outside your control. A third party can be served over it. And the boundary that makes legal advice privileged in the first place has been quietly tested.

Mickai is built for the opposite posture. It is a sovereign AI operating system (SIOS): artificial intelligence that regulated businesses own and run inside their own walls, on-premise and air-gapped, with every action written to a tamper-evident, post-quantum-signed audit record we call the OAR. Built and live. This article is the GC case for that architecture, in the terms that survive a deposition.

Privilege is a property of custody, not a checkbox

Legal professional privilege protects confidential communications made for the dominant purpose of legal advice or litigation. Privilege depends on confidentiality being maintained. Confidentiality, in turn, depends on custody.

When a model provider processes your prompts, retains them for abuse-monitoring, or routes them through sub-processors across borders, you have introduced parties to a confidential communication who were never meant to be inside the circle. Even where a vendor contract promises no training on your data, the GC still has to answer a harder question in front of a court: could that data be produced by someone other than us, under legal compulsion, without our consent or even our knowledge?

With an air-gapped deployment the answer is structurally no. The model runs on hardware you control. The retrieval corpus, the RAG index over your matters and precedents, never leaves the estate. There is no sub-processor to subpoena because there is no sub-processor. Privilege survives because custody was never surrendered.

The CLOUD Act problem is a jurisdiction problem

The US CLOUD Act lets US authorities compel a US-headquartered provider to produce data in its possession or control, wherever in the world that data is stored. Choosing an EU region on a US hyperscaler does not remove that reach. The obligation follows the corporate parent, not the datacentre postcode.

For a UK or EU General Counsel that is a live conflict-of-laws exposure. You can be simultaneously bound by UK GDPR and by a foreign production order over the same material, with your provider caught in the middle and you not necessarily the one deciding the outcome. Roughly 0.85M UK businesses and around 5M across the EU sit in regimes where this exposure is not tolerable, whether through UK GDPR special-category handling, sector rules, or contractual confidentiality owed to their own clients.

Data residency on a sovereign, self-hosted system removes the conflict at the root. If the data is in your possession and control and no foreign-reachable intermediary holds a copy, there is no third party for a foreign order to compel. The GC moves from managing an exposure to not having one.

Discovery: own the record, do not inherit someone else's

Discovery and disclosure obligations run to material in your control. AI changes what that material is. Prompts, retrieved documents, model outputs and the decisions taken on them are all potentially disclosable. If those artefacts live in a vendor's logging pipeline, you have a duty to preserve and produce data you cannot fully see, in a format you did not design, held by a party you do not direct.

Mickai inverts this. Every action, every retrieval, every model decision is written to the OAR, an append-only audit record signed with ML-DSA-65 post-quantum signatures and bound to hardware-verified identity. That gives the GC three things a cloud log cannot. Completeness, because the record is generated at the point of action, not reconstructed later. Integrity, because tampering is cryptographically detectable rather than a matter of trusting a supplier's assurance. And custody, because the record is yours from the first byte.

When you own the record, discovery becomes a controlled disclosure of a defensible chain of evidence rather than a scramble to extract fragments from someone else's infrastructure. It also maps cleanly onto a GDPR DPIA, where the DPO needs to show exactly what data was processed, by which component, and on whose hardware.

Why the arbiter design matters to a lawyer

Under the hood Mickai runs 50 specialist brains under a deterministic arbiter. For legal work that determinism is the point. A single stochastic model is hard to attest to, because the same question can yield materially different reasoning on different days. A deterministic arbiter that governs which brain answers, records why, and can execute a compensating rollback if a downstream action needs unwinding, produces something a GC can actually stand behind. Legal reasoning is handled in the Astraea studio, compliance in Nomos, audit in Aletheia, each writing to the same signed record. Voice-biometric quorum can gate the most sensitive actions so that no single credential releases privileged material.

This is not AI as an unaccountable oracle. It is AI as an attestable process, which is the only kind a General Counsel can safely put near privileged and litigation-sensitive material. It is also the kind of evidence the Board and its non-executive directors can be shown when they ask how AI is governed under operational resilience and, for in-scope firms, DORA.

An ally to your existing stack, not a rip-and-replace

Mickai is an ally, not a replacement for the legal team's judgement or the tools they already trust. It sits alongside your document management, matter management and e-discovery systems as the sovereign reasoning layer that keeps sensitive work inside the perimeter. Public-cloud AI has a role for non-confidential, low-sensitivity tasks. The GC case is simply that privileged, litigation-bearing and special-category material should never be one of them. This is a dual-buyer posture: the same architecture that satisfies the General Counsel also satisfies the CISO and the DPO, because all three are asking a version of the same custody question.

As a third-party momentum signal, our founder was ranked number 4 on Crunchbase as verified in June 2026, with Mickai among the top one to two percent of companies globally. We read that as external validation of a thesis regulated buyers already understood, that ownership beats access when the material is privileged.

The estate behind the claim

Underpinning all of this is 104 filed UK patent applications spanning roughly 2,340 claims, held by Mickai LTD, inventor Micky Irons. Filed, not granted: a priority date and a prior-art position, not a marketing line. For a GC evaluating a long-term dependency, the relevant point is that the sovereign architecture, the signed audit record and the deterministic arbiter are documented inventions with a defensible provenance, not features that can be quietly deprecated.

Closing note for General Counsel

The system is built and live, and we are building to scale with a first cohort of regulated organisations for whom privilege, discovery integrity and data residency are board-level obligations rather than preferences. This is a deliberate, capacity-managed selection.

If you are the person who has to answer where the privileged material lives and who can compel it, I would welcome the conversation directly. micky@mickai.co.uk

Frequently asked questions

Does on-premise sovereign AI actually protect legal professional privilege?

Privilege depends on maintaining confidentiality, which depends on custody. With an air-gapped Mickai deployment the model and the retrieval corpus run on hardware the organisation controls, with no sub-processor holding a copy, so custody is never surrendered and privilege is not tested by a third party being in the circle.

How does this address the US CLOUD Act for a UK or EU General Counsel?

The CLOUD Act reaches data in the possession or control of a US-headquartered provider regardless of storage region, so an EU region on a US hyperscaler does not remove the exposure. A sovereign, self-hosted system with no foreign-reachable intermediary leaves no third party for a foreign production order to compel.

What makes the OAR useful for discovery and disclosure?

The OAR is an append-only audit record generated at the point of action, signed with ML-DSA-65 post-quantum signatures and bound to hardware-verified identity. It gives the GC completeness, cryptographically detectable integrity, and custody from the first byte, turning discovery into controlled disclosure of a defensible chain rather than a scramble through a vendor's logs.

Why does a deterministic arbiter matter for legal work?

A single stochastic model can produce materially different reasoning on the same question, which is hard to attest to. Mickai runs 50 specialist brains under a deterministic arbiter that records which brain answered and why and can execute a compensating rollback, producing an attestable process a General Counsel can stand behind.

Does Mickai replace existing legal technology?

No. Mickai is an ally that sits alongside document management, matter management and e-discovery systems as the sovereign reasoning layer keeping privileged and special-category material inside the perimeter, while public-cloud AI remains fine for non-confidential tasks.