For a corporate legal department, the right legal AI is the one with the right content library and citation check behind it. The model underneath changes every quarter; the library doesn't. If your research lives in Westlaw, buy CoCounsel for research. If your AI spend is mostly contracts and diligence, buy Legora, the only one that lets you meter agent work by project and cap the spend. Harvey loses for most GCs, not on product (it posted the best results in the only independent multi-vendor test) but on buyer terms: no published price, case law through a separate LexisNexis integration, and a model stack in flux. Whichever you buy, a lawyer still signs the redline, the task every tool struggled with in that test.
| CoCounsel Legal (Thomson Reuters) | Legora | Harvey | |
|---|---|---|---|
| Pick it if | Your research runs on Westlaw | Your AI spend is contracts and diligence | You are an Am Law firm with KM staff and a Lexis subscription |
| Grounded on | Westlaw + Practical Law | Own US primary-law corpus (since Aug 2026) + Wolters Kluwer legislative data | LexisNexis case law + Shepard's, via a separate alliance |
| Citator | Westlaw's own | AI-native citator in limited beta, GA Q4 2026 | Shepard's, inside Harvey via LexisNexis |
| Models | Anthropic, OpenAI, Google + its own Qwen-derived "Thomson" model | Anthropic Claude (per Anthropic's case study) | Multi-model; own Kimi K3-based Tenet; Opus for the hardest tasks |
| Price (checked 27 Sep 2026) | Online configurator, new customers only; 1/2/3-year terms | Not published; seats, plus metered credits for Agent Pro | Not published; contact sales |
| Do NOT pick it if | You are a Lexis shop or won't sign multi-year | You need US case-law research with a citator today | You want a published price or a metered option |
| Verdict | Best for research | Best for most legal departments | Strongest product, worst buyer terms for a GC |
Why Does the Content Library Matter More Than the Model?
The content library matters more because all three vendors have changed or added models in the past year, and none of those changes needed your consent. Thomson Reuters launched its own model in August. It is a Qwen derivative, and it became the default in CoCounsel's Tabular Analysis. CoCounsel's model providers had been Anthropic, OpenAI and Google when it passed one million users in February. Harvey post-trained Tenet on Moonshot's Kimi K3 in August, and its multi-model design post says it can route work to another model "without disrupting the user's workflow." Legora builds on Claude, according to Anthropic's customer case study. We covered the provenance side of those swaps and the margin pressure that drove Harvey's.
What does not move is the library an answer is checked against. Grounding means the tool retrieves the authority from a curated corpus and cites it, rather than generating it from the model's memory. When a model invents a case, the citator catches it or it does not. That depends on who owns the reporters.
- CoCounsel says it is "grounded in trusted Westlaw and Practical Law content" on its product page. No rival can license that corpus.
- Harvey got its case law through a strategic alliance with LexisNexis announced in June 2025. It brings LexisNexis primary law and Shepard's Citations into the Harvey workspace. That is excellent content, but it is another company's content, reached through a partnership.
- Legora put its own US primary-law corpus live on 5 August 2026. It covers federal appellate case law from 1924, state appellate from 1950 and the Supreme Court from 1790, with legislative data from Wolters Kluwer. Its citator announcement on 14 September is unusually candid about the gap: "In the US, access to citable case law is constrained by the companies that own the official reporters and won't sell a digital feed." The citator is in limited beta, with general availability planned for Q4 2026.
That last sentence is the whole US research market in one line. Two companies own the reporters. One of them sells CoCounsel, and the other rents its content to Harvey.
How Do They Handle Hallucinated Case Law?
None of the three is hallucination-free, and the only peer-reviewed measurement of this category found the grounded tools still invented or mis-cited law a sixth to a third of the time. Stanford RegLab's study, published in the Journal of Empirical Legal Studies in 2025, found hallucination rates of about 17% for Lexis+ AI, 33% for Westlaw AI-Assisted Research, and 43% for GPT-4 with no retrieval. Those were 2024 versions of the Thomson Reuters and LexisNexis research products, the ancestors of what now sits under CoCounsel and inside Harvey. Treat the numbers as proof that retrieval narrows the problem without closing it, not as today's rates.
The cost of the residual is public. Damien Charlotin's AI Hallucination Cases database listed 2,079 court decisions involving fabricated or AI-mangled citations when checked on 27 September 2026. Catching them is hard even for machines. A June 2026 paper, Who Checks the Citations?, found the best detector (GPT-5 in an agentic loop) reached 84.4% recall and a 55.0% F1 score, averaging 15.3 steps per excerpt, and noted that "all models struggle with subtle error categories."
So compare how each tool puts a checkable authority in front of the reviewing lawyer:
- CoCounsel checks answers against Westlaw, the same citator your associates already trust. This is its real edge. It also means a hallucination that slips through came from the vendor whose citator you would have used anyway.
- Harvey routes case-law questions to LexisNexis Protégé and Shepard's inside the workspace. For document work, it has rebuilt its review tables around sentence-level citations and visible reasoning, which makes a cell auditable instead of taken on faith.
- Legora has the corpus but not yet a GA citator. Until Q4, its US case-law answers need checking in whatever citator you already license. That is the concrete reason not to make it your research tool this year.
Which One Actually Performs Best?
On the one independent multi-tool benchmark, Harvey performed best, which is why calling it the loser needs defending. The Vals Legal AI Report of 27 February 2025 tested four tools against a lawyer control group on seven tasks. Harvey opted into six, took the top AI score on five, and beat the lawyer baseline on four. CoCounsel entered four tasks and took the top score on document summarization (77.2%), with 89.6% on document Q&A. Legora was not among the tools tested.
The same report found the AI tools struggled with redlining and EDGAR research. Redlining is exactly the task a GC most wants to delegate. That is the title of this piece in one data point: document Q&A and summarization are ready for delegation with review, and the redline is not.
Two caveats keep this from settling the question. The study is 19 months old, and all three products have changed models since. And Vals' later legal research study, from October 2025, found "little differentiation" in accuracy between specialist legal tools and a generalist one. The specialists pulled ahead on authoritativeness. Accuracy converges fast; sourcing does not. Run your own matched test before you trust anyone's leaderboard, and size it properly. Our Astra for Law analysis shows how a vendor number can answer a narrower question than the one you are asking.
Does Using Them Waive Privilege?
Not on the current record, as long as it is an enterprise tool used at a lawyer's direction. The first ruling on the question turned on exactly those two facts. On 10 February 2026, Judge Jed Rakoff held in United States v. Heppner that 31 documents a defendant generated with consumer Claude were protected by neither privilege nor work product. The court pointed to a privacy policy that allowed training on inputs, and to the fact that no lawyer directed the work. The ruling left open whether a lawyer's own use of a non-public tool is protected. Legora published a response arguing that enterprise controls and attorney direction distinguish it. That is a reasonable reading, but it is a vendor's reading, not a holding.
All three vendors make the commitments Heppner found missing:
- Harvey's security page commits that "we don't use inputs, outputs, or uploaded documents to train underlying models." It adds zero data retention by model providers, EU/Switzerland/Australia processing options, synced ethical walls, and SOC 2 Type II, ISO 27001 and ISO 42001.
- Legora's security page says it "will not use your data to train or fine tune any AI models." It deletes customer data at contract end, offers EU and US hosting, and holds ISO 42001, ISO 27001 and SOC 2 Type 2.
- Thomson Reuters says CoCounsel's zero data retention policies mean "your queries aren't stored or used for training."
On paper, privilege posture does not separate them. What separates them is what the paper leaves out: which model processed a given matter, and whether you are told when that changes. Harvey's ethical-wall sync is the most useful control for a department that walls off M&A or investigations work.
Who Reviews the Output, and What Does It Replace?
A licensed lawyer reviews every output that leaves the building. ABA Formal Opinion 512 of July 2024 applies the existing duties of competence, confidentiality, communication, supervision, candor and reasonable fees to AI use. None of these tools changes who is accountable. They change how long the review takes and what the reviewer is looking at.
That makes the economics straightforward to set up and easy to get wrong. To compare like for like, use one workload: a 15-lawyer corporate legal department that reviews about 600 commercial contracts a year, runs diligence on two or three acquisitions, and does case-law research a few times a month.
Harvey and Legora publish no price at all. CoCounsel publishes one only through a configurator that asks for your sector, attorney count, jurisdiction and 1-, 2- or 3-year term. It is open to new customers only, and existing customers must call sales. So the only honest cost model is a break-even:
- Brightflag's billing data put the Am Law 100 blended rate at $961 an hour for the first nine months of 2023. Thomson Reuters Institute's index shows worked rates up 7.1% year on year in Q2 2026.
- At roughly $1,000 an hour, every $1,000 of monthly seat cost has to displace about one outside-counsel hour a month to pay for itself.
- Only work that would have gone out counts. Time your own lawyers save is capacity, not cash. It is real, but it does not reduce the invoice your CFO reads.
This is where Legora's June change matters for this buyer. On 23 June 2026 it introduced consumption-based pricing for Agent Pro, where "the cost can be attributed to the project that drove it," with real-time dashboards and spending controls, a move Law360 reported as a shift away from flat subscriptions. For a department that already tracks outside-counsel spend by matter, a per-project AI line item makes the displacement argument checkable. Legora has published no credit price, so negotiate a hard monthly cap before you sign. Our agentic pricing comparison explains why.
Thomson Reuters' own 2026 Future of Professionals report is a useful warning: 91% of professionals said their organisations fall short of AI's potential value. Buying a seat is the easy part. Proving it displaced an invoice is the part most teams skip.
Who Should Not Buy Each One?
Don't buy Harvey if you are a corporate legal department that wants a published price, a metered option, or case law without a second vendor relationship. Harvey closed a $550 million round at a $15.5 billion valuation this month with more than 3,000 customer organisations, so it is not going anywhere. But its model now runs through a router that decides which of your matters get the frontier model, and nothing you receive tells you where that line sits. What changes the answer: you are an Am Law firm, you already license Lexis, and you have KM staff to run your own evaluations. Then Harvey's benchmark lead is worth the terms.
Don't buy CoCounsel if your lawyers research on Lexis, or if you will not commit to a multi-year bundle. The configurator itself says longer terms "typically offer greater savings," which is the lock-in stated as a discount. Also check your tenant's default model: the Qwen-derived Thomson model shipped as the Tabular Analysis default, and administrators can switch it. What changes the answer: nothing, if you are on Westlaw. For research, it is the default choice.
Don't buy Legora as your US case-law research tool before its citator reaches general availability. Don't buy Agent Pro without a written spending cap. Legora raised $550 million at a $5.55 billion valuation in March and is moving fast, which also means the product you evaluate in October is not the one you renew. What changes the answer: if the Q4 citator ships and holds up on your own matter set, Legora becomes a credible single tool for departments without a Westlaw habit.
The Decision Criteria That Predict Regret
The regret in this category rarely comes from the model. It comes from the contract terms and the content you did not check. Four questions predict it:
- Which citator will your lawyers use to check the output? Buy the tool whose grounding matches that citator. A mismatch means every answer gets checked twice.
- Will the vendor tell you, in writing, which model served each feature, and give notice before it changes? None of the three offers this by default. Ask for it as a clause.
- Can you attribute AI spend to a matter? If not, you cannot show it displaced outside counsel, and the renewal becomes a faith argument.
- Who reviews redlines, and is that review time-boxed? The independent data says redlining is where these tools are weakest. Budget the lawyer, not just the seat.
This Week:
- Ask each vendor for a written quote on your defined workload. For CoCounsel, run the configurator with your real attorney count and save the result.
- Send one question to all three: for each feature we would use, name the model, its base, and the notice you give before it changes.
This Month:
- Pull 50 closed matters: contracts whose final redline you have, and research questions whose controlling authority you know. Run them through two finalists with the same reviewer.
- Score citations in their own column: real, controlling, correctly characterised. Score redlines on whether a partner would send them unedited.
Before You Sign:
- Put model-change notice, a usage allowance or hard spending cap, and data deletion at termination into the order form, not the policy page.
- Set a six-month checkpoint that compares outside-counsel hours on the targeted work types against the prior year. If they have not moved, the tool is capacity, not savings. Renegotiate on that basis.
The Bottom Line
Legal AI is repeating what happened with legal research databases a generation ago. The interface was never the moat; the reporters were. The models under all three tools are converging and changing faster than any contract tracks. The corpus, the citator and the review workflow are what you are actually buying. Pick on those, contract for the model, and keep a lawyer on the redline.
The model is rented. The library is owned. Buy the library.
Continue Reading
- CoCounsel's New Model Runs on Qwen. Go Read the Card.
- Harvey Swapped Models After Agents Sank Its Margin to -50%
- Astra for Law Scored 54% on a Set Vals Never Publishes
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
- Guardrails' Hub Died Aug 25. Harvey Bought the Team.
- $1.2B Legal AI: Why Blackstone Pays for Outcomes, Not Hours
- Only 3 of 36 Model Gaps Were Real. Size Your Eval Set.