CoCounsel (Thomson Reuters) Review for Lawyers
CoCounsel started life as Casetext's GPT-4-powered assistant, was acquired by Thomson Reuters in 2023, and has since been folded into the broader Westlaw and Practical Law stack. It now sits inside Thomson Reuters' AI platform alongside CoCounsel Legal for law firms and CoCounsel for corporate legal departments. This review is aimed at practicing lawyers deciding whether to license it — and, equally important, whether the confidentiality posture holds up under the standards most bar authorities are now applying to generative AI.
What CoCounsel (Thomson Reuters) Actually Does
CoCounsel is positioned as an AI legal assistant rather than a single-purpose tool. In current form it clusters around a handful of skills that matter to litigation and transactional lawyers:
- Legal research assistance grounded in Westlaw content, with citations back into the Westlaw corpus rather than the open web.
- Document review — uploading a set of contracts or discovery documents and asking questions across the set, extracting terms, or flagging deviations from a playbook.
- Contract analysis and drafting against Practical Law templates and firm-supplied precedents.
- Deposition and transcript summarization, timeline building, and issue extraction.
- Deep Research, a longer-running agentic workflow that plans a research task, runs multiple Westlaw queries, and returns a structured memo.
The important architectural point is that CoCounsel is a wrapper over frontier models (historically OpenAI, with Anthropic and others added over time) plus Thomson Reuters' own retrieval layer over Westlaw and Practical Law. That means answers are usually grounded in a proprietary, curated corpus — which materially reduces the raw hallucination risk that plagues general-purpose chatbots when they are asked legal questions. It does not eliminate it; TR itself warns users to verify outputs, and independent testing (notably the 2024 Stanford HAI/RegLab study on legal RAG systems) found that even retrieval-augmented legal AI products still produce meaningful error rates.
The Confidentiality Lens
This is the section that matters most for lawyers, because everything downstream — Rule 1.6, Rule 1.1, engagement letter language, whether you can use it on a matter subject to a protective order — turns on where the data goes.
Based on Thomson Reuters' published materials and its CoCounsel data security and privacy page, the relevant architectural facts are:
- Cloud processing, not on-device. Documents you upload and prompts you enter are transmitted to Thomson Reuters' cloud environment and, from there, to the underlying model providers via enterprise API arrangements.
- No training on customer data. Thomson Reuters states that customer inputs and outputs are not used to train the underlying foundation models, and that its contracts with model providers prohibit that use.
- Enterprise API tenancy with OpenAI/Anthropic. Data is processed under zero-retention or short-retention enterprise terms with the model vendors, so prompts are not retained by the model provider for training or product improvement.
- Retention at Thomson Reuters. Matter data, uploaded documents, and chat history are retained inside the customer's CoCounsel tenant so lawyers can go back to prior work. Retention length and deletion controls are set by the customer's administrator.
- DPA available. Thomson Reuters offers a Data Processing Addendum for GDPR/UK GDPR customers and a standard set of enterprise security certifications (SOC 2 Type II, ISO 27001) referenced on its privacy program pages.
Now the honest part: this is a materially better posture than pasting client facts into a consumer ChatGPT window, and it is roughly on par with other enterprise legal-AI offerings. It is not, however, the same thing as local processing. A subpoena served on Thomson Reuters could in principle reach content stored in the customer tenant — prompts, uploaded documents, and generated outputs — subject to whatever objections TR and the customer raise. The same is structurally true of any cloud legal-AI product, and it is a reason to think carefully about what you upload for highly sensitive matters (internal investigations, trade-secret work, matters under seal). The lesson of US v. Heppner (S.D.N.Y. Feb. 2026), where Judge Rakoff held that a litigant's chats with a public AI platform were not privileged, is narrower than headlines suggested — it turned on the public/consumer nature of the tool — but it is a useful reminder that the vendor's terms and architecture do the work, not the label "AI."
Disclosure: Basil, which publishes this review library, competes with Thomson Reuters in the meeting-notetaker category via Basil for Law. CoCounsel is a research and drafting assistant rather than a notetaker, so we are reviewing it on its own merits; Basil is not a substitute for what CoCounsel does, though the two can sit side by side in a practice where meetings are captured on-device and research runs in the cloud.
Ethics-Opinion Fit (ABA 512)
ABA Formal Opinion 512 (July 2024) is now the default lens most bar counsel apply to generative AI in practice. It does not bless or ban any specific product; it requires lawyers to understand the tool, protect confidentiality under Rule 1.6, supervise outputs under Rules 5.1 and 5.3, communicate with clients where appropriate, and bill reasonably under Rule 1.5. CoCounsel maps onto that framework reasonably well:
- Competence (1.1) and understanding the tool. Thomson Reuters publishes documentation of what CoCounsel does and does not do, and offers CLE-style training. That helps satisfy the "reasonable understanding" component, provided you actually read it.
- Confidentiality (1.6). Enterprise tenancy, no-training terms, and a DPA are the baseline the opinion contemplates. They are architectural risk reduction, not a guarantee.
- Supervision (5.1/5.3). CoCounsel outputs still need to be checked. The Stanford legal-RAG work and TR's own disclaimers make clear that grounding reduces but does not eliminate error. Cite-checking is not optional.
- Client communication. Opinion 512 stops short of a blanket disclosure requirement, but a growing number of state opinions — including NYC Bar Formal Opinion 2025-6 — push toward disclosure when AI is doing substantive work. Engagement letters should reflect that.
- Billing (1.5). If CoCounsel compresses a four-hour research task into forty minutes, you generally cannot bill the four hours. That is not a CoCounsel problem; it is a workflow one.
Also worth reading in this vicinity: West Technology Group v. Sundstrom (D. Conn. 2024), which continues to be cited for the proposition that sanctions follow AI-hallucinated citations regardless of the tool used. CoCounsel's grounding in Westlaw reduces that risk relative to open-web chatbots, but the professional obligation to verify each citation before filing is unchanged.
How It Actually Feels To Use
In day-to-day use, the strongest parts of CoCounsel are the ones tied most closely to Westlaw. Research questions that would otherwise require assembling a KeyCite pass, a secondary-source scan, and a jurisdictional survey can be posed in a single prompt and returned as a structured memo with linked authority. For lawyers already paying for Westlaw, that integration is genuine value rather than a marketing bullet.
Document review across a folder of contracts is competent and fast, and the ability to run a playbook — "flag any indemnity that survives termination for more than two years" — across dozens of agreements is a real time-saver on diligence. Deposition summarization is useful for first-pass work, though most litigators will still want a human associate to produce the version that goes into a brief.
The weaker edges are the ones that show up in every large-model product: occasional overconfident summaries, occasional missed nuance in state-specific doctrine, and the ever-present need to open the underlying case rather than trust the paraphrase. None of that is disqualifying; it is the reason supervision under 5.1 exists.
Pricing and Who It's For
Thomson Reuters does not publish a public list price for CoCounsel; pricing is quoted per user, typically bundled with Westlaw and Practical Law entitlements, and negotiated at the firm or department level. The current pricing and packaging options are on the CoCounsel product page. In practice, CoCounsel makes the most sense for:
- Firms and in-house teams already committed to Westlaw, where the integration is the point.
- Litigation-heavy practices that will actually use document review and deposition tooling, not just the chat interface.
- Transactional groups running repeated diligence or playbook work at volume.
It is a heavier lift, and probably poor value, for solo practitioners who do not already subscribe to Westlaw and who mostly need drafting help; a lighter-weight assistant plus a disciplined intake process may serve better at that scale.
Verdict
CoCounsel is a serious, enterprise-grade legal AI product with a confidentiality posture that stands up to the framework in ABA Formal Opinion 512, provided the firm configures it responsibly and supervises outputs. It is strongest where it leans on Westlaw and Practical Law, and it is not a substitute for local, on-device workflows where those are called for.
| Pros | Cons |
|---|---|
| Grounded in Westlaw and Practical Law, reducing raw hallucination risk | Cloud processing; a subpoena to the vendor could in principle reach tenant data |
| Enterprise no-training terms with model providers; DPA available | No published list pricing; requires negotiation and usually a Westlaw commitment |
| Genuinely useful document review, contract analysis, and deposition tooling | Outputs still require careful cite-checking and human supervision |
| SOC 2 Type II and ISO 27001 program controls | Heavier than solos or very small firms typically need |
This review is for information only and is not legal advice.
Frequently asked questions
Does Thomson Reuters train its AI models on CoCounsel customer data?
Thomson Reuters states that customer inputs and outputs in CoCounsel are not used to train the underlying foundation models, and that its enterprise contracts with model providers such as OpenAI and Anthropic prohibit that use. Retention of matter data inside the customer tenant is a separate question and is governed by the customer's administrator settings and the applicable agreement.
Is CoCounsel processed on-device or in the cloud?
CoCounsel is a cloud product. Prompts and uploaded documents are transmitted to Thomson Reuters' cloud environment and to model providers under enterprise API terms. It is not a local or on-device tool, which is a meaningful architectural difference from products designed for highly sensitive matters where data cannot leave the device.
Does CoCounsel satisfy ABA Formal Opinion 512?
Opinion 512 does not bless or ban specific products; it requires lawyers to understand the tool, protect confidentiality, supervise outputs, communicate with clients, and bill reasonably. CoCounsel's enterprise data terms, no-training posture, and Westlaw grounding help on confidentiality and competence, but supervision and cite-checking remain the lawyer's responsibility.
Can CoCounsel be trusted for legal citations?
CoCounsel's grounding in Westlaw reduces citation hallucinations relative to general-purpose chatbots, but independent testing of legal retrieval-augmented systems has shown non-zero error rates. Every citation should still be opened and verified before it appears in a filing, consistent with cases like West Technology Group v. Sundstrom.
How is CoCounsel priced?
Thomson Reuters does not publish a public list price. CoCounsel is quoted per user, typically bundled with Westlaw and Practical Law entitlements, and negotiated at the firm or department level. Current packaging is described on the Thomson Reuters CoCounsel product page.
Should a solo practitioner buy CoCounsel?
For solos not already committed to Westlaw, CoCounsel is often heavier and more expensive than the practice needs. Lighter-weight drafting assistants combined with disciplined intake and confidentiality practices may serve better at that scale, with CoCounsel becoming more compelling as Westlaw usage and matter volume grow.
Meeting notes with no server to subpoena
Basil transcribes and summarizes entirely on-device — privilege-safe by architecture. See Basil for Law →
This review is for information only and is not legal advice.