Law firms are discovering that zero data retention agreements don't stop frontier AI labs from learning their secrets, prompting a strategic shift toward building proprietary AI infrastructure that keeps competitive advantages locked inside the firm. When Anthropic launched Claude Cowork in January 2026, the market reacted with shock. Within three days, roughly $285 billion evaporated from software and professional services stocks in what analysts called the "SaaSpocalypse." Thomson Reuters fell about 16 percent, RELX dropped about 14 percent, and LegalZoom plummeted about 20 percent. The reason was clear: law firms' own clients were shifting work directly to AI competitors, cutting into firm revenues.

How Are Frontier Labs Learning Your Secrets Without Your Permission?

Most law firms believed that zero data retention (ZDR) agreements, which prevent vendors from keeping copies of documents and training on them, would protect their proprietary knowledge. But security experts and industry leaders now argue this protection is incomplete. Three distinct mechanisms allow AI labs to extract competitive intelligence without ever touching a single client document.

The first mechanism is straightforward: data extraction through prompts and outputs. ZDR prevents this directly. But the other two are far more subtle and harder to defend against. The second is capability absorption. Even without reading a single contract, a frontier lab can watch which workflows a firm spends the most time and money on. Usage data, feature adoption, and integration patterns reveal where the value sits, and none of that is covered by a retention clause because it was never classified as "data" in the traditional sense. Once a lab identifies where the value sits, it executes the third mechanism: vertical entry. The lab ships that workflow itself and becomes a direct competitor.

This is precisely what happened in legal. The pattern accelerated throughout 2026. Microsoft launched its own Legal Agent for Word in April. Claude for Legal followed in May with a dozen practice-area plugins and more than 20 legaltech connectors built in. OpenAI opened a legal vertical and hired Jason Boehmig, founder and former CEO of Ironclad, to lead it. Anthropic saw everything it needed from its own usage data to justify building Claude for Legal, without ever needing to train on a single customer document upload.

Why Training Your Own Model Isn't Enough

Some law firms have responded by training their own models, hoping to own their intelligence instead of renting it. But this approach creates new problems that many firms haven't anticipated. If a vendor trains and hosts the model, the firm's workflows end up locked inside an artifact it cannot export. A document corpus is portable; a fine-tuned model generally is not. Switching costs shift from "re-integrate" to "start over," which is also known as lock-in.

There's also the question of what ends up in the model. Training it on how a firm handles a specific client means that privileged material, possibly the client's own confidential information, is now baked into the model permanently. Additionally, the benchmark used to prove the model works is usually built by the same vendor selling the model, which is not an independent test and makes it difficult to build internal expertise.

Kirkland & Ellis recognized these gaps and decided to build a $500 million solution in-house with Palantir. Rather than a model fine-tuned and hosted by an outside vendor, it's a platform backed by an AI team of more than 180 engineers. But getting there doesn't require that scale of investment for most firms.

How to Build AI Sovereignty Without Breaking the Bank

• Create a Structured Ontology: Build a structured representation of your firm's world, including funds, limited partners, side letters, obligations, counterparties, negotiation rounds, clauses, definitions, and the relationships between them. This allows lawyers, AI agents, and vendors to work using a shared language of quality data instead of uploading raw documents to external labs.
• Make the Model Swappable: An ontology layer functions like an API for your data, not a firehose into someone else's model. Your firm doesn't have to hand a lab your entire contract history to get a useful answer. Only the narrow slice of data needed for a given task ever reaches a model, creating a much smaller surface area than shipping full documents and prompts to external vendors.
• Retain Ownership of New Insights: When a firm produces a new negotiated outcome, it gets captured straight back into its own system and ontology. That means it sits within the firm's ownership rather than the model's, so the firm can immediately enhance the value of its own solutions. This asset becomes something the firm definitively owns, in a data format that is open and exportable.
• Start Narrow and Expand: Firms that achieve impact fastest start with a narrow domain where institutional knowledge already compounds, such as precedent, negotiated positions, clause libraries, or playbooks. Once that foundation exists, the ontology expands alongside the firm's work rather than requiring everything to be modeled up front.

The key insight is that an ontology is not the same as a document management system. Documents are the source material. The ontology is the structured representation of the concepts inside them and the relationships between them. A document management system stores files; an ontology models the business.

What Does AI Sovereignty Actually Mean for Enterprises?

NVIDIA CEO Jensen Huang first used the term "AI sovereignty" in 2024 to describe countries building and controlling their own AI infrastructure. Palantir's Alex Karp has since applied the same idea to enterprises facing a critical threat: when you outsource your AI to a vendor, you're not just paying with money, you're revealing your most valuable competitive secrets. The concept reflects a fundamental tension in the enterprise AI market. As frontier labs like Anthropic and OpenAI expand into vertical markets, they gain not just revenue but also competitive intelligence about how their customers work.

"You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it," said Satya Nadella, CEO of Microsoft.

Satya Nadella, CEO at Microsoft

The firms that thrive will be those that understand when, how, and where to build their own infrastructure versus when to partner with external vendors. For law firms, that investment increasingly means building the ontologies and AI infrastructure that keep their competitive edge proprietary and portable, ensuring they remain masters of their own data and insights rather than becoming unwitting suppliers of intelligence to their competitors.