The math alone is eye-catching: Thomson Reuters has poured more than $200 million into artificial intelligence over the past two years. But the real wager isn't in the dollars. It's in the belief that professionals—lawyers, accountants, tax advisors—are ready to hand off entire workflows to machines that think through problems on their own.
This isn't another chatbot promising to draft your emails faster. Starting last June, the legal and tax information giant began rolling out what it calls "agentic intelligence"—AI systems designed not to answer questions, but to plan research strategies, execute multi-step tasks, and navigate the kind of complex professional work that used to demand hours of billable time. By October, Thomson Reuters had launched agentic products across legal research, tax preparation, audit functions, and corporate workflows, powered partly by custom models from OpenAI.
The bet reflects a broader, quieter transformation in enterprise AI. Generative assistants that wait for careful prompting? That's already starting to feel dated. The new frontier involves autonomous agents capable of handling end-to-end assignments. Whether this shift delivers on its considerable promise is another question entirely—Gartner warned in June that more than 40% of agentic AI projects could be abandoned by 2027. But Thomson Reuters is moving with unusual speed: five major product launches in five months.
David Wong, the company's Chief Product Officer, put it bluntly at the June launch event: "Agentic AI isn't a marketing buzzword. It's a new blueprint for how complex work gets done."
Perhaps more than he intended, that statement captures both the ambition and the uncertainty.
From Prompt Engineering to Outcome Engineering
The difference between what Thomson Reuters is building and a standard AI assistant comes down to delegation. Traditional generative AI demands that users craft prompts, review outputs step by step, and essentially guide the system through each stage of a task. It's collaborative, sure, but labor-intensive.
Agentic AI takes a different tack. Users describe an outcome—draft a tax memo analyzing a client's situation, research case law on a narrow regulatory issue, comb through client files for compliance gaps—and the system does the rest. It plans a multi-step process, executes it, adjusts course if needed, and delivers a final work product.
CoCounsel for Tax, Audit & Accounting, the first agentic product Thomson Reuters shipped last June, automates client file review, memo drafting, and compliance checks by weaving together firm knowledge bases, Checkpoint tax research, IRS code, and internal documents into a single AI-guided workspace. Ready to Review, announced in July and entering general availability this fall for a subset of 1040 returns, goes even further. It drafts tax returns, adapts to feedback from the GoSystem Tax Engine, and resolves diagnostics on its own.
At least in theory. Early adopters have reported significant time savings—BLISS 1041's CIO noted at the June launch that state filing code comparisons used to consume "half a week" per jurisdiction. With CoCounsel templates and agentic research, the work now takes under an hour.
Still, there's a gap between pilot programs and scaled adoption across an industry. And that gap is where most enterprise AI initiatives go to die.
OpenAI's Custom Models and the Multi-Model Hedge
Thomson Reuters built its agentic platform on multiple large language models, but OpenAI occupies a central role. Last November, the company announced it was testing a custom version of OpenAI's o1-mini model for CoCounsel—the first enterprise customization of that particular model. The custom build emphasizes reasoning, which matters in legal and tax work where a single error can trigger liability, missed deadlines, or worse.
Joel Hron, Thomson Reuters' Chief Technology Officer, told VentureBeat the company isn't "dogmatic" about models. The platform also integrates Anthropic's Claude, Google's Gemini 1.5 Pro, and Thomson Reuters' own fine-tuned models. It's a strategy that mirrors Microsoft's multi-model approach and reflects a broader enterprise trend: companies want the flexibility to swap models as capabilities improve and costs fluctuate.
Olivier Godement from OpenAI endorsed the partnership at Thomson Reuters' June launch, saying OpenAI was "thrilled to power" the agentic workflows being introduced. The collaboration extends beyond model access—Thomson Reuters leverages AWS SageMaker HyperPod for infrastructure and integrates CoCounsel directly into Microsoft 365, including Word, Teams, and Outlook.
The technical architecture matters less to most users than the results. But for enterprises placing multimillion-dollar bets on AI, having multiple model options provides a hedge against obsolescence. Or at least the illusion of one.
Deep Research: Slow by Design
The most technically ambitious product Thomson Reuters has shipped is Deep Research, launched in August with CoCounsel Legal. Unlike chatbots optimized for speed, Deep Research prioritizes rigor. The system generates a multi-step research plan, explains its reasoning at each stage, searches Thomson Reuters' curated corpus of over 20 billion documents, and delivers citation-backed reports.
Average query time? About 10 minutes. That's glacially slow by chatbot standards. But it's dramatically faster than the 20-hour manual research tasks it's designed to replace.
VentureBeat described Deep Research in September as a multi-agent system, meaning multiple AI models work in parallel. Each handles different aspects of the task—planning, retrieval, synthesis, citation checking. Thomson Reuters grounds the system in Westlaw and Practical Law, its flagship legal research platforms, to ensure accuracy and reduce hallucinations.
Morgan Lewis, an Am Law 100 firm, highlighted the reasoning transparency in public comments at launch. That transparency matters in professional services, where clients expect to see how conclusions were reached, not just what they are.
By October, Thomson Reuters had expanded Deep Research to Practical Law (in beta), integrated it more deeply with HighQ, its collaboration platform, and added French, German, and Japanese language support. The international rollout signals confidence, though it also spreads risk across markets with different regulatory environments and professional standards.
Tax and Accounting: Three Products in Six Weeks
If the legal rollout felt aggressive, the tax and accounting push was even faster. Thomson Reuters shipped three agentic products between June and July. CoCounsel for Tax, Audit & Accounting launched first. Ready to Advise, which automates advisory workflows, reached general availability in the U.S. in July. Ready to Review, for tax preparation, entered early access the same month, with full availability planned for fall.
Both Ready to Advise and Ready to Review are described as "AI-native"—built from the ground up around agentic workflows rather than retrofitted onto legacy software. The products connect to firm knowledge bases and Thomson Reuters' Checkpoint tax research, allowing the AI to pull jurisdiction-specific rules, adapt to firm-specific processes, and handle compliance checks without constant supervision.
The speed of the rollout raises questions. Developing enterprise-grade software typically takes quarters, if not years. Thomson Reuters accelerated the timeline through acquisitions—five in the past two years, each designed to add specific AI capabilities or infrastructure.
The company acquired Casetext in August 2023 for its CoCounsel legal assistant, originally powered by GPT-4. That established the brand and initial product foundation. It followed with Safe Sign (August 2024) for legal-specific LLMs, Materia (October 2024) for agentic tax and accounting AI, SafeSend (January 2025) for tax workflow automation, and Additive (September 2025) for AI-powered tax document processing.
The acquisition strategy is textbook enterprise playbook: buy talent and technology, integrate quickly, and beat competitors to market. Whether the integrations hold up under stress—when clients are relying on these systems during tax season or major litigation—remains an open question.
The Platform Underneath
Thomson Reuters didn't build its agentic platform on a whim. The infrastructure sits on more than 15 petabytes of data, supported by 4,500 subject matter experts and 180-plus AI engineers. The company holds ISO 42001 certification, the global standard for AI management systems, and operates what it describes as a "secure, zero-retention architecture." Client data used for queries isn't retained for model training, a critical assurance for firms handling sensitive legal and financial information.
The human-in-the-loop design is central to the company's pitch. Agents don't operate autonomously in production; they surface plans for review, flag uncertain conclusions, and provide citations so users can verify reasoning. Wong framed it this way in August: "This is where AI starts to feel less like a tool and more like a teammate."
Whether professionals want a teammate or a subordinate is an interesting cultural question. Law firms, in particular, have long operated on hierarchies where junior associates do the grunt work before progressing. If AI handles that layer, the implications for hiring, training, and career progression aren't entirely clear.
Early corporate adopters include Hutchinson and Bloodgood LLP for audit work, NRMA's in-house legal team for research and contract review (documented in a Guardian case study), and multiple law schools that gained access to CoCounsel Legal and Deep Research in September. The rollout spans five new international markets announced in March: Australia, New Zealand, Hong Kong, Japan, Southeast Asia, and the UAE. Additional language support has arrived throughout the year.
The Competition Isn't Sitting Still
Thomson Reuters isn't operating in a vacuum. LexisNexis launched Protégé, its own agentic assistant, in January, using a multi-model approach that includes OpenAI's GPT-5/4o and Anthropic's Claude. Protégé emphasizes voice interaction and integrates with Lex Machina and Lexis' document management systems.
Then there's Harvey, the startup collaborating with OpenAI on custom legal models. Harvey has been winning law firm deployments and recently secured deeper integration with LexisNexis content—a partnership that puts competitive pressure on both incumbents. When a startup and an incumbent form an alliance, it usually signals that neither feels secure.
PwC announced its "agent OS" in March, a platform designed to orchestrate multiple AI agents across enterprise workflows. That signals professional services firms themselves see agent coordination as a strategic capability, not just something they'll buy from software vendors.
The race is on, in other words, and the finish line keeps moving.
What Buyers Should Actually Know
Pricing for Thomson Reuters' agentic products hasn't been disclosed publicly, which is typical for enterprise software sold through existing relationships. The company has emphasized deployment through its established base—law firms already using Westlaw get access to CoCounsel Legal; accounting firms on Checkpoint can adopt Ready to Advise or Ready to Review.
That bundling strategy makes sense from a sales perspective, but it also means firms may face pressure to adopt agentic tools whether they're ready or not. And readiness is a real issue. Capgemini research in August found low scaled adoption and significant trust gaps among IT leaders, even as the potential financial value appears substantial.
The human oversight model mitigates some risk. But it also means these systems aren't truly autonomous—they're assistants with more initiative. Whether that's enough to justify the investment depends on how much time they actually save and how much trust firms are willing to extend.
The Reality Check
Enterprise enthusiasm for agentic AI is colliding head-on with some sobering analysis. Gartner's warning about 40% project failure rates by 2027 reflects real concerns: many so-called agentic systems are essentially chatbots with better marketing. The technology is legitimately advancing, but so is the hype, and distinguishing between the two isn't always straightforward.
Thomson Reuters' approach—grounding agents in curated, domain-specific content and maintaining human oversight—may address some of the risk. But the company is also making a massive financial and technical bet that professional services workflows are ready for delegation, not just augmentation. That's a bet on culture as much as technology.
The next 18 months will reveal whether agentic AI lives up to its billing or joins the long list of overhyped enterprise technologies that promised transformation and delivered incremental improvement. Thomson Reuters, with more than 180 AI engineers, five acquisitions, and a growing product portfolio powered by some of the most advanced models available, has placed its chips on the table.
Whether it wins depends not just on what the technology can do, but on whether professionals are willing to let it.
