Analyze 10,000 Legal Documents Without Breaking Privilege
Two weeks into a major discovery review, a litigation partner received a message from the client's general counsel: "The board needs written confirmation that no privileged documents were sent to any third-party AI service during this matter." The partner had used a cloud-based AI tool for initial privilege screening. The next 48 hours were spent reading vendor terms of service instead of reviewing documents.
It's already happening. Firms building their answer before the question arrives are the ones whose next 48 hours go differently.
Any document you upload to a cloud AI tool, you've handed to a third party. Courts are still deciding whether that constitutes privilege waiver — and it's not a comfortable position to be in during active litigation.
Why Cloud AI Creates Structural Privilege Risk
Privilege waiver isn't about intent. Courts don't ask whether you meant to waive attorney-client privilege — only whether voluntary disclosure to a third party occurred. Uploading documents to a cloud AI service is a voluntary act of disclosure. Your intent was efficiency. What happened was disclosure.
Legal ethics boards in the US, UK, and EU have all issued guidance flagging this since 2023. A 2024 American Bar Association survey found that 42% of lawyers reported using AI tools on client matters; only 18% had received formal guidance from their firm on privilege implications. The AI is already in the building — it's just not the AI the firm chose to deploy, supervised the way the firm intended, or processing documents the way partner letters contemplate.
Bar associations in New York, California, and Florida all issued AI ethics opinions in 2023-2024 specifically flagging the privilege risk from third-party AI tools. The profession is running ahead of its own guidance.
Vendors respond to this concern with security certifications. Certifications address whether documents are encrypted in transit and at rest. They don't address whether documents left your perimeter — they did — or whether your jurisdiction's ethics rules permit that disclosure on privileged material. Those are separate questions with separate answers.
How Sovereign Document Review Works
AI document review that doesn't create privilege risk runs on your own infrastructure. Documents load into your servers. AI models process them there. Determinations are generated on your equipment. Nothing moves outside your perimeter.
A litigation boutique in Luxembourg processed privilege review on 12,000 documents for a cross-border regulatory investigation using Leeloo. Every document remained on the firm's servers throughout. The review completed in 14 hours. The same work by traditional methods would have required four associates working full-time for three weeks.
Each document received a classification across four categories: privileged (recommend withhold), potentially privileged (flag for attorney review), waiver risk (privilege may have been broken by earlier disclosure), and producible (no privilege indicator). Each classification included the textual basis for the determination — machine-readable and producible as a privilege log in the format required by the relevant jurisdiction.
Every document, every determination, and every model decision in the sequence that produced it went into a processing log. When the general counsel asked for a technical description of the review process, the answer was accurate and complete before the email finished loading.
The Technology Behind It
Leeloo deploys for legal document review at SL2 — data sovereign level — meaning the law firm's infrastructure is the only compute involved. Open-weight language models fine-tuned on common law precedent run on the firm's servers. The Vault holds the document collection, indexed for semantic search — the AI can find relevant passages across 12,000 files without sending a single document to an external system. The Recorder captures every classification decision with its textual basis.
Attorney supervision is built into the workflow. Borderline determinations flag for human review rather than forcing an automated call on close cases. The AI handles the 80% of documents where privilege status is clear — and focuses attorney attention on the 20% where judgment is genuinely required.
Economically, this reallocation is significant. Relativity's 2024 survey found document review accounts for 70% of discovery costs in complex litigation. Firms using AI complete matters 60% faster than those using manual review. The economic gap between firms that can offer AI-assisted pricing and those that cannot is now visible in pitch outcomes — clients specifically ask how AI will be used and what the data handling policy is.
Law firms that answer both questions from their own infrastructure win those pitches.
What the Economics Actually Look Like
At current Leeloo licensing — €30,000-€80,000 per month — the math on a single major discovery matter is clear. A review that previously required 400 associate hours at €200-€300 per hour runs at infrastructure cost instead. Two or three matters justify the implementation investment of €300,000-€2 million.
Second-order effects are more significant than the cost comparison alone. When document review no longer scales linearly with associate headcount, the firm can take large discovery matters it previously had to decline because the economics didn't work. A partner who could previously run one complex discovery at a time can run three. Cases that required staffing up — and absorbing the risk that the matter doesn't pay what was expected — become matters the firm can commit to with fixed cost certainty.
Third-order: associates who spent months on mechanical privilege review shift to analysis and strategy work. The judgment they were hired for gets applied to the determinations that require it, rather than the ones that don't.
Firms calculating this correctly are already adjusting their staffing models and pricing structures.
Answering the Board Question in Advance
Stress test: opposing counsel files a motion to compel in a major matter. The motion includes evidence that privileged documents were processed by a cloud AI service. You have 48 hours to file a technically accurate opposition explaining why that processing doesn't constitute waiver. That opposition is going to be filed in the first significant AI privilege case, and it will arrive at a firm without warning.
On sovereign AI infrastructure, the processing log exists, the model decisions are documented, and the affidavit writes itself: every document was processed on the firm's servers, no document touched an external service, and here is the technical record showing exactly what happened to each file.
Using cloud AI produces a different answer — one that depends on vendor contract terms, what the security certification actually covers versus what it implies, and what a federal judge finds persuasive when the issue hasn't been ruled on yet.
The General Counsel Conversation
Sovereign AI doesn't eliminate all privilege risk. Attorneys still supervise, borderline determinations still require judgment, and courts may disagree with individual calls. What sovereign AI removes is the one question that derails active litigation: "Did documents go somewhere they shouldn't have?" Answering it in advance, with documentation, is the entire point.
Managing partners who have been carrying this question since AI document review became possible now have a complete technical answer. The board asks whether privileged documents were sent to a third-party service. The answer is ready, accurate, and accompanied by the processing log: no document was processed on any server except the firm's own. The audit trail confirms it.
That's not a compliance checkbox. It's the infrastructure that makes large AI-assisted discovery defensible — and the foundation for the pricing model, the staffing model, and the client relationships that come after it.
Eight to twelve weeks to deploy. The processing log is ready for the first matter.