CoCounsel's Generational Refresh: Capability vs. Control

Thomson Reuters quietly opened early access to a substantially reworked CoCounsel in July 2026, and the legal technology industry is paying attention. The company is calling it the most significant overhaul of the platform since the $650 million Casetext acquisition closed in 2023—a claim that carries weight when you look at what has actually changed. This is no longer a research assistant with a polished interface. The new CoCounsel is being positioned as an agentic legal work platform: multi-step task orchestration, document-centric workflows, tighter Westlaw integration, and an architecture built to string together the kind of reasoning chains that previously required a paralegal to supervise.

The feature set is genuinely impressive. And it should prompt a conversation at your firm that has nothing to do with whether the features work.

The question worth asking in July 2026 is not "Is CoCounsel good?" It almost certainly is. The question is: what does your firm's infrastructure look like on the day Thomson Reuters changes its pricing, its data policy, or its API terms? That conversation is harder, less exciting, and more consequential.

The Agentic Turn Is Real—and It Changes the Stakes

For the past three years, the primary competition among legal AI platforms was on research quality. Which system surfaced the right cases faster? Which hallucinated less? Which had better coverage of secondary sources? Those are still relevant questions, but they are no longer the differentiating ones.

The major platforms—CoCounsel, Harvey, Lexis+ Protege, Legora, and others—have converged on research quality to a degree that makes it a threshold requirement rather than a competitive advantage. The new battleground is agentic capability and workflow depth: how many steps can the system handle autonomously? How deeply is it integrated into matter management, document review, and drafting pipelines? How much of a lawyer's daily work can it absorb before requiring human intervention?

CoCounsel's generational refresh is the clearest signal yet that Thomson Reuters understands this shift. The platform is now competing not just on what it knows but on what it can do—autonomously, across a sequence of tasks, with your documents at the center.

This shift matters architecturally, not just functionally. When a platform moves from "answer this research question" to "manage this workflow across these 400 documents," the data relationship between the firm and the vendor changes fundamentally. More of your matter data lives inside the platform's processing pipeline for longer periods. The retrieval index that surfaces relevant precedents and clauses is built and maintained on vendor infrastructure. The audit trail of what the agent did, and why, is stored in vendor logs.

For BigLaw firms with dedicated technology counsel and enterprise agreements that include data processing addenda, audit rights, and negotiated egress terms, this is manageable. For a 60-attorney regional firm in Nashville or a 120-attorney full-service firm in Phoenix, it is often a risk that does not get priced into the subscription decision.

What the CoCounsel Refresh Actually Signals

To be precise about what Thomson Reuters has built: the next-generation CoCounsel is not simply a UI refresh or a model upgrade. Based on what the company has disclosed in its early access communications, the rework centers on three structural changes.

First, agentic task orchestration. The platform can now decompose complex legal tasks into sequences of subtasks, execute them in order, and synthesize results—without requiring the attorney to manually prompt each step. Ask it to analyze a portfolio of commercial leases for landlord-favorable provisions, flag deviations from a standard form, and draft a summary memo, and it will attempt all three as a connected workflow rather than three separate sessions.

Second, document-centric architecture. Earlier versions of CoCounsel were primarily retrieval-forward: you asked a question, it searched. The new architecture inverts this to some degree—your uploaded documents become the primary context, with Westlaw research supporting rather than leading. This is more useful for transactional work, due diligence, and contract analysis, where the client's documents matter more than published precedent.

Third, deeper platform integration. CoCounsel now connects more directly to Thomson Reuters' broader product ecosystem, including HighQ and Elite 3E, moving toward the vision of an AI layer that sits across the firm's full technology stack rather than in a separate research window.

Each of these is a meaningful capability advancement. Each also deepens the firm's operational dependence on Thomson Reuters' infrastructure decisions.

The TCO Question Mid-Market Firms Are Not Asking

For firms in the 20–200 attorney range—the segment that represents the largest share of the AmLaw 200's client firms and the majority of independent regional practices—the total cost of ownership conversation around legal AI tends to stop at the subscription price. It should not.

Consider what a firm actually surrenders when it adopts a fully hosted platform at the depth CoCounsel is now operating:

The subscription fee for a platform like CoCounsel at the firm level is real—estimates for mid-market firms using enterprise legal AI platforms range from $50,000 to $250,000 annually depending on seat count and tier, a figure that has trended upward as platforms add agentic capabilities that justify premium pricing. But the subscription fee is the visible cost.

The invisible costs are infrastructure lock-in, retrieval dependency, and the negotiating position the firm occupies when the vendor's next pricing letter arrives. A firm that has run 18 months of matter documents through a vendor's document-centric platform, built workflows around its connectors, and trained its attorneys on its interface is not in a strong position to push back on a 30% price increase.

This is not a hypothetical. Legal SaaS pricing has followed the same trajectory as every other enterprise software category: competitive pricing during the land-and-expand phase, followed by consolidation-driven increases once switching costs are high. Thomson Reuters is a publicly traded company with margin targets. The current pricing reflects competitive pressure from Harvey, Lexis+ AI, and others. That pressure will not be constant.

For a deeper breakdown of how these costs compound over a three-to-five year horizon, see our legal AI total cost analysis.

The Architectural Alternative: Capability Without Surrender

The honest framing here is not "use RAGbase Legal because we never touch your data." That would be misleading. Any serious legal AI system that uses frontier language models—whether GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, or the models underlying platforms like CoCounsel—involves sending something to an external model provider. The question is what gets sent, under whose terms, and what stays home.

RAGbase Legal's architectural position is specific: the agentic scaffolding, retrieval index, vector stores, document permissions, connectors, workflows, and full client document corpus remain on the firm's infrastructure—on-premise, in the firm's private cloud, or in a dedicated environment the firm controls. What may leave that perimeter is only the minimal retrieved chunk needed to answer a specific query, transmitted to the firm's selected LLM provider under the firm's own API agreement.

This distinction matters in practice:

• Full corpus vs. minimal chunk: A vendor-hosted platform processes your entire uploaded document set within its environment. A firm-controlled architecture sends only the retrieved passage most relevant to the specific query. If you're analyzing a 500-page acquisition agreement, the vendor-hosted system processes all 500 pages in its environment. The firm-controlled system retrieves the three relevant clauses and sends those.

• Vendor API terms vs. firm API terms: When Thomson Reuters selects the underlying LLM for CoCounsel, it negotiates terms at its scale for its use case. When a firm uses RAGbase Legal, it selects the model provider and signs its own API agreement—which can include zero-retention terms, specific data processing addenda, and model selection that matches the sensitivity of the workload.

• Vendor audit logs vs. firm audit logs: In a firm-controlled architecture, the log of what the agent retrieved, what it sent to the model, what it returned, and who authorized the query lives on the firm's infrastructure. This matters for privilege analysis, malpractice defense, and client confidentiality obligations.

For private AI deployment at mid-market firms, this architecture is now operationally achievable. The infrastructure requirements that made on-premise AI prohibitive in 2021—GPU clusters, specialized MLOps teams, custom model training—have been largely eliminated by the API-based model access that frontier providers now offer. A firm does not need to run its own model. It needs to own its own retrieval layer, and that is a meaningfully different technical requirement.

What Sovereignty-Critical Workloads Actually Look Like

Not every task a firm runs through a legal AI platform carries the same data sensitivity. A research query about whether a specific circuit has adopted the economic loss doctrine does not implicate client confidentiality. A document analysis workflow running across a client's internal communications during pre-litigation discovery absolutely does.

Firms should be mapping their AI use cases to their data sensitivity profile—and the result of that mapping should inform architecture decisions, not just policy memos.

High-sovereignty workloads where firm-controlled infrastructure is most defensible:

• M&A due diligence document review (client financial records, internal communications)
• Pre-litigation document analysis (potentially privileged material)
• Regulatory investigation support (government-sensitive or confidential business information)
• Employment matter document review (HR records, individual personnel files)
• Patent prosecution document analysis (trade secrets embedded in technical documents)

Standard workloads where hosted platforms perform well and the data risk is lower:

• Statutory and case law research
• Publicly available regulatory analysis
• Legal drafting from standard form templates
• CLE research and continuing education support

The practical answer for most mid-market firms is not to choose between hosted and firm-controlled architecture across the board—it is to use hosted platforms for standard workloads while maintaining firm-controlled infrastructure for sovereignty-critical matters. This is the same tiered approach that enterprise IT has applied to cloud adoption for a decade: public cloud for non-sensitive workloads, private infrastructure for regulated or sensitive data.

For a broader view of how this tiered approach applies across practice areas, see our AI for law firms guide and the case search architecture overview for retrieval-specific implementation details.

What the Next 18 Months Will Reveal

Thomson Reuters is not alone in this generational push. Harvey has been building toward agentic workflows for transactional practices. Lexis+ AI's Protege is deepening its integration with LexisNexis' content universe in ways that parallel CoCounsel's Westlaw integration. Legora is advancing multi-agent collaboration features aimed at larger team workflows. The direction is consistent across all major players: deeper, more autonomous, more document-centric, more integrated with the firm's operational data.

This convergence will produce better tools. It will also produce greater switching costs, higher renewal pricing for firms that have deeply integrated these platforms, and more complex data relationships that are difficult to unwind once established.

The firms that will negotiate from strength in 2027 and 2028 are not the ones that adopted the best platform in 2026—they are the ones that adopted capable platforms while maintaining architectural options. That means owning the retrieval layer. It means knowing exactly what data is processed where. It means having an audit trail that belongs to the firm, not the vendor.

CoCounsel's generational refresh is impressive. The conversation your firm should also be having is about what your infrastructure looks like on the day the terms change—because in enterprise software, they always do.

If your firm is evaluating how to layer agentic AI capability against the infrastructure decisions that determine your long-term position, the private AI deployment architecture overview is a practical starting point. The question is not whether to use the best available models and platforms—it is whether the orchestration layer, retrieval index, and document custody sit on your infrastructure or someone else's. That distinction is worth mapping before the next subscription renewal, not after.