The story of how CoCounsel built trust in legal AI by prioritizing reliability over novelty
Highlights
- CoCounsel launched as the first GPT-4-powered legal AI assistant in March 2023.
- Integration with Westlaw and Practical Law enabled trusted, end-to-end legal workflows.
- Agentic AI architecture now allows intent-driven outcomes rather than task-based instructions.
The night CoCounsel’s team got early access to GPT-4, most people didn’t yet know what large language models could do. Inside Casetext, the question was different: not what AI could do, but whether it could meet the standards of real legal work.
Matthew Heagerty, now VP of Product for CoCounsel Skills, remembers the moment clearly. “It felt like a wizard playing with magic,” he says. “But we immediately started stress-testing it against what lawyers actually need.”
From the start, CoCounsel pushed the frontier while holding itself to the standards of real legal work. That tension has defined its evolution. From an early breakthrough in generative AI to a deeply integrated legal platform to what comes next, the team has consistently been ahead, building solutions not just for the sake of showing what AI can do, but for what legal professionals actually need and can trust.
Jump to ↓
Early innovation, grounded in legal reality (2023)
From isolated tasks to legal workflows (2023-2024)
Raising the bar on legal tech (2024-2025)
The shift to a more intuitive, agentic experience (2026 and beyond)
Early innovation, grounded in legal reality (2023)
Heagerty joined Casetext, pre-Thomson Reuters acquisition, driven by a question every practicing attorney asks themselves mid-task: is there a better way to do this? He got his answer before most of the industry knew the question had changed. When Casetext launched CoCounsel on March 1, 2023, it was the first AI legal assistant of its kind, built on early, exclusive access to GPT-4 before the model was publicly available. While the rest of the legal industry was still debating whether generative AI belonged in professional practice, CoCounsel was already in production at major law firms.
But being first only matters if you get it right. Possibility and reliability are different things, and the team knew it. Before launch, Casetext’s dedicated Trust Team ran every skill on the platform through thousands of internal tests — nearly 4,000 hours of training and fine-tuning based on over 30,000 legal questions. Context limits, latency, and hallucination risk were not hypothetical. They were design constraints that shaped the product from day one: document handling, grounding, verification, and synthesis all working in concert before a single output reached an attorney’s screen.
From the start, CoCounsel treated those constraints not as obstacles, but as design requirements. Legal AI had to meet the standards of legal work.
From isolated tasks to legal workflows (2023-2024)
CoCounsel’s early capabilities were not chosen for novelty. They were mapped directly to the moments in legal work where time pressure is highest and the cost of error is real. Tasks such as timeline creation, document review, comparison against policy, and contract data extraction were first on the list. “We looked at different skills that would be useful for litigators and transactional attorneys,” Heagerty explains. “Different pieces of the attorney workflow where we could apply this technology.”
That work deepened after Thomson Reuters’ acquisition of Casetext, which brought a broader pool of legal expertise into the development process. Attorneys from across practice areas helped the team pressure-test every capability against how legal work actually flows — which, as Heagerty notes, is rarely linear. “A lot of thought went into what the flow of a case looks like and the different steps involved. We had to break it apart to understand where we could genuinely help.”
The result was a platform that earned trust incrementally, through consistent performance in real workflows. Not by overpromising what AI could do but by demonstrating what it could do reliably. And critically, that reliability wasn’t built on generic web content. By grounding outputs in Westlaw’s primary law and Practical Law’s expert guidance, the platform could produce work that reflected the standards of the profession, not just the capabilities of the model.
Raising the bar on legal tech (2024-2025)
The integration of CoCounsel with Thomson Reuters content in early 2024 marked an inflection point. By bringing together CoCounsel’s generative AI capabilities with trusted sources like Westlaw and Practical Law, the platform moved beyond standalone outputs to connected, end-to-end workflows. Research could inform drafting. Expert guidance could shape analysis. Authoritative content could be applied throughout the process, not added after the fact.
The underlying technology evolved alongside it. Moving from discrete skills to full workflows required more than better prompts — it required a fundamental leap in model capability. “First and foremost, the foundational models have just gotten much more capable.” Heagerty explains. That improved model capability meant that CoCounsel didn’t have to stop at retrieving the right information. It could now link separate skills and features to create end-to-end workflows, cutting down on the need to perform each part of that work in a silo. Tasks that once required multiple steps could now be executed within a single workflow, with CoCounsel showing its work through citations and sources grounded in Westlaw and Practical Law throughout. With model improvements also came the beginnings of agentic AI being applied to legal research, with CoCounsel’s Deep Research planning research steps, analyzing legal sources, and applying its findings to answer complex and layered legal questions.
Legal expertise remained central to every step. Many CoCounsel engineers bring prior legal practice experience and work directly with Westlaw and Practical Law attorney editors while building and testing new capabilities — always asking, as Heagerty puts it, “‘Is there a better way to do things based on this new technology? And how do we deliver the best value to our customers?'”
The shift to a more intuitive, agentic experience (2026 and beyond)
Today, CoCounsel is evolving again — this time from a collection of powerful skills and workflows to a more intuitive, agent-driven system.
Thomson Reuters’ partnership with Anthropic has been a significant catalyst — not just for the models powering CoCounsel, but for the architecture underneath it. ” The latest technology enables an architecture that was not possible when CoCounsel first launched,” Heagerty says. The Claude SDK (software development kit) gives the team an agentic harness that fundamentally changes how CoCounsel operates: rather than executing discrete tasks in sequence, the system can now plan, delegate, and orchestrate across multiple tools and content sources in real time.
This changes the interaction model entirely. Instead of asking the system to complete a task, professionals can define an outcome and rely on CoCounsel to plan and execute the steps required to get there.
That shift in architecture enables a shift in experience. The next iteration of CoCounsel focuses on intent rather than instruction. Users can articulate what they want to accomplish, and the system determines how to get there by planning, executing, and synthesizing across tools and content dynamically. For users, the experience becomes simpler and more aligned with how legal professionals actually work.
Heagerty has seen the excitement for these developments firsthand. At Legalweek earlier this month, he watched attorneys encounter the next generation of CoCounsel for the first time.
Multiple times I saw jaws hit the floor. You could see people immediately connecting how they would incorporate it into their practice. And also, how it will make their lives easier, how they could be more efficient and get better outcomes for their clients.
VP of Product, CoCounsel Skills, Thomson Reuters
Innovation that earns trust
CoCounsel’s journey reflects a broader truth about legal technology: progress is not defined by novelty. It is defined by confidence..
What has always differentiated CoCounsel isn’t just access to cutting-edge models. The difference is what surrounds the technology: legal expertise embedded into every layer of development, from engineers who’ve practiced law to attorney editors who validate outputs to the authoritative content that grounds every answer. When the people building the system understand what’s actually at stake when a lawyer works with an AI output, the product reflects that. The bar for easy verification, accuracy, and professional judgment gets built in from the start, not bolted on after the fact.
That balance between innovation and trust is what has always set Thomson Reuters apart. And it’s what continues to shape the future of CoCounsel.
As CoCounsel enters this next phase, it represents more than a new set of features. It signals a shift toward legal AI that is more intuitive, more integrated, and more aligned with the realities of legal work — while remaining firmly rooted in the standards that define the profession. The question was never just what AI could do. It was whether it could be trusted to do it.
That is the standard CoCounsel was built to meet. And as legal AI moves from capability to confidence, it is the standard that will define the category.
