data sovereignty

Data Sovereignty Is Now a Legal AI Deployment Requirement

The Heppner ruling and Swiss-Noxtua's €80.7M raise show data sovereignty in legal AI is now a privilege issue, not a preference. Here's the architecture that answers it.

RAGbase Legal Research TeamNovember 6, 2026 10 min read

In February 2026, a single court ruling did what three years of legal AI marketing could not: it turned data sovereignty from an IT preference into a malpractice exposure. The Heppner ruling examined whether work product generated through a third-party generative AI platform retained privilege when the underlying prompts, source documents, and draft outputs had passed through — and remained resident on — infrastructure the firm did not control. The reasoning was unsentimental: privilege and work-product protection depend on demonstrable custody and control, and a vendor's shared cloud that retains copies of client material beyond the engagement undermines exactly that.

Four months later, Swiss-Noxtua closed an €80.7 million raise to build sovereign, jurisdiction-controlled legal AI infrastructure for European firms and institutions. Two events, one signal: enterprise buyers — and now courts — are treating the physical and contractual location of AI processing as a legal variable, not a procurement footnote.

For managing partners evaluating AI vendors in 2026, this changes the diligence question. It's no longer "is the model good enough?" It's "can we prove, in a privilege dispute or a regulatory audit, exactly what left our infrastructure, where it went, and who could access it?" Most legal AI platforms were not architected to answer that question cleanly.

The Heppner Ruling: Privilege Now Turns on Architecture, Not Intent

The significance of Heppner isn't that it banned AI-assisted drafting — no court has done that, and none is likely to. The significance is narrower and more consequential: the court treated data custody as a factual predicate for privilege, not an abstraction. Where a firm couldn't show that prompts, source documents, and generated drafts stayed within a controlled, auditable boundary, the privilege claim weakened on the facts, independent of whether anyone actually misused the data.

That's a different legal standard than most firms have been operating under. Up to now, the working assumption at many AmLaw 200 firms has been: if the vendor contract says the data isn't used for model training, we're covered. Heppner suggests that contractual assurances about use don't fully answer the question of custody — regulators and opposing counsel can still probe where data sat, how long, and who had technical access, regardless of what the vendor promised not to do with it.

This is why the ruling is being cited well outside the jurisdiction where it was decided. It crystallizes a risk that litigators have been flagging informally since 2023: generative AI outputs built on client confidential information, processed through infrastructure a firm doesn't own, create a discoverable seam. Heppner gave that seam a case name.

Swiss-Noxtua's €80.7M Raise: The Market Already Priced This In

Swiss-Noxtua's raise is best read as a market signal rather than a single company's success story. Investors are backing sovereign legal AI infrastructure at scale because European financial institutions, government legal departments, and increasingly AmLaw 200 firms with EU-regulated clients are asking a question their US-only stacks can't satisfy: where, specifically, does our data reside, and under whose legal jurisdiction? 

This isn't a Europe-only concern dressed up as GDPR compliance theater. The same question is now standard in RFPs from US financial services clients, government contractors, and health systems — all sectors with regulatory regimes (GLBA, FedRAMP, HIPAA) that treat data residency and processing custody as compliance-critical, not optional. A firm that can't answer the sovereignty question loses those clients' instructions before the AI tool's output quality is ever evaluated.

The capital flowing into sovereign AI infrastructure — Noxtua's raise is one of several similar rounds across 2025–2026 — confirms that "keep it in the EU" or "keep it on our servers" is now a line item buyers will pay for, not a nice-to-have feature. For law firm leadership, the lesson isn't "go build sovereign infrastructure from scratch." It's: the deployment model matters as much as the model itself, and that's now underwritten by real capital and real case law, not vendor talking points.

What Sovereignty Actually Means in Legal AI — A Precise Definition

"Data sovereignty" gets used loosely enough that it's worth defining precisely, because the honest answer is more nuanced than "never send data to a cloud provider." Every usable generative AI system, including sovereign ones, ultimately routes some text to a language model for inference. The question that matters is what, how much, and under whose control.

LayerWhat it includesWhere it typically lives
Full document corpusContracts, filings, correspondence, discovery setsFirm infrastructure (or should be)
Vector indexes & embeddingsSearchable representations of the full corpusFirm infrastructure (or should be)
Permissions & access controlsWho can query what, matter-level ethical wallsFirm infrastructure (or should be)
Audit logsFull record of queries, retrievals, outputsFirm infrastructure (or should be)
Retrieved chunks sent for inferenceMinimal text snippets needed to answer one queryLLM provider, under firm's chosen API terms

This is the distinction that actually matters in a privilege analysis, and it's the one most shared-cloud legal AI platforms blur. If a platform ingests entire matters into a vendor-hosted environment to build its retrieval index, the firm has already lost exclusive custody of the full corpus — regardless of how careful the vendor is about model training opt-outs. If, instead, the index and full documents never leave firm infrastructure and only small retrieved fragments are sent per query, the custody question answers itself far more favorably in a discovery dispute.

The Architecture That Actually Answers Heppner

This is the core engineering decision RAGbase Legal made, and it predates Heppner by design rather than in response to it — but the ruling is the clearest validation yet that the decision was correct. The system separates two layers that most legal AI products collapse into one:

  • The retrieval and agent layer — full documents, vector indexes, matter-level permissions, connectors to DMS and practice management systems, workflow orchestration, and complete audit logs — runs on the firm's own infrastructure, whether on-premise or in a firm-controlled private cloud.
  • The inference layer — the actual language model call — receives only the minimal set of retrieved chunks necessary to answer a specific query, sent to whichever LLM provider the firm has selected, under the firm's own API and data-processing terms.

This is not "we never touch an external model," which would be both inaccurate and, frankly, a weaker product. It's the more defensible and more honest position: the full client file never leaves firm control; only narrowly scoped fragments do, and the firm chooses the terms under which that happens. For a firm answering a sovereignty questionnaire or defending a privilege claim, that means the custody chain for 95%+ of the sensitive material — the actual documents, who accessed them, what was searched, what was retrieved — stays entirely within infrastructure the firm owns and can produce records for. Only small, task-specific fragments have a documented, controllable exit point.

This matters practically in something like case search: a partner querying thousands of pages of discovery material never uploads that corpus to a third party. The index lives in-house. The model sees only the handful of passages relevant to the specific question asked — and that exchange is logged, scoped, and auditable.

Deployment Models Compared

Firms evaluating AI platforms in 2026 are effectively choosing among four architectural postures, each with different sovereignty and privilege implications:

ModelExample categoryFull corpus locationWhat reaches the modelCustody story for privilege disputes
Consumer AI assistantsGeneral-purpose chat tools used ad hoc by staffUploaded, often unmanagedEntire uploaded documents, no firm oversightWeak — no firm-level audit trail
Per-seat legal SaaSWorkflow-specific tools billed per userVendor-hosted cloudFull documents ingested for processingMixed — depends entirely on vendor's retention terms
Shared-cloud legal AIMulti-tenant platforms with broad agentic featuresVendor-hosted, multi-tenant cloudFull documents and context, vendor-managed indexesMixed to weak — custody shared with vendor by design
Private/on-premise AIRAGbase Legal's modelFirm-owned infrastructureMinimal retrieved chunks onlyStrong — firm retains documentable custody throughout

The middle two categories aren't necessarily bad products — tools in the Harvey, CoCounsel, Lexis+ Protege, and Legora category have real workflow value and sophisticated agentic features. But their architecture generally requires ingesting client documents into vendor-managed environments to power retrieval and agentic reasoning, which is precisely the custody arrangement Heppner put under a legal microscope. Claude Cowork and similar consumer-grade assistant products sit even further from a defensible custody position, since document handling there is often informal and outside any firm-level governance layer at all.

None of this requires treating every shared-cloud tool as unusable — plenty of workflows (first-draft brainstorming, non-privileged research, internal knowledge management) carry lower custody risk. The point is matching workload sensitivity to architecture, deliberately, rather than defaulting to whichever tool is most convenient.

What This Means for Procurement at AmLaw 200 Firms

General counsel and innovation leads evaluating legal AI vendors post-Heppner should be asking three questions that most RFPs still don't cover:

  • Where does the full document corpus sit during indexing and retrieval — vendor cloud, or firm-controlled infrastructure? If it's the former, what jurisdiction governs that infrastructure, and does it match the matter's governing law and the client's regulatory requirements?
  • What exactly is sent to the underlying LLM per query — the full document, a summarized version, or a minimal retrieved fragment? Vendors should be able to answer this precisely, not with a general data-use policy.
  • Can the firm produce a complete, independent audit log of every query, retrieval, and output tied to a matter, without relying on the vendor to generate or preserve that record? In a discovery dispute, the firm needing to ask its vendor for logs it doesn't already control is itself a custody problem.

These questions matter most for the highest-stakes workloads — privileged litigation strategy, pre-merger diligence, regulatory investigations, anything where a privilege log might eventually be tested in court. They matter less for lower-sensitivity tasks, which is exactly why a one-size-fits-all deployment decision is the wrong instinct. The right posture is tiered: sovereignty-critical matters on firm-controlled infrastructure, lower-risk workflows on whichever tool is most productive. Our AI for law firms guide walks through how to map specific practice workflows to the deployment model each one actually requires.

What Comes Next

Expect two things to accelerate through 2026. First, more rulings like Heppner — probably in multiple jurisdictions, since the underlying custody logic isn't specific to one legal system. Courts asking "who controlled this data, and can you prove it" is a generalizable line of inquiry, not a one-off. Second, expect vendor marketing to get noisier about "sovereignty" and "privacy" as a result, which makes architectural specificity more valuable, not less. The firms that will navigate this well are the ones that can answer custody questions with infrastructure diagrams and audit logs, not policy documents.

The capital behind Swiss-Noxtua and the reasoning in Heppner are pointing at the same underlying shift: AI deployment architecture has become a matter of legal risk management, evaluated with the same rigor firms apply to conflict checks and data breach response plans. That's a higher bar than most legal AI procurement processes were built for eighteen months ago.


If your firm is reassessing AI vendor contracts in light of Heppner, the diligence worth doing now is architectural: ask every vendor, in writing, exactly what data leaves your infrastructure, in what form, and under what jurisdiction — then weigh that against the workflow value each tool provides. Private AI deployment is built around giving a documentable answer to that question for exactly the matters where it matters most.

Frequently Asked Questions

Does using a cloud-based generative AI tool automatically waive attorney-client privilege?
Not automatically, but the Heppner ruling (February 2026) confirmed that courts will scrutinize who retains custody of prompts, source documents, and outputs when assessing whether privilege and work-product protections hold. If a vendor's infrastructure stores and can access that material beyond the engagement, firms may struggle to demonstrate the exclusive control privilege doctrines typically require.
What does 'data sovereignty' mean in the context of legal AI, specifically?
In legal AI, data sovereignty means the firm — not a third-party SaaS vendor — controls where full documents, indexes, permissions, and logs are stored and processed, including jurisdiction of data residency. It does not necessarily mean avoiding large language model providers entirely; it means minimizing and governing exactly what leaves firm infrastructure.
How is RAGbase Legal's architecture different from sending documents to a cloud AI assistant?
RAGbase Legal keeps the full document corpus, vector indexes, permissions, and audit logs on the firm's own infrastructure. Only the minimal retrieved text chunks needed to answer a specific query are sent to the selected LLM provider under the firm's own API terms — a materially smaller and more controllable data exposure than uploading entire matters to a shared-cloud platform.

Related Articles

R
RAGbase Legal Research Team
Research

RAGbase builds private AI systems for law firms: deployed on the firm's own infrastructure, zero data retention, full ownership.

See How RAGbase Works on Your Data

30-minute call. We scope your use case and show the system live.

We use audience and marketing cookies (Google Analytics, LinkedIn). No tracker loads without your consent. Learn more