Legal AI that never leaves your control.
Request accessLaw firms have always traded on what they know. Palladium puts the best AI available to work on that knowledge, owned by and under your firm's control.
Lawyers are using unapproved AI tools to move faster. With no oversight, risk goes up, and partners spend more time reviewing junior work.
Palladium is designed to run inside your firm, on your premises or in your firm's own cloud, on a model your firm owns. Your lawyers can work on real client files without their details ever leaving your walls, and bring in frontier models when a question needs more. Most AI vendors offer zero data retention (ZDR): they delete what you send. Palladium is built on Zero Intelligence Retention™ (ZIR™): nothing your lawyers do, including their reasoning, is kept outside your firm.
Your firm's data was never the valuable part. The value is in the reasoning: how each question was worked through, and how your lawyers checked it, corrected it and made the call. That is the art of being a lawyer, and it is the way your firm thinks. With Palladium, every matter your lawyers review becomes firm knowledge, and your firm's own model gets sharper for the next one. The flywheel starts in Phase 2, from Q3 2027.
When a question needs more, the Adversarial Frontier Council sends it to Anthropic, OpenAI, Google and xAI at the same time, with client details removed. A final review compares their answers and gives your lawyer one answer, checked against all four, with a Defensibility Score.
✓Secure
One answer for Lakeland Industries, cross-checked across all four models.
Every matter starts on a secure internal model. When a question needs more, lawyers can bring in a frontier model from the same menu.
Ask about this document, or ask for a change — every edit arrives as a tracked change for you to accept or reject.
Before anything goes to an outside model, names, parties and figures are replaced with placeholders, and anything that could be inferred from what is left is removed. The model gets the question without your client's details.
Edits from the assistant arrive as tracked changes for a lawyer to accept or reject. Each decision is also written to the matter's audit trail, which stays with the firm.
Replaced "Published Deal" with "Publishing Deal"
Most legal AI runs on the vendor's infrastructure. Your documents travel to their servers, their model does the work, and whatever the system learns about how your firm practises law stays with them.
The protection you get is a contract and a certification. That's a promise about behaviour, not a property of the architecture, and it leaves your data reachable by a breach at the vendor or a subpoena served on them. Courts have already compelled AI providers to preserve and produce user conversations in litigation.
Palladium runs inside your firm, on your premises or in your firm's own cloud, on keys you hold. There's no Palladium server holding your matters. The model, the redaction map for every matter, and every reasoning trace it builds are assets of the firm. Nothing leaves your perimeter unless a lawyer chooses to send a redacted request to an outside model, and when they do, the request carries no client identity.
The other difference is what accumulates. Every prompt into a rented tool improves the vendor's product for the next firm that signs. Every matter worked in Palladium improves a model only your firm can use.
No. Before anything leaves the firm, an internal model running on your infrastructure reads the request and replaces client names, counterparties, individuals, addresses, account numbers, dates of birth, and other details that could identify a real person or company with consistent placeholders.
Consistent matters: if Apex Data Solutions appears as “Apex,” “the Company,” and “ADS” across a matter, all three resolve to the same placeholder so the outside model can still follow the argument.
A second, separate check then reads the outgoing request as an adversary would and confirms it can't be reconstructed from what remains, including from context that isn't a name but points to one (a distinctive deal size, a named building, a rare job title). Only then does the request leave.
The key that links each placeholder back to its real value lives only inside your firm. The outside model works on the structure and substance of the problem without knowing whose problem it is. When the answer returns, placeholders are restored inside your perimeter, and your lawyer sees real names in a document the outside model never saw. The internal model, which does most of the work, sees your documents in full and never leaves.
Your firm, in three layers. The first is the redaction map: which entities appear in each matter, how they're referenced, and what your lawyers decided about each one.
The second is the working record: which clauses your lawyers accept, what they flag, how they redline, what they send back and why. The third is the model itself, which is deployed in your environment and, as it's tuned on your work over time, becomes a representation of how your firm practises that no one else holds.
All three live on your infrastructure. None of it trains a model outside your firm, and none of it is visible to us or to any other Palladium customer. If you stop working with us, the deployment, the data, and the tuned model stay with you and keep running. The firm-owned model arrives in Phase 2, from Q3 2027.
Your firm's precedent library took decades to build. This is the same asset in a new form, and it should compound for you rather than for a vendor.
It waits for a lawyer. Every candidate for redaction carries a confidence score. High-confidence matches are applied automatically and shown in the document so a lawyer can see what changed.
Anything below the threshold is held in a review queue, with the surrounding text and the engine's reasoning, and the matter can't be sent to an outside model until each held item has been confirmed, corrected, or dismissed by a person. A held item might be an ambiguous name, a person who shares a surname with a party, or a term the engine can't tell is a company or a common noun.
Once a lawyer decides, that decision is written into the matter's redaction map, so the same entity is handled the same way across every document in the matter without being reviewed again. Correcting the engine is the same motion as reviewing its work, and each correction makes the next document faster. Every decision is logged with who made it and when, and every decision can be reversed.
A secure internal model, by default. It runs on your infrastructure, sees your documents in full, and handles most day-to-day drafting, review, summarisation, and search without anything leaving the firm. For most matters, that's the whole workflow.
When a lawyer needs a second opinion on a near-final document, they can send a redacted request to Claude, GPT, Gemini, or Grok, or put the same question to all four at once through the Adversarial Frontier Council. The Council doesn't average the answers. It returns where the models agree as confidence and where they disagree as exceptions, so a lawyer's attention goes to the specific clauses or claims the models contest rather than to the whole output.
Council requests are the expensive step and the only step that leaves the perimeter, so the product is built to make them rare and deliberate. Your firm decides which outside models are permitted, for which practice groups, and for which clients, and those rules are enforced before a request is assembled.
Tell us a little about your practice and we'll show you Palladium on documents like yours.