How we tested
Testing ran between June 10 and June 24, 2026, on current paid tiers (or vendor-published trials). Scores are weighted toward citation trust and research depth, with security posture and total cost weighted heavily for firm and in-house buyers. Every citation returned by every tool was independently verified against the underlying authority.
Citation Trust & Hallucination Rate
Each tool ran the same 30 research questions (10 federal circuit-split questions, 10 state-specific statutory questions across five jurisdictions, and 10 recent regulatory questions from 2024–2026). We clicked through every citation the tool returned, verified the case or statute existed, verified the holding matched the tool's characterization, and recorded citations that were fabricated, misattributed, or materially mischaracterized as errors.
Research Depth & Authority Grounding
For each of the 30 research questions we scored coverage against a gold-standard answer prepared by two licensed attorneys: primary law surfaced, secondary sources cited (treatises, practice guides, Practical Law or Practical Guidance notes), jurisdictional coverage, and whether the tool flagged negative treatment via KeyCite or Shepard's.
Drafting & Document Analysis Quality
Each tool drafted the same three work products from the same fact pattern (a summary-judgment brief section, a commercial NDA redline against a counterparty draft, and a memo on a mixed-jurisdiction employment question), and reviewed the same 40-document data room for change-of-control provisions. Two reviewers scored outputs blind against an attorney-written reference on structure, accuracy, and usable first-draft quality.
Security & Data Posture
We read each vendor's trust page, security addendum, and DPA and recorded SOC 2 Type II status, ISO 27001, GDPR/HIPAA posture, whether a zero-data-retention agreement with the underlying LLM provider is published, whether customer data is used to train models, and whether the vendor supports iManage / NetDocuments / SharePoint integration for firm documents.
Total Cost of a Working Seat
We priced one attorney at each vendor's real working configuration (the tier a practicing lawyer needs to actually do the job, including any required primary-law or Practical Law bundle, on a one-year term) and recorded published pricing where available, third-party-reported pricing where not, and the minimum seat count required to buy.
We ran every tool through the same research questions, contract reviews, and drafting tasks, and every citation returned was clicked through and verified. The full battery and the per-criterion marks are above; the notes here cover where the ranking turned.
Why CoCounsel Legal leads
CoCounsel Legal wins on the dimension that decides this category: the substance underneath the chat box. It’s the only legal AI that reasons from authoritative Westlaw primary law, trusted Practical Law guidance, and the customer’s own knowledge, developing the right legal approach, working through each step, and delivering polished work product with citations. Its agentic Deep Research feature, in our testing, planned and executed multi-step research the way an attorney would, using features like Deep Research, which emulates the workflow of a seasoned legal researcher, formulating and executing a multistep plan to cover every base of your legal questions. Every citation was linked and one click away for verification, and the reasoning path was visible as the model worked.
The trade-off is cost structure. CoCounsel by Thomson Reuters has 4 pricing plans ranging from $104 to $639/user/month. The CoCounsel Essentials plan requires contacting sales for a custom quote and is designed for solo and small firms wanting CoCounsel’s AI without Westlaw research, and the Westlaw Advantage with CoCounsel Essentials plan costs $639/user/month, best for litigators who need Westlaw primary-law research plus CoCounsel AI. For firms already inside the Thomson Reuters ecosystem, that math works. For firms that would need to buy Westlaw from scratch to unlock it, the answer is likely different.
When to choose Lexis+ with Protégé instead
Lexis+ with Protégé is the tool we recommend when citation trust and secondary-source depth are the decisive criteria. As of February 2026, Lexis+ AI has been renamed Lexis+ with Protégé, and the platform has evolved into a fully legal AI workflow solution for legal drafting, research, and analysis powered by the LexisNexis Protégé AI assistant. The mechanical advantage is Shepard’s: Lexis+ with Protégé is designed to provide verifiable legal answers using authoritative content, linked citations, and Shepard’s citation validation, though attorneys should still review AI-generated results, confirm cited authorities, and apply independent professional judgment. That validation loop is why a Stanford study found Lexis+ AI had a 17% error rate in legal citations, compared to 34% for Westlaw’s AI, a meaningful accuracy advantage for litigation work.
The buying experience is the weakness. AI features require a separate purchase. Lexis+ with Protégé sits outside every base tier and requires a separate sales conversation, and firms subscribing to Lexis+ and expecting full AI capabilities discover the gap only after signing. And the total is not small: the full Lexis+ with Protégé bundle ranges from $128 to $494/user/month depending on scope, and for a 10-attorney firm, the full Lexis+ with Protégé bundle can add $15,000 to $59,000 per year on top of the base subscription.
When Harvey is the right call
Harvey is genuinely capable enterprise legal AI, and its Vault repository-scale analysis is the differentiator. Vault is a secure document repository and bulk analysis engine built for large-scale review across large document sets simultaneously. The customer roster is real: Harvey has achieved rapid growth, reaching $190M ARR by January 2026 and serving 1,000+ clients globally, including 50% of AmLaw 100 firms.
The price is where most buyers stop. The price reflects that positioning. Harvey charges approximately $1,000 to $1,200 per seat per month. Sales cycles run six months or longer. Implementation requires enterprise procurement, IT involvement, and often a dedicated onboarding team. For AmLaw 100 firms with technology budgets in the millions, this is a reasonable investment. For everyone else, it is not an option. And the substance is not immune to the category’s core failure mode. In one documented user account of a real matter, a lawyer asked Harvey to review the distribution waterfall in each LPA and determine whether it was whole fund or deal by deal waterfall; Harvey gave an excellent analysis, but when the lawyer reviewed the specific sections it used, the conclusion was wrong for 2 of the 3 documents. That’s the failure mode any buyer should assume until proven otherwise.
When Spellbook is the answer
Spellbook is a focused tool that does one thing well. It’s an AI contract drafting, review, and negotiation tool for lawyers and legal teams. It works inside Microsoft Word and helps users review contracts, draft clauses, ask questions about agreements, compare terms to market standards, and apply reusable playbooks. Spellbook is designed mainly for transactional lawyers, in-house legal teams, and law firms handling commercial agreements. It is not primarily a case law research platform. That last sentence is the point. If the daily job is redlining commercial agreements in Word, Spellbook is faster than any general-purpose tool and cheaper than any full-stack legal AI. If the daily job is finding case law, it’s the wrong tool.
The security posture is also solid: Spellbook has industry-leading security, compliance and privacy standards, and is compliant with major international regulations such as SOC 2 (Type II), GDPR and CCPA.
What did not make the cut
Paxton is the one product in our test that we mark Not Recommended at its current published price. The capabilities are real. Paxton is an all-in-one AI legal assistant built for lawyers and legal professionals, which helps users research law, draft legal documents, analyze files, summarize materials, and generate legal work product from prompts and uploaded documents, with access to U.S. federal regulations, state laws, and case law across all 50 states. The problem is the sticker. Paxton’s own pricing page, as of June 6, 2026, lists an Individual plan at $499 per user per month, or $2,999 per user per year, with Enterprise at custom volume-based pricing. Several 2025-era third-party roundups still cite Paxton in the $49 to $99 per user per month range. We could not reconcile that gap: the current vendor page says $499/month, so either Paxton repriced significantly, the older roundups are stale, or both. At that number, Paxton is competing directly against CoCounsel Essentials and Lexis+ with Protégé, and on research depth we do not swap it for either. The exception is personal injury: the medical-chronology and medical-billing summary tooling is genuinely differentiated for that practice, and a personal-injury shop should evaluate it on that basis.
The larger point across every product on this list: legal AI does not remove the verification obligation. It moves it. The tool that produces the sharpest-looking answer is not the tool that produces the most accurate one, and the number a vendor prints on a marketing page is rarely the number a firm pays. Read the citations. Read the contract.
