Lexis+ with Protege

Protégé is not a standalone product. It is the generative-AI assistant inside Lexis+, and the thing that makes it work is the LexisNexis content licence sitting underneath — primary law, secondary sources, Practical Guidance, and Shepard's citation validation. Legaltech Hub describes the product as combining exactly those elements, and Lawyerist's directory entry lists it as "an add-on product" (https://www.legaltechnologyhub.com/vendors/lexis-ai-by-lexisnexis/, https://lawyerist.com/reviews/artificial-intelligence-in-law-firms/lexis-ai-review-artificial-intelligence-for-lawyers/). Practically, this means your negotiation is about two things at once — which content packages you licence and which AI tier you sit on — and the content is usually the larger number.

The naming has changed twice in eighteen months, which makes older research confusing. It was Lexis+ AI; Protégé launched in preview in August 2024 and went generally available in January 2025 (https://www.lawnext.com/2025/08/lexisnexis-launches-protege-general-ai-expanding-the-agentic-capabilities-of-its-ai-assistant-to-general-ai-models-such-as-gpt-5.html); the USC Gould law library guide records the product being renamed from Lexis+ AI to Lexis+ with Protégé in February 2026, and warns that "older training materials and videos may still use the former name" (https://lawlibguides.usc.edu/c.php?g=1447399&p=10757499). If you are reading a review or a benchmark, check which product generation it tested.

As of the May 2026 release the platform covers: Ask, Draft, Summarize, Documents and Workflows (per the USC guide); Protégé Work, a skills-and-orchestration layer; Protégé Agentic Drafting for contracts, motions, briefs and deal documents; Protégé Workrooms for permissioned collaboration; Protégé Vault, now supporting up to 100,000 documents and accepting audio, video and image files; and Shepard's Verify Trust Markers (https://www.lawnext.com/2026/05/lexisnexis-expands-lexis-with-protege-adding-agentic-skills-collaboration-workrooms-and-customer-held-encryption-keys.html).

Three things stand up to checking.

First, accuracy relative to its direct competitor. The Stanford RegLab / HAI study — the first preregistered empirical evaluation of proprietary legal AI research tools, later published in the Journal of Empirical Legal Studies — found Lexis+ AI answered 65% of queries accurately with a 17% hallucination rate, against Westlaw AI-Assisted Research at 42% accurate and 33% hallucination. Westlaw hallucinated at roughly double the Lexis rate (https://www.lawnext.com/2024/06/in-redo-of-its-study-stanford-finds-westlaws-ai-hallucinates-at-double-the-rate-of-lexisnexis.html, https://reglab.stanford.edu/publications/hallucination-free-assessing-the-reliability-of-leading-ai-legal-research-tools/). This is a study that set out to puncture vendor marketing and did — and Lexis still came out ahead of the field it tested. The authors also noted that "both Westlaw and Lexis often offer high-quality responses."

Second, model flexibility that most legal AI vendors don't offer. Protégé General AI lets users select among frontier models from more than one provider — Claude Sonnet 4 and 4.5 from Anthropic and GPT-4o, o3, GPT-5 and GPT-5.1 from OpenAI — with a "Best Fit" mode that selects a model per task, and an agent architecture split across an Orchestrator Agent, Legal Research Agent, Web Search Agent and Customer Document Research Agent (https://www.lawnext.com/2025/12/lexisnexis-unveils-the-next-generation-of-its-protege-general-ai-callling-it-the-most-integrated-legal-ai-workflow-solution.html). For a firm that would otherwise run a shadow ChatGPT deployment alongside its research tool, consolidating that into one governed environment is a real procurement argument.

Third, enterprise security controls that are unusually concrete. Protégé BYOK gives customers their own encryption keys via AWS KMS, Azure Key Vault, Google Cloud KMS or HashiCorp Vault, and Ambrogi reports it has been "deployed in AmLaw 100 firms" (https://www.lawnext.com/2026/05/lexisnexis-expands-lexis-with-protege-adding-agentic-skills-collaboration-workrooms-and-customer-held-encryption-keys.html). Customer-held keys are a specific, auditable control rather than a security bullet point, and for firms with client outside-counsel guidelines that restrict AI processing, it is often the clause that unblocks the deal.

Ambrogi's read on the competitive position is fair and worth repeating: LexisNexis is competing on "ownership of authoritative content ... and enterprise-grade controls" rather than on "speed and product-velocity." If those are the axes you care about, this is a strong fit.

The Stanford figures above are the best independent numbers available, and they are also the only ones. They were collected in 2024 against Lexis+ AI, before Protégé, before the agent architecture, and before the current model line-up. A buyer should treat 65%/17% as a historical data point about an earlier product, not a current spec.

What has not happened since is a public re-test. LexisNexis has not published or commissioned an independent re-benchmark of the Protégé-generation product, and we could not find one from any third party. Several sites circulate a claim that current Lexis hallucination rates sit in the 4–7% range; we could not trace that to any study, dataset or primary source, and it should not be relied on.

The useful non-vendor evidence that does exist is qualitative. At a Southern California Association of Law Libraries session, librarians Mark Gediman, Cindy Guyer and Tanya Livshits ran three California/9th Circuit questions through Lexis+ AI, Westlaw Precision AI and vLex Vincent. Lexis did well on the first: it correctly identified the 10-day interlocutory-appeal window with controlling statutes cited inline, and its Shepard's integration and follow-up explanations were rated strengths. On the class-certification appeal question it missed the "death knell" doctrine — which Guyer noted was "kind of important" — and its later paragraphs became repetitive, citing the same cases again. Vincent gave what Guyer called a "great answer" on that question. The panel's overall framing was that these tools are "starting points, not ending points" (https://www.lawnext.com/2025/02/in-ai-smackdown-law-librarians-compare-legal-ai-research-platforms-finding-distinct-strengths-and-limitations.html).

The Vals Legal AI Report is the most serious attempt so far at a vendor-neutral benchmark of legal AI. LexisNexis was initially evaluated but withdrew before publication: "Lexis+AI (from LexisNexis) was initially evaluated but withdrew from the sections studied in this report" (https://www.artificiallawyer.com/2025/02/27/vals-publishes-results-of-first-legal-ai-benchmark-study/). The company gave a reason, and it is a reasonable one: "The timing didn't work for LexisNexis to participate in the Vals AI legal report. We deployed a significant product upgrade to Lexis+ AI, with our Protégé personalized AI assistant, which rendered the benchmarking analysis out-of-date and inaccurate." The published February 2025 results covered Harvey, Thomson Reuters' CoCounsel, vLex and Vecflow, with Harvey taking top scores on five of six tasks it entered.

When Vals later split legal research into its own benchmark, published October 2025, LexisNexis again did not take part — nor did Thomson Reuters. Legal IT Insider headlined the absence and reported that "although they had vendors opt into legal research as part of the first report, after it was separated out, they chose not to be included" (https://legaltechnology.com/2025/10/16/vals-ais-benchmarking-report-for-legal-research-is-out-but-the-market-leaders-are-absent/). That benchmark ran 200 US legal research questions across Alexi, Counsel Stack, Midpage and ChatGPT (https://www.lawnext.com/2025/10/vals-ais-latest-benchmark-finds-legal-and-general-ai-now-outperform-lawyers-in-legal-research-accuracy.html).

This is context rather than an accusation — the incumbents in this market have generally not submitted to third-party benchmarking, and a fast-moving product genuinely does go stale between test and publication. But the practical consequence for a buyer is real: you cannot compare Lexis+ with Protégé against Harvey, CoCounsel or Vincent on any common independent yardstick. That makes your own pilot, on your own matters, the only comparison you will get. Budget for it.

Shepard's is the strongest asset here and also the most commonly over-read. The May 2026 release added Shepard's Verify Trust Markers, which "identifies legal citations ... checks them against LexisNexis authoritative sources, and flags citations that cannot be verified as existing." Ambrogi's caveat is the important sentence for any litigator: it "is not capable of confirming that a cited authority actually supports the proposition for which it is cited — which is a common failure in AI hallucination cases" (https://www.lawnext.com/2026/05/lexisnexis-expands-lexis-with-protege-adding-agentic-skills-collaboration-workrooms-and-customer-held-encryption-keys.html). Existence-checking and support-checking are different problems, and only the first is automated.

There is a documented instance of a lawyer relying on a Lexis AI check and it not catching the problem. In Gauthier v. Goodyear Tire & Rubber Co. (E.D. Tex., 25 November 2024), Judge Marcia Crone sanctioned plaintiff's counsel $2,000, plus a CLE requirement and an obligation to send the order to the client, after a summary-judgment response cited two nonexistent cases and multiple fabricated quotations. The drafting tool was Claude, not Lexis. But the order records: "Monk further recounted that he attempted to check the content of the Response by using a feature available through Lexis AI. According to Monk, the Lexis AI feature failed to flag the issues with the Response" (https://www.courthousenews.com/wp-content/uploads/2024/11/attorney-sanctioned-for-using-ai-hallucinations.pdf).

Three caveats belong with that, and they matter. This is the attorney's own account offered at his own sanctions hearing; the court made no finding about the tool and did not verify the claim, and the sanction rested on his failure to read the cases at all — the court noted it was "unclear what legal research, if any" he did. Which feature he used, and whether correctly, is not established. And it predates Protégé and Shepard's Verify by well over a year. It is not evidence that the product is broken. It is a well-documented illustration of the gap Ambrogi identifies, and a reason to write your verification policy around reading the authority rather than around a green tick.

The USC law librarians reach the same place independently: "Verification means more than confirming that cited authorities exist; check that each authority is characterized accurately," and on Shepard's specifically, "A signal is a starting point, not a substitute for reading the citing decisions yourself." On the agentic workflows they add that "because these workflows chain multiple AI steps together, the output warrants the same verification as any other AI-generated result, at every step" (https://lawlibguides.usc.edu/c.php?g=1447399&p=10757499). That last point is worth weighing against the agentic drafting features: longer chains mean more places to check, not fewer.

LexisNexis publishes no rate card for the AI layer. Lawyerist lists pricing as "Contact Vendor" and flags "No Published Pricing" and "Expensive" as cons (https://lawyerist.com/reviews/artificial-intelligence-in-law-firms/lexis-ai-review-artificial-intelligence-for-lawyers/). None of the vendor announcements from August 2025, December 2025 or May 2026 disclose price, and none of the trade coverage of them does either.

We attempted to verify pricing directly and could not. The LexisNexis online store and small-firm pages block automated access, and we found no bar association, university consortium or public procurement document publishing an actual figure. A specific warning: several per-seat figures circulate widely — ranges around $128–$494/user/month, and a separate $125–$275 AI add-on figure. Every instance we found appeared on third-party aggregator sites that cite no source and largely repeat each other. We could not corroborate any of them against LexisNexis or a named buyer, and we are not repeating them as fact. Treat any such number you encounter, including on this site's competitors, as unsourced.

What is verifiable is that discounting exists through bar associations — Ohio State Bar and Massachusetts Bar member benefit pages reference a 15% discount on Lexis+ AI — though the underlying prices those apply to are not published either (https://www.ohiobar.org/membership/Practice-Resources/legal-research/lexis-ai/). If you are a small firm, ask your state bar before you call sales.

The practical guidance: because the AI tier and the content licence are negotiated together, ask for them priced separately, ask what happens to the AI entitlement if you drop a content module, and get the per-seat step-ups in writing at the headcounts you actually expect to reach rather than today's.

This is a comparative strength and the claims are specific enough to test in diligence. LexisNexis states it does not use customer data to tune or train its large language models, that customer prompts and documents are encrypted at rest using AWS KMS and AES-256 with TLS 1.2 in transit, and that it undergoes an annual SOC 2 Type 2 examination, alongside ISO 27001 (see the LexisNexis Trust Center, https://trust.lexisnexis.com/). These are vendor statements; the SOC 2 report itself is the thing to request under NDA rather than take on trust, and it is worth confirming the scope of that report covers the Protégé components specifically rather than only the underlying Lexis and Lexis+ platforms.

The BYOK capability described above is the more meaningful control, because it changes who holds the keys rather than describing how the vendor holds them. Independently reported as deployed in AmLaw 100 firms (https://www.lawnext.com/2026/05/lexisnexis-expands-lexis-with-protege-adding-agentic-skills-collaboration-workrooms-and-customer-held-encryption-keys.html). Protégé Workrooms adds role-based permissions and audit trails, which is the other thing outside-counsel guidelines tend to ask for.

One question to put to sales: Protégé General AI routes work to third-party frontier models from OpenAI and Anthropic, and Protégé General AI "draws from open-web sources." Establish in writing which data handling and no-training commitments follow the request out to those providers, and whether the web-search agent can transmit matter content externally.

Unusually for this category, there is no viability question. LexisNexis Legal & Professional is a division of RELX, a public company. RELX reported the division's underlying revenue growth accelerating from 9% in 2025 to 10% in the first half of 2026, with revenues of £959m up from £900m and adjusted operating profit up 13% to £208m — described as the fastest growth in its history and attributed substantially to AI products, with a claim that 90% of new business came from AI-related products (https://legaltechnology.com/lexisnexis-posts-fastest-growth-in-its-history-powered-by-ai-products/).

Read that as reassurance about continuity and support, not as evidence of product quality. Note also the corollary: an incumbent growing this fast on AI attach has limited commercial incentive to compete on price, which is consistent with the quote-only pricing above.

Good fit: firms and legal departments that already licence LexisNexis content and would otherwise pay twice to bolt a separate AI tool onto it; organisations whose blocker is client or regulatory data-handling requirements, where BYOK, workroom permissions and audit trails do real work; and teams whose core need is research and citation-grounded drafting rather than high-volume contract review. Legaltech Hub positions the target as corporate legal departments and firms from 20 lawyers up to 500+ (https://www.legaltechnologyhub.com/vendors/lexis-ai-by-lexisnexis/), which matches the enterprise shape of the security feature set.

Weaker fit: solo and very small firms, where the opaque quote process and add-on structure work against you and where the base content licence is the dominant cost regardless of AI; teams already committed to Westlaw content, since switching cost is the content, not the assistant; and buyers who want to compare shortlisted tools on published third-party benchmark scores, which for this product do not exist.

Worth knowing if you are cross-shopping: in the one qualitative head-to-head we found, vLex's Vincent beat both Lexis and Westlaw on the hardest of three questions (https://www.lawnext.com/2025/02/in-ai-smackdown-law-librarians-compare-legal-ai-research-platforms-finding-distinct-strengths-and-limitations.html), and in the Vals task benchmark Harvey led on five of six tasks it entered (https://www.artificiallawyer.com/2025/02/27/vals-publishes-results-of-first-legal-ai-benchmark-study/) — though neither comparison included the current Protégé build.