Thomson Reuters Builds Its Own AI Model as Revenue Accelerates

Thomson Reuters is moving its first production-ready in-house AI model into CoCounsel after quarterly revenue rose 9 percent. The timing creates a sharper conflict than the headline circulating through Google News suggests. The company is not simply adding another AI feature. It is testing whether authoritative professional content can reduce its dependence on models supplied by frontier AI laboratories.

The model, called Thomson, will initially power Tabular Analysis, a bulk document review capability within CoCounsel Legal. Thomson Reuters expects that migration to begin in August 2026. Management says the model offers lower latency and operating costs than third-party alternatives, although its comparative results have not been independently validated.

That puts Thomson Reuters on a different path from software vendors that treat OpenAI, Anthropic, Google, or another model provider as a permanent foundation. The company still uses external models, and CoCounsel remains a multi-model product. However, Thomson gives it an owned alternative for selected work where legal knowledge, predictable costs, and customer control matter more than general conversational range.

What Thomson Reuters Actually Changed

Thomson Reuters has turned its proprietary model from a research project into production infrastructure.

The company completed the first production-ready version of Thomson after combining teams from Thomson Reuters Labs and Safe Sign Technologies. Thomson Reuters acquired Safe Sign, a developer of legal-specific language models, in 2024. That acquisition provided expertise for adapting open-source model foundations to professional legal work.

Thomson is a domain-specific large language model, meaning its training and evaluation focus on a defined field rather than every possible consumer task. The company says it used expert-curated material from Westlaw, Practical Law, and Reuters alongside input from attorney editors and practice specialists.

Thomson Reuters invested approximately $40 million in the model and trained it on less than 10 percent of its legal content. Chief Executive Steve Hasker told investors that internal benchmarking placed Thomson near leading frontier models across broad general tasks. He also said it performed strongly on legal evaluations.

Those claims deserve careful attribution. The company has not published enough information for outsiders to reproduce the complete comparison. It has not disclosed every underlying model, test set, scoring method, or production constraint involved.

The important change is therefore not a claimed benchmark victory. It is the decision to deploy Thomson in a real customer workflow. According to the company’s earnings transcript, Tabular Analysis will be the first meaningful production test.

Tabular Analysis reviews information across groups of documents and organizes the findings into a structured form. Lawyers can use that capability when comparing contract terms, reviewing evidence, or locating recurring provisions across a document collection.

This task gives Thomson Reuters a contained starting point. Outputs can be checked against source documents, while speed and inference costs can be measured against existing systems. A failure would also be easier to diagnose than an error inside a broad, open-ended legal assistant.

Management says the company will consider moving additional CoCounsel functions onto Thomson after evaluating this first deployment. It is not planning an immediate replacement of every external model. Hasker acknowledged that CoCounsel already works well and said the company wants to avoid disrupting it unnecessarily.

That measured rollout matters. The story surfaced widely through Google News as an in-house model launch, but the operational reality is narrower. Thomson Reuters is beginning a staged transfer of selected workloads, with broader adoption dependent on actual production results.

The deployment also creates a new option for large law firms. Hasker said Thomson Reuters is discussing environments where firms could run a version of Thomson alongside their own protected information. CoCounsel could then operate above that combined knowledge layer.

Such an arrangement is sometimes described as sovereign AI. In this context, the term means a firm retains greater control over its data, model environment, and operating policies. It does not mean the system functions without external software, infrastructure, or ongoing vendor support.

Rising Revenue Gives the AI Strategy Room to Run

The company is funding its model strategy from a growing subscription business, not from an isolated experimental budget.

Thomson Reuters reported second-quarter revenue of $1.954 billion, compared with $1.785 billion one year earlier. Total revenue increased 9 percent, while organic revenue grew 8 percent. Recurring revenue also rose 9 percent and represented 82 percent of quarterly revenue.

Its Legal Professionals, Corporates, and Tax, Audit and Accounting Professionals divisions form what the company calls its Big 3. Those segments produced $1.62 billion in revenue during the quarter and accounted for 83 percent of the company total.

Big 3 organic revenue increased 10 percent. Legal Professionals generated $772 million, up 10 percent on a reported basis and 10 percent organically. Thomson Reuters identified Westlaw and CoCounsel as primary contributors to recurring legal revenue growth.

Corporates revenue reached $537 million, while Tax, Audit and Accounting Professionals contributed $311 million. The quarterly results show growth across all three segments, although their individual revenue mixes and seasonal patterns differ.

Operating profit rose 28 percent to $558 million. Adjusted EBITDA increased 10 percent to $745 million, while the adjusted EBITDA margin moved from 37.8 percent to 38.1 percent. Free cash flow increased 29 percent to $727 million.

These figures do not prove that generative AI caused the quarter’s overall growth. Westlaw, transactional products, acquisitions, international businesses, tax software, and other established services also contributed. Thomson Reuters does not publish enough product-level revenue information to isolate CoCounsel’s exact effect.

Nevertheless, the results weaken one concern surrounding the company’s AI transition. Thomson Reuters is not trying to rebuild its product foundation while its central professional businesses shrink. Revenue growth and recurring subscriptions give it time to test new architecture without depending on immediate standalone model sales.

Management raised its full-year outlook to approximately 8 percent total and organic revenue growth. It also lifted expected organic growth for the Big 3 to a range of 9.5 percent to 10 percent.

The company’s portfolio changes reinforce that priority. Thomson Reuters agreed to sell a 51 percent interest in its Global Print operation to a joint venture controlled with KKR-advised funds. It expects approximately $500 million in gross proceeds when the deal closes.

Global Print revenue declined 3 percent in the quarter, while its professional software segments expanded. Moving control of print into a joint venture lets Thomson Reuters concentrate management attention and investment on content-based software and AI.

This is the central reversal inside the earnings story. Publishers and information providers were initially viewed as likely victims of generative AI. Models could summarize their material, weaken traditional search traffic, and place a conversational interface between information owners and customers.

Thomson Reuters is instead trying to make its accumulated content the model’s differentiating input. The same legal collections that supported subscription research products could become training, grounding, and evaluation assets for AI systems.

The company is not abandoning its established products. It is attempting to move their authority into new interfaces before general-purpose assistants capture the professional workflow.

For enterprise buyers, the revenue context matters more than a prominent headline on Google News. It suggests Thomson is part of a continuing product transition backed by recurring income, rather than a temporary demonstration built for attention.

Thomson AI Challenges Dependence on Frontier Models

The primary contest is not Thomson versus one AI laboratory. It is owned domain infrastructure versus permanent reliance on external frontier models.

Most enterprise AI products began with a practical shortcut. Vendors connected a capable general-purpose model to proprietary data through retrieval-augmented generation, or RAG. RAG finds relevant source material at query time and supplies it to the model before an answer is produced.

That approach remains useful. It lets a software provider benefit from improvements made by large model laboratories without financing a complete training program. It can also preserve citations and restrict answers to approved information.

However, dependence on third-party models introduces strategic limits. A provider does not fully control model changes, inference prices, availability, latency, or future access conditions. It must also distinguish its product when competitors can buy access to similar underlying intelligence.

Thomson Reuters already works with external models and is not presenting Thomson as an immediate universal substitute. The company says its advantage comes from selecting open-source foundations and adapting them with carefully curated professional material.

Management argues that data quality can compensate for a smaller training budget. Hasker said the company achieved competitive internal results with approximately $40 million of investment and less than one-tenth of its legal content.

The claim is plausible as a strategy, but it is not yet established as a general rule. A specialized model does not need to answer every question better than a frontier model. It needs to perform selected professional tasks accurately enough, quickly enough, and cheaply enough to improve the full application.

For Thomson Reuters, that application includes authoritative retrieval, workflow logic, legal citations, user permissions, document handling, and auditability. The model is only one layer.

This is where the company’s position differs from that of a new legal AI startup. Thomson Reuters already controls Westlaw, Practical Law, and a large network of editors and subject specialists. It also has distribution through products embedded in law firms, corporate legal departments, tax teams, and accounting practices.

That does not make the company immune to competition. Anthropic’s expansion into Claude for Legal shows that frontier model developers can move closer to professional customers. Legal AI companies such as Harvey and Legora are also building interfaces, workflows, and firm integrations around external models.

LexisNexis has its own legal content, professional distribution, and AI products. Large law firms can also assemble internal systems that combine commercial models with private work product.

The threat therefore comes from several directions, but they all create the same pressure. Thomson Reuters must show that owning more of the model layer produces an advantage customers can feel. Lower internal cost alone will not persuade a lawyer to change tools.

The company expects Thomson to reduce latency, which is the delay between a request and a response. That benefit matters when an agent performs many sequential operations. A small delay repeated through 20 or 30 steps can make a workflow frustrating.

Control over inference costs could also become important as usage increases. An AI assistant used occasionally for research has a different cost profile from an agent reviewing thousands of documents or completing repeated compliance work.

CoCounsel has already passed one million users, according to Thomson Reuters. Management says interactions and session duration continue to rise, although it has not disclosed active-user rates or model-level economics.

Higher engagement turns model cost into a strategic issue. Every additional task creates compute expense. An owned model that performs suitable work at a lower cost could protect margins while allowing broader use.

The tradeoff is maintenance. Thomson Reuters must continue improving the model, evaluating new open-source foundations, securing data pipelines, and responding to rapid advances elsewhere. Model ownership exchanges one form of dependency for an ongoing engineering obligation.

What the Thomson Reuters Benchmarks Do Not Show

Internal evaluation results are encouraging evidence, but they are not a substitute for transparent production measurements.

Thomson Reuters says Thomson performs on par with leading frontier models across a broad set of general tasks. It also reports strong legal performance, lower latency, and meaningfully lower costs.

The company has released directional findings, but readers should separate three questions. How did Thomson score in controlled evaluations? How reliably does it perform inside CoCounsel? Do customers receive better outcomes than they receive from competing systems?

Only the first question has a preliminary company answer. The second is being tested through Tabular Analysis. The third will require comparative usage, retention, accuracy, and purchasing evidence over time.

Legal AI benchmarking is especially difficult because common public tests rarely reproduce professional work. A model might identify a legal rule correctly but fail to account for jurisdiction, procedural posture, an exception, or a later authority.

Long documents create another problem. A model may accept a large amount of text while still overlooking decisive language buried inside it. Thomson Reuters has previously discussed the challenge of evaluating long-context performance, particularly when legal work involves extensive records.

The company’s proprietary tests may better reflect real work than public benchmarks. Yet proprietary evaluation also limits outside scrutiny. Customers cannot independently determine whether the selected tests favor Thomson’s training data or product design.

Training on Thomson Reuters content creates further questions. The company has enormous legal and professional collections, but more content does not automatically produce proportional improvement. Data must be cleaned, structured, weighted, updated, and connected to jurisdiction-specific metadata.

Thomson Reuters worked with DatologyAI on legal domain adaptation and reported improvements in legal reasoning, retrieval, and downstream tasks. The partner’s data curation case describes better results with a limited data and compute budget. However, that account is also supplied by companies involved in the project.

Production use will expose harder cases. Lawyers may upload inconsistent agreements, incomplete records, scanned documents, or private templates. A system must distinguish authoritative law from client material and clearly reveal which source supports each conclusion.

CoCounsel’s rebuilt agentic interface is designed to show its work. Agentic AI refers to software that plans and executes a sequence of actions toward a goal, rather than returning one response to one prompt.

Hasker said users can inspect the steps, citations, and references produced during a CoCounsel task. That transparency is valuable because a professional can challenge an intermediate decision before relying on the final output.

Still, visible reasoning steps do not guarantee correctness. A well-organized chain can contain an early classification error that shapes everything after it. Citations can be real while failing to support the proposition attached to them.

Human review therefore remains part of the product’s value proposition. Thomson Reuters calls its approach “fiduciary-grade AI,” a company-defined standard emphasizing authority, auditability, security, and accountable professional use.

That label should not be mistaken for an independent certification. Buyers need evidence tied to their own risk requirements, including error rates, source coverage, access controls, logging, retention practices, and escalation procedures.

Firms considering these systems also need a durable knowledge layer. Matter records, internal guidance, and prior work must remain organized before any model can retrieve them reliably. A searchable knowledge base can reduce the gap between impressive model demonstrations and dependable daily work.

The skeptical case is therefore straightforward. Thomson Reuters owns unusually valuable data and has reported promising results. It still must prove that these assets translate into repeatable customer outcomes outside tests designed by the company.

Why Law Firms and Enterprise Buyers Should Care

Thomson creates new purchasing options, but it also makes AI architecture a more important part of vendor evaluation.

Professional buyers once compared research coverage, software features, support, and contract terms. They now must ask which models perform each task, where those models run, how customer data is handled, and what happens when a provider changes the underlying architecture.

Thomson Reuters can offer a hybrid answer. CoCounsel may direct selected tasks to Thomson while continuing to use other models where they perform better. This can make the application more adaptable than a single-model system.

A hybrid design can also make evaluation harder. A customer may see one product interface without knowing which model generated a particular result. Buyers should request clear documentation covering model routing, material updates, testing standards, and incident response.

Large firms have another option. Hasker said some firms are interested in combining Thomson with their protected internal information inside a controlled environment. That arrangement could give a firm greater authority over data and deployment while preserving access to Thomson Reuters content and CoCounsel workflows.

The appeal is understandable. Law firms hold confidential client records, negotiation histories, internal precedents, and attorney work product. Sending every task through a broadly shared external service can create governance concerns even when contractual protections exist.

An isolated or dedicated environment does not solve every issue. Firms still need access controls, data classification, retention rules, security reviews, and a clear division of responsibility between vendor and customer.

Smaller organizations face a different calculation. They are less likely to operate dedicated AI infrastructure or maintain extensive evaluation teams. They need a packaged service with understandable controls and predictable behavior.

For those buyers, model ownership matters only if it improves the final product. Faster document review, clearer citations, fewer unsupported conclusions, and consistent availability are meaningful. The model’s name is not.

Knowledge workers should also expect the interface to change. An agent that performs research, reviews documents, and drafts an output shifts the user’s role from assembling every step to supervising a process.

That can save time, but it also concentrates risk. A user who manually conducts research encounters evidence throughout the task. A user supervising an agent may see only selected sources and a summarized trail.

Training must therefore focus on verification rather than prompt tricks. Professionals need to know how to challenge scope, check authority, identify missing documents, and review the system’s sequence of actions.

The one-million-user milestone suggests CoCounsel already has meaningful distribution. However, Thomson Reuters has not disclosed how many of those users are active each day, how activity varies by profession, or what share uses the newest agentic version.

Enterprise adoption cannot be measured through account access alone. Sustained value appears when users return, complete substantial work, and renew without creating unacceptable review burdens.

This distinction is easy to lose when following AI stories through Google News. Product access generates an announcement, while dependable adoption develops slowly inside organizations.

The company’s quarterly results offer an early commercial signal because Westlaw and CoCounsel contributed to legal recurring revenue growth. They do not yet separate expansion from pricing, bundling, new customers, or migration among existing products.

Buyers should treat Thomson as a meaningful architectural change, not proof of a settled winner. It gives Thomson Reuters greater control over cost and product development. Customers must determine whether that control improves their own work.

Three Signals That Will Decide Whether Thomson Works

The next phase will be judged by production performance, customer deployment, and measurable financial contribution.

The first signal is the August migration of Tabular Analysis. Thomson Reuters must show that Thomson can handle bulk document review with acceptable accuracy, speed, and stability under real workloads.

This deployment will test the company’s central mechanism. If latency falls and users complete more document-heavy tasks without losing reliability, the case for shifting additional CoCounsel capabilities becomes stronger.

A quiet delay, limited rollout, or return to third-party models would weaken that case. So would a lack of specific performance reporting after management emphasized the model’s cost and speed advantages.

The most useful disclosure would compare Thomson with the previous production setup across task completion, reviewed-document volume, latency, error escalation, and compute cost. Even partial operational data would be more informative than another general benchmark claim.

The second signal is whether Thomson Reuters announces a sovereign deployment with a major law firm. Management says conversations are underway, but no named agreement or production configuration has been disclosed.

A real customer deployment would expand Thomson beyond an internal CoCounsel component. It would indicate that firms see value in combining a professional model, Thomson Reuters content, and private institutional knowledge inside a controlled environment.

The details will matter. A pilot involving a narrow practice group would signal experimentation. A broader deployment covering protected firm material and recurring workflows would support the company’s claim that model ownership creates a durable business opportunity.

The third signal is financial. Investors need clearer evidence that CoCounsel and other AI products add recurring revenue without eroding margins through compute, development, sales, and support costs.

Thomson Reuters reported strong overall growth and a higher adjusted EBITDA margin in the second quarter. It also raised its revenue outlook. Those results provide room for investment, but they do not isolate Thomson’s economics.

Future earnings calls should reveal whether AI increases customer spending, redirects existing software budgets, or mainly protects established subscriptions. Each outcome has different implications.

New spending would support management’s description of legal AI assistants as a white-space opportunity. Budget substitution would make Thomson more defensive, preserving customer relationships while changing the underlying cost base.

The competitive response also deserves attention. Frontier laboratories will continue improving their models, and specialized legal products will add deeper integrations. Large law firms will keep experimenting with private systems and multiple vendors.

Thomson Reuters does not need Thomson to defeat every alternative. It needs the model to remain good enough for defined professional tasks while giving CoCounsel better economics, control, and access to proprietary knowledge.

That is a demanding target, but it is more realistic than competing with frontier laboratories across every capability. The company can route other work to external models and reserve Thomson for areas where its content changes the result.

Readers arriving from Google News should therefore look beyond the apparent model launch. The consequential event is a production test of whether a long-established information provider can turn trusted content into owned AI infrastructure.

Watch the Tabular Analysis migration first. Then watch for a named law-firm deployment and clearer AI revenue economics. Those signals will show whether Thomson becomes a real platform advantage or remains an interesting option inside a much larger software portfolio.