Jun 22, 2026 | AI and product innovation

As AI becomes more capable, a new divide is emerging. Not between models, but between work that can tolerate a lack of precision and work that cannot.

General-purpose AI is advancing quickly and delivering real value. It will continue to augment the operational layer: workflow automation, routine document processing, and open-ended brainstorming. Then there is the work that must be cited, audited, and defended. Research that shapes case strategy. Regulatory submissions. Transactions where a partner puts their name on the closing certificate. This is the work that demands something general-purpose AI was not designed for and can’t reliably deliver through prompting alone. The AI to support high-stakes work needs authoritative grounding, citation integrity, and professional governance built into the architecture from the first step, not added at the end. It needs depth.

Those two categories complement each other, but they are separating in terms of their roles in the stack. And the question for every professional organization is no longer whether to adopt AI; it is which kind the work in front of them requires. Legal is where that question is sharpest, and where we are answering it first. But the same divide is opening across tax, accounting, audit, and risk, wherever professionals are expected to produce work that can withstand scrutiny. That is the work we built for, and today the next generation of CoCounsel Legal is here for early access.

The profession has already reached this conclusion

Our 2026 Future of Professionals research, drawing on more than 1,800 legal and tax professionals across 62 countries, makes the shift plain. Most now use AI several times a week. The open question is no longer adoption. It is trust. Asked what makes AI accountable to professional standards, 95% pointed to safeguarding confidential data, 94% to outputs grounded in verified content rather than the open internet, and 87% to work built on human expertise that can be explained and defended.

The stakes are no longer hypothetical. Courts have sanctioned lawyers for filings built on AI-generated citations that did not exist and have been clear that responsibility sits with the professional, not the tool. Adoption is accelerating anyway. That combination is exactly why the category of AI now matters as much as the capability of AI.

Agentic AI is not a counting exercise

The word agentic is everywhere, usually used to count things: how many tools a system can call, how many agents it can name. Those capabilities matter, but the number of agents is the wrong measure. We built the next generation of CoCounsel Legal on the principles that make agentic systems genuinely powerful, the same ones that make coding agents work: reason through the problem in testable steps, call the right tools as the work requires, adapt when new information changes the analysis, and keep the reasoning traceable enough to verify. Agentic capability is not about enumerating and building a deterministic set of skills; it is about building an intelligent system that is adaptable while being rooted in the tools and information that underpin the business and the industry.

What makes that reasoning trustworthy is where the knowledge sits. There is a meaningful difference between a system that calls Westlaw at the end of a task to check citations, and one where Westlaw and Practical Law are wired into how the agent reasons from the first step, content that took Thomson Reuters 175 years to assemble. The first generates and then checks. The second reasons against authoritative knowledge throughout. The outputs are not slightly different; they are structurally different. CoCounsel Legal is the second model. The attorney does not approve the work on faith. They inspect it. Building professional AI requires more than integrating frontier models. It also requires specialized models built for the work itself.

We are strengthening this from underneath, too. CoCounsel Legal is and will remain multi-model by design. Alongside the frontier providers we work with, we are building our own: Thomson, our vertically specialized foundation model, built for depth and accountability rather than breadth. It ships later this year, first in Tabular Analysis inside CoCounsel Legal, where extracting structured answers across large document sets rewards precision a general model handles inconsistently.

A standard you cannot measure is just a claim

If professional AI is moving from answers to defensible work product, evaluation has to move with it.

That is why we built CoCoBench, our lawyer-led evaluation framework: more than 1,000 tasks written by practicing attorneys and scored against gold-standard answers also written by attorneys, from Big Law, government, in-house, and boutiques. It does not ask whether an output sounds like legal analysis. It asks whether the reasoning is right for the jurisdiction, whether the correct framework was applied to the facts, and whether an attorney could defend the result under scrutiny.

Measured against representative legal tasks reviewed by licensed attorneys, CoCoBench shows what professional-grade evaluation requires. These are not abstract benchmark questions. They are legal tasks, informed by practice experience, held to the standard of work a professional would need to review, rely on, and defend. That bar gets set when verified content, agentic reasoning, expert evaluation, and deep workflow integration work as one system.

Built for where work begins, wherever that is

Professional work does not start in one place. It begins in research platforms, drafting tools, document systems, and increasingly in general-purpose AI assistants. So we are not trying to be every interface a professional touches. Through APIs, connectors, Model Context Protocol support, and partner integrations across the AI and enterprise ecosystem, CoCounsel Legal can be reached from the tools customers already use. When a lawyer hands off a task through MCP, they are not querying a content database. They are handing it to a system where authoritative content is wired into the reasoning, and the result can be inspected and defended. The general-purpose model starts the work. CoCounsel Legal finishes it to a standard that holds up.

This is what we mean when we describe CoCounsel evolving into an agentic operating system for professional work. Not a chatbot, not a list of agents, not a collection of plugins, but a trusted system that brings content, context, tools, models, verification, and human oversight together to move work to completion, while keeping people in the parts that matter most: judgment, strategy, review, and accountability.

Professionals will adopt AI when they trust it more than they fear it. That trust will not come from speed alone. It will come from systems that help them finish the work correctly and prove they did.

The future of professional AI will not be defined by who generates the most answers. It will be defined by who can help professionals produce work they can defend. That is the standard we believe matters, and the standard we are building toward.

That is what we built. The next phase of professional AI is here.

Read more: