Mar 10, 2026 | AI and product innovation
Rebuilding for the Agent Era: The Next Generation of CoCounsel Legal
When we first brought generative AI into Thomson Reuters products, the goal was practical. Help professionals complete real work faster, with systems they could trust.
One million users now have access to CoCounsel. That scale made something clear. The opportunity is larger than a collection of AI-powered skills.
We are at an architectural inflection point. Generative AI is not another feature layer in professional software. It enables automation by simulating human judgment in the execution of a workflow. When that happens, incremental iteration is not enough. Core systems must be rebuilt.
Every major professional software platform will face this decision. Over the past year, we made the decision to rebuild CoCounsel Legal for the agent era.
This is not a feature update. It is a foundational transformation.
From Delegation to Planning
The first generation of CoCounsel Legal was built around specialized skills and guided workflows. That model worked well for clearly defined use cases. But professional work does not stay within predefined lanes.
As new needs emerged, we added more skills. Routing grew more complex. Multi-step work requires chaining workflows together. The system was more capable, but the burden on users grew – to learn what skills were available, how to call them, and how to combine them into a workflow.
The next version of CoCounsel removes that burden while simultaneously handling a much wider range of legal tasks. The system understands objectives as the user describes them, and it comes up with the plan for users to adjust if they wish.
Building a Legal AI Platform
The next generation of CoCounsel Legal is being built as a highly capable, customizable legal AI platform.
Instead of routing questions to isolated skills, the system designs a solution specific to each request. It plans. It selects tools. It retrieves authoritative content. It adapts as new information emerges. In the agent era, AI will not simply sit beside professional workflows. It will guide them. This is not a feature enhancement. It is a platform re-architecture.
The first wave of generative AI in SaaS attached models to existing workflows. The next wave requires redesigning the system around reasoning, orchestration, and verification. That is the shift we made.
Our objective is to build an extensible operating system for legal work, one that allows new tools, workflows, and models to evolve on top of it. We are leveraging modern agent frameworks and investing heavily into domain specific differentiators – grounding responses in authoritative content, easing the task of fiduciary-grade verification, and leveraging the full breadth of domain expertise from practical law.
The new system is content native by design. It directly accesses Westlaw, Practical Law, KeyCite, and customer documents, where permissioned, through structured content tools built for agentic reasoning.
We are applying the same taxonomy, citation, and metadata discipline to customer content that is brought into the workflow for a specific customer, with appropriate permissions, as we do for authoritative content. That enables the agent to reason not only over authoritative external sources, but also over that customer’s own precedents and institutional knowledge within a clearly defined and governed boundary.
Language models do not store up-to-date, accurate, or verifiable knowledge of the law. Trust in high-stakes domains requires grounding in authoritative content and customer context.
This is what makes AI professional-grade, and with this next evolution, it moves even further toward fiduciary-grade AI, designed to meet the highest standards of responsibility and client duty.
We remain model agnostic. The model layer will evolve. Our responsibility is to ensure that whichever models we use are paired with strong content intelligence, evaluation discipline, and enterprise guardrails.
Measuring the Shift
Architectural ambition must be matched with disciplined evaluation.
We constructed an internal evaluation set representing hundreds of real-world legal tasks across research, drafting, and analysis. These reflect the work professionals actually attempt to complete.
Thousands of evaluation runs have been reviewed by subject matter experts, representing tens of thousands of hours of expert review and annotation. In high-stakes domains, evaluation isn’t a quality gate for release; it’s how you build the product.
Where we previously measured the success rate of individual skills, we now measure whether the system can complete a natural task end to end, even when there is no prebuilt skill designed for it. This evaluation is intentionally unforgiving: unconstrained, multi-step requests that mirror real work, reviewed by subject matter experts. Under this stricter methodology, CoCounsel completes roughly three quarters of these workflows, 76% in the latest evaluation runs, with virtually no learning curve for users. This is a step change in capability that will feel substantially different to users on day one.
What This Signals
We believe professional-grade AI platforms require four elements:
- Advanced reasoning models and an agentic harness to maximize their capabilities
- Domain-specific content and intelligence
- Deep customer context and workflow integration
- Human expertise, evaluation, and governance
At scale, this is not theoretical. It is already supporting more than one million users and evaluated across hundreds of natural, end-to-end workflows under structured subject matter expert review.
Models alone are not enough. Content alone is not enough. Without all four working together, systems fail under real professional pressure.
The organizations that define this era will not be those that add AI features fastest. They will be those that build well-governed, extensible platforms capable of harnessing model improvements while ensuring seamless human oversight and built-in verification
What Comes Next
We are expanding structured content tools, deepening the integration between customer context and authoritative content for complex workflows, building matter-based workspaces, enabling firm-level customization, and extending third-party integrations so the agent can operate across the systems our customers use every day.
Broader availability of CoCounsel Legal in its next generation is planned for later this year.
A Different Standard
AI will not replace professional judgment. But platforms architected for real work, grounded in authoritative content, and evaluated with rigor can materially change how and where that judgment is applied.
For a company with more than 170 years of history, transformation is not about chasing trends. Rearchitecting CoCounsel for the agent era sets the foundation not just for our next release, but for how professional work will be executed in the years ahead.
We believe this architectural shift will define the next generation of enterprise software. The question is not whether agents will participate in complex work, but which platforms will be trusted to run them.
Read more: CoCounsel Legal, Reimagined. Fewer Steps. Bigger Outcomes
