Thomson Reuters is taking a counterintuitive approach to legal artificial intelligence: making it work for minutes rather than answering immediately. Its Deep Research capability, associated with Westlaw and CoCounsel, is designed to investigate a legal question, follow authorities, compare competing arguments and return a citation-linked report for attorney review.
VentureBeat reported on September 15, 2025, that a typical run took about 10 minutes for work lawyers might otherwise spend 10–20 hours researching. That is a reported use-case comparison, not an independently validated benchmark showing that every 20-hour assignment can be completed in 10 minutes.
What Thomson Reuters actually built
Deep Research is a legal-research workflow, not simply a chat window with a larger prompt. The system is intended to:
- Interpret the legal question and relevant facts.
- Break the issue into researchable sub-questions.
- Form lines of inquiry or working hypotheses.
- Search relevant primary and secondary authority.
- Follow citations and related decisions.
- Compare favorable, adverse and distinguishable authorities.
- Revise the research plan when new material raises another issue.
- Produce a structured answer with links or citations to the underlying sources.
VentureBeat described the process as iterative: the system updates its plan and follows “breadcrumbs” from one authority to another. Thomson Reuters has not publicly documented a fixed internal diagram or a specific number of production agents, so “multi-agent” should be understood as a description of coordinated processes and roles, not a disclosed roster of autonomous legal experts.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →#1 Best Overall
The company’s broader explanation of agentic AI describes systems that plan, reason, act, react and remain within professional workflows, with people guiding judgment and validating results. See Thomson Reuters’ agentic-AI announcement.
Why this is different from an ordinary chatbot or basic RAG
A conventional retrieval-augmented-generation (RAG) application retrieves documents, places excerpts in a model’s context and generates an answer. That can be useful, but the user often has to decide what to search next, chase citations and test the answer’s limitations.
| Feature | General chatbot | Deep Research-style legal workflow |
|---|---|---|
| Main objective | Fast conversational response | Thorough, source-linked research report |
| Data | Broad model knowledge or web results | Licensed, curated legal content |
| Process | Usually one turn or a short chain | Planned, iterative investigation |
| Analysis | Summarizes retrieved material | Compares authorities and revises searches |
| Verification | User must locate and check sources | Links to source material and legal research tools |
| Remaining risk | Fabrication, omission and stale knowledge | Misinterpretation, omission and model error still require review |
Deep Research still relies on retrieval and language models. The difference is the surrounding loop of planning, tool use, evaluation and revision. Thomson Reuters has described a multi-model strategy involving OpenAI, Anthropic, Google, frontier models and experiments with open-source models, but available reporting does not establish which model handles each production stage.
How a representative legal question would be handled
VentureBeat used a trade-secret example: whether a customer list qualifies as a trade secret under a particular fact pattern and jurisdiction. A useful investigation would need to:
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11- Identify the governing statute and applicable jurisdiction.
- Find cases addressing customer lists and comparable facts.
- Determine which facts courts treated as significant.
- Locate decisions granting and denying protection.
- Compare the client’s facts with both lines of authority.
- Surface contrary, limiting or newer decisions.
- Provide citations that the lawyer can inspect.
The value is not merely finding the statute. It is helping the lawyer develop an argument by analogy while exposing unfavorable precedent that needs to be distinguished.
Why Westlaw’s content is the moat
The model is only one part of Thomson Reuters’ proposition. The company says its platform draws on more than 20 billion documents, over 15 petabytes of data and more than 500 trusted content assets, including opinions, statutes, administrative rulings, treatises, books and practitioner commentary. It also cites 4,500 subject-matter experts and more than 180 AI engineers. These are company-reported platform figures; they do not mean every query searches every document or that every document has equal relevance.
The practical advantages of a legal database are authority and context:
- Editorial classification: material is organized by jurisdiction, subject and authority type.
- Currency: updates, treatment information and invalidation signals can identify changes in the law.
- Primary and secondary sources: researchers can move from a rule to commentary and practical guidance.
- Auditability: source links let attorneys inspect the text rather than trust a fluent paragraph.
- Workflow integration: research can connect with drafting and document-analysis tools.
That is a vertically integrated information product, not simply an LLM interface.
Free tools Windows power users keep installed
One-click scans. No signup required.
What “20 hours to 10 minutes” does—and does not—prove
According to the VentureBeat interview, lawyers may spend approximately 10–20 hours on complex research, while Deep Research’s default run was described as taking about 10 minutes. Three- and seven-minute options were also discussed, and a 20-minute mode was reportedly being explored at the time. Current availability of those specific modes should be confirmed in the product interface.
Thomson Reuters did not publish definitive performance data proving that all comparable matters fall from 20 hours to 10 minutes. The headline should therefore be read as a typical or illustrative workflow comparison. It is not evidence of a universal 95% time reduction, and a faster first report may still require substantial attorney checking and revision.
Rank #3
Fewer hallucinations does not mean no legal errors
Legal-AI reliability has at least three separate dimensions:
Fabricated authority
A system invents a case, statute, quotation or pinpoint citation. Curated content and linked citations can make this risk easier to detect and may reduce it.
Incorrect interpretation
The cited authority exists, but the system misstates its holding, scope, procedural posture or relevance.
Incomplete research
The citations are genuine, yet the report omits controlling, adverse, newer or jurisdictionally important authority.
An earlier independent study reported hallucination rates of 17% to 33% for the proprietary legal-AI systems it tested, depending on the product and task. That study evaluated earlier versions and is not a direct measurement of Deep Research in 2025–2026. It is nevertheless a warning against “hallucination-free” claims. See the study on arXiv.
What human oversight means
Human-in-the-loop does not mean an attorney approves every intermediate search. It means the professional remains responsible for:
- Framing the question and supplying jurisdiction, dates, posture and material facts.
- Checking the cited source, quotation and treatment history.
- Testing the report against adverse authority and the client’s evidence.
- Applying confidentiality, privilege, conflicts and professional-conduct rules.
- Making the final legal judgment and filing or client-communication decision.
Deep Research is an accelerator for research, not independent legal advice.
Where Deep Research sits in the current product portfolio
Product names matter because Deep Research is not included in every CoCounsel package.
| Offering | What the current pages indicate | Deep Research |
|---|---|---|
| CoCounsel Legal | Broad legal AI offering spanning research, drafting, document analysis and workflow integrations | Included in the package positioning |
| Westlaw Advantage with CoCounsel Essentials | Westlaw research combined with CoCounsel capabilities, including litigation-oriented tools | Included |
| CoCounsel Essentials | Document analysis, drafting, playbooks, knowledge search and Microsoft Word workflows | Not included in the current comparison table |
| Practical Law Dynamic Tool Set with CoCounsel Essentials | More focused on transactional and advisory work | Check the selected plan and feature table |
See CoCounsel Legal plans and Westlaw plans and pricing. Firms with more than 10 attorneys are directed to contact sales. Public pages emphasize “view pricing” and demo flows rather than a single universal price, so buyers should model the full Westlaw, Practical Law and CoCounsel commitment.
Who should evaluate it
Likely fit
- Firms already invested in Westlaw or Practical Law.
- Litigation teams handling complex, jurisdiction-specific questions.
- Legal departments that need traceable authority and an auditable workflow.
- Organizations willing to run a controlled pilot and require attorney review.
Likely poor fit
- Individuals seeking a cheap, occasional general-purpose chatbot.
- Teams whose priority is general writing rather than legal research.
- Firms whose key sources fall outside Thomson Reuters’ collection.
- Buyers without an AI-governance process or willingness to verify citations.
How it compares with alternatives
Lexis+ with Protégé
Lexis+ AI was renamed Lexis+ with Protégé in February 2026. Its offering combines legal research, drafting, summarization, document analysis and workflow automation using LexisNexis content and Shepard’s citation validation. See Lexis+ with Protégé and the core Lexis+ page. The decisive comparison is usually coverage, editorial tools, existing contracts, integrations and performance on the firm’s own matters—not a generic claim that one model is smarter.
Recommended Free Tools
Harvey
Harvey is an enterprise legal-AI alternative with emphasis on custom workflows and organization-specific deployment. Buyers may need a separate legal-research data strategy if they require the depth, editorial structure and citator functionality associated with Westlaw or LexisNexis. Public pricing is not a dependable apples-to-apples comparison.
General-purpose AI
General tools can help with brainstorming, plain-language explanations and nonconfidential draft restructuring. They are not substitutes for current jurisdiction-specific authority, treatment history, source verification and professional auditability. Do not upload privileged or confidential material without reviewing the enterprise contract, data controls and firm policy.
A practical buyer’s evaluation checklist
- Test authority coverage: use the jurisdictions, agencies, regulations and secondary sources your lawyers actually need.
- Use difficult matters: include adverse authority, conflicting precedent, statutory interpretation and misleading facts.
- Score the output: record citation correctness, completeness, treatment of adverse cases, factual fidelity and attorney edits.
- Measure the whole workflow: compare time to usable work product, not time to a polished first draft.
- Review security terms: check retention, deletion, training use, tenant isolation, permissions and ethical walls.
- Model economics: include seat utilization, existing subscriptions, implementation and the review time that remains.
Thomson Reuters says its platform uses secure, zero-retention architecture and does not repurpose customer data to train third-party models. Those are vendor claims that should be verified against the contract, data-processing terms and security documentation.
What changed after the 2025 launch story
Thomson Reuters has since positioned agentic AI across legal, tax, audit, accounting, risk and compliance workflows. In a February 24, 2026 announcement, it said one million professionals had chosen CoCounsel across 107 countries and territories. That is a company-reported adoption milestone, not an independent measure of active usage, customer satisfaction or answer quality; see the company announcement.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesThe durable idea is not that slower reasoning is always better. Longer execution can still accumulate irrelevant sources, overfit to an early hypothesis or produce a longer report without better legal judgment. The attempted advantage is a controlled loop around authoritative content, editorial metadata, multiple models, tools, citations and professional review.
The Bottom Line
Thomson Reuters’ Deep Research is best understood as an orchestrated legal-research process, not an “anti-LLM.” Its reported 10-minute run can be dramatically faster than a 10–20-hour manual assignment, but that comparison is illustrative rather than independently proven. The defensible advantage is the combination of Westlaw’s curated content, iterative investigation, linked citations and workflow integration—while attorneys remain responsible for checking sources and conclusions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
