What's Happening?
Legal AI vendors, including Harvey and Thomson Reuters, are increasingly developing their own artificial intelligence models, such as Harvey's Tenet and Thomson Reuters' Thomson. This strategic shift aims to reduce their reliance on external frontier
model providers like OpenAI and Anthropic. By building vertically integrated AI stacks, these companies seek to gain greater control over their technology and operational costs. This move represents a significant departure from the previous trend where many legal AI products functioned as 'model routers,' switching between various external AI platforms depending on the task. The new approach leverages open-source foundation technology that can be customized, allowing vendors to manage their own model inference and associated expenses. This development is driven by a desire for improved unit economics and greater control over their AI infrastructure.
Why It's Important?
This shift by legal AI vendors has significant implications for enterprise legal operations leaders, transforming how contracts and governance are managed. When a vendor operates its own model, critical aspects such as pricing mechanics, security ownership, and data portability are altered. The primary benefit for vendors is enhanced cost control, as they can reduce payments to external AI providers for inference, potentially improving profitability. For buyers, this could translate into more stable pricing and fewer usage limits or feature gating in the long run. However, this also means vendors assume greater responsibility for security and other obligations previously handled by frontier labs. Legal departments will need to scrutinize vendors' model lifecycles, including red-teaming cadences, evaluation sets for hallucination, citation behavior, and the speed of bug fixes, as these responsibilities now fall more directly on the vendor.
What's Next?
The legal AI procurement landscape is evolving into a 'stack decision,' where legal departments must carefully consider who bears the cost of inference, who is responsible for patching models, and who ensures the safety of using these models for sensitive work. Contracts will need to be updated to reflect these changes, especially if a vendor switches from a named subprocessor to an in-house model, as this alters the risk profile. Budgeting will also be affected, as vendor-hosted models can change cost curves for heavy users, particularly in tasks like contract review and large document summarization. While average attorneys may not immediately notice changes in model quality, legal operations teams must remain vigilant, as underlying costs, security postures, and auditability can materially change even if the user interface remains the same. This necessitates stronger governance when model swaps occur.
Beyond the Headlines
This trend highlights a broader movement within the AI industry towards vertical integration and proprietary model development, driven by the desire for cost efficiency and strategic independence. It underscores the growing maturity of AI technology, where companies are moving beyond simply leveraging existing models to building their own specialized solutions. This could lead to increased innovation within specific industry verticals, as vendors tailor AI models more precisely to their unique needs. Furthermore, it raises questions about data sovereignty and intellectual property, as more sensitive legal data will be processed by vendor-controlled models. The shift also signals a potential power rebalance in the AI ecosystem, as reliance on a few dominant frontier model providers may diminish over time, fostering a more diverse and competitive landscape for AI development and deployment.
