# How to Choose GenAI Review, Drafting & Assistants Tools in 2026 > A buyer's guide to legal AI assistants: what they handle well, where testing shows they slip, and the questions to settle before you sign. By The Circle Desk | Legal Ops | Published 2026-10-23 | Updated 2026-10-23 Source: https://legaltechcircle.com/insights/choose-genai-review-drafting-assistants-tools/ Short answer: buy a legal AI assistant for the jobs independent testing says it does well, like extraction, summarizing and first-pass review, and keep a lawyer's eyes on redlines and multi-jurisdiction research. Independent testing such as the [February 2025 Vals Legal AI Report](https://www.vals.ai/industry-reports/vlair-2-27-25) backs that split, and the tools are already inside most large firms, so test any shortlist on your own documents before you sign. What we hear from readers is less "does it work?" and more "which parts can I trust, and how do I catch the rest?" ## What Can a Legal AI Assistant Do Well Today? These tools take the slow first pass: summarizing a data room, pulling terms from agreements, marking up a contract against a playbook, or drafting a memo from prior work. In the February 2025 Vals report, AI tools collectively beat a lawyer baseline on document analysis, extraction, summarization and transcript tasks, while lawyers still outperformed the tools on redlining and on multi-step EDGAR research. In Vals' [October 2025 legal research study](https://www.vals.ai/industry-reports/vlair-10-14-25), every AI product tested beat the lawyer baseline overall, but scores on multi-jurisdiction questions were 14 points lower on average than on single-jurisdiction ones. Redlining is still the hard part. Ivo, a contract review vendor, published [ContractBench](https://www.ivo.ai/research/ivo-micro1-contractbench) in October, testing seven frontier models on 350 contract review tasks. No model passed 60% of weighted attorney-written criteria, models did worse responding to a counterparty's markup than on an opening redline, and attorneys made most edits inline while every model relied mostly on block rewrites. It tests underlying models rather than finished products, but it maps where to look during a trial. ## Who Is Buying, and What Does It Mean for the People Doing the Work? Big firms have mostly picked their platforms. [Harvey says](https://www.harvey.ai/blog/harvey-raises-dollar550m-at-a-dollar155b-valuation-to-help-legal-teams-own-their-intelligence) 80% of Am Law 100 firms use it, and Legora says more than half of the AmLaw 50 are customers. The hard part now is people: training, workflows, and making sure juniors still learn the craft the tool is speeding up. We know that last point is personal for a lot of readers. A&O Shearman cut an undisclosed number of business-team staff after significant investment in technology and data, the [Law Gazette reported](https://www.lawgazette.co.uk/news/business-team-redundancies-at-aando-shearman/5126721.article) on May 7, 2026, and Ropes & Gray now asks summer program applicants what they do day to day to keep up with AI, [per Law.com](https://www.law.com/americanlawyer/2026/04/22/for-law-students-ai-literacy-and-enthusiasm-may-impact-big-law-employment-prospects/) on April 22, 2026. Worry about jobs is reasonable, and the firms that handle it well will treat training as part of the rollout. In-house teams get the most from contract review against a playbook, because the same agreement types come back every week. Look for tools that work in Word, apply your fallback positions, and escalate rather than decide. Legora says in-house teams made up more than 40% of its new customers in the third quarter, according to its [Series E announcement](https://legora.com/newsroom/legora-raises-550-million-series-e-at-8.8-billion-valuation-to-accelerate-growth). At a mid-size firm, the build-versus-buy question is sharpest, because Microsoft made [Copilot for legal](https://techcommunity.microsoft.com/blog/microsoft-copilot-blog/introducing-microsoft-copilot-for-legal/4564246) generally available on October 15 with more than 20 legal skills for drafting, playbook review and obligation tracking. Already paying for Microsoft 365 Copilot? Test those skills before paying for a second assistant. And if you run a boutique, look for quick setup and per-seat pricing without large minimums, then judge the tool on the one or two high-volume tasks in your practice. ## What Should You Ask, and What Should Make You Walk Away? Which tasks has the product been independently tested on, and by whom? Ask whether the vendor took part in a Vals study or another third-party evaluation, and read the results rather than the summary. Then ask which models run underneath and whether you will be told when they change, since several vendors route tasks across multiple models and performance can shift after an update you never saw. Ask what happens to your data: whether prompts and documents are used for training, where data is stored, and which security certifications the vendor holds. Insist on a trial with your own documents. A demo on the vendor's sample contract tells you little, so run a pilot on agreements your team has already negotiated, where you know the right answer. The ABA's [Formal Opinion 512](https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/) identifies competence, confidentiality, client communication and reasonable fees as the model rules principally in play when lawyers use generative AI. Check what your own jurisdiction's rules and bar guidance say, and ask the vendor what controls it offers to support your review process. Be wary of an accuracy percentage with no test set, baseline or named evaluator, and of benchmarks the vendor ran on itself presented as comparisons with rivals. Ask to see an output the tool gets wrong and how a user would catch it, because a demo that never fails is a performance. Insist that every authority links to its source: in its October 2025 study, Vals noted more than 370 reported instances globally of AI-assisted court submissions with hallucinated citations or incorrect analysis. And get multi-year pricing, usage charges and seat minimums in writing before rollout, so the price does not move after the pilot. ## Which Vendors Should Be on Your List, and Are These Tools Worth It? A short starting list: [Harvey](https://www.harvey.ai) describes itself as legal AI for law firms and corporate legal teams, covering contract analysis, due diligence, compliance and litigation. [Legora](https://legora.com) sells a collaborative AI platform with an agent, tabular review and Word and Outlook add-ins. [Spellbook](https://spellbook.com) focuses on AI contract review and drafting, and [Luminance](https://www.luminance.com) offers a contract intelligence platform for legal and business teams. [Ivo](https://www.ivo.ai) sells AI contract review for in-house legal teams, and [Noxtua](https://www.noxtua.com) markets itself as Europe's legal AI for legal professionals and businesses. Our verdict is **Real**. We would buy for extraction, summaries and first-pass review today, because the independent testing backs that up, while redlining and multi-jurisdiction research should be piloted with a lawyer in the loop. ## FAQ **What tasks are legal AI assistants good at?** Extraction, summarization and first-pass review, according to independent testing cited in this guide, including the February 2025 Vals report. **Which ethics rules apply when lawyers use generative AI?** The ABA's [Formal Opinion 512](https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/), issued July 2024, points to competence, confidentiality, client communication and reasonable fees. Check your own jurisdiction's rules. **Should a firm with Microsoft 365 Copilot buy a second assistant?** Test Microsoft's [Copilot for legal](https://techcommunity.microsoft.com/blog/microsoft-copilot-blog/introducing-microsoft-copilot-for-legal/4564246) skills first. It became generally available October 15, 2026 with more than 20 legal skills. **What is the biggest red flag in a demo?** Accuracy claims with no method: no test set, no baseline and no named evaluator. *Updated Oct 23 to refresh the story and add an FAQ.* --- Legal Tech Circle (https://legaltechcircle.com/). News and analysis, not legal advice.