September 2026 legal tech recap for in-house teams: OpenAI's Astra for Law, Harvey's raise, TR v. ROSS, California SB 574 and No Robo Bosses, and consolidation.

Legal Tech Rewind is our month-by-month look back at 2026 for in-house legal teams: what happened, why it mattered, and what it looks like with the benefit of hindsight. Published October 2026.
OpenAI entered legal research, the specialist vendors raised money and bought guardrails, and courts and legislators turned AI best practices into binding rules.
On September 17, OpenAI launched Astra for Law. It isn't a new model. It wraps OpenAI's GPT-6 Astra in a legal search index of more than 230 million URLs covering US case law, statutes, regulations and court rules. On a 200-question legal research benchmark from Vals AI, OpenAI reported 54% correctness, compared with 38.7% for the same model using plain web search. It launched with 26 partner plugins, including Thomson Reuters, iManage, Relativity and Clio. Harvey and Legora were named as API customers, and Latham & Watkins helped design governance for ethical walls and permissions.
Why it mattered: with Anthropic in May, Google in August and OpenAI in September, three of the biggest AI model companies built legal products within five months. Microsoft followed in October. 54% is a big improvement and also a reminder that legal research AI still needs a lawyer checking the work.
On September 9, Harvey raised $550 million at a $15.5 billion valuation, up from $11 billion in March, co-led by Diffusion and Lightspeed. The same day it acquired Guardrails AI, an agent reliability company whose framework tests and monitors how AI agents behave. Harvey CEO Winston Weinberg framed the deal around a question clients keep asking: "how do you know what it will do?"
Why it mattered: the top legal AI vendors now have years of runway, while many smaller vendors don't. And agent testing is becoming a selling point. Ask every agent vendor how it tests behavior before the agent touches your matters.
On September 29 and 30, the Third Circuit affirmed Thomson Reuters' win over ROSS Intelligence. It held that 2,243 Westlaw headnotes are copyrightable and that copying them to train ROSS's AI research tool was not fair use, calling the use "minimally transformative, at best." The court recognized a developing market for licensing content as AI training data, but a footnote limits the ruling to ROSS's non-generative tool. ROSS said it will ask the Supreme Court to review the case.
Why it mattered: it's the first federal appellate word on AI training and fair use. It strengthens licensing arguments for proprietary content, including yours, and raises the stakes on vendor IP indemnities. We broke down what it means for your vendor contracts in our TR v. ROSS post.
On September 30, Gov. Gavin Newsom signed SB 574, the first state law governing how lawyers and arbitrators use AI. Lawyers must verify citations, including AI-generated ones, disclose generative AI use in court filings, and keep confidential information out of open AI tools, and they may not delegate the practice of law to AI. Arbitrators may not delegate decisions to AI.
He also signed a 13-bill AI package that included the "No Robo Bosses Act," which bars employers from relying solely on automated systems to fire or discipline workers.
Why it mattered: citation checks and AI disclosure moved from "good outside counsel guidelines" to state law in the biggest legal market in the US. And employers using automated tools to discipline or fire California workers will need human review in the loop.
Why it mattered: neither story is about AI, and both are about the basics: privilege discipline and data handling by the firms that hold your most sensitive information.
Repario acquired UnitedLex, combining two private equity-owned legal services providers. Epiq bought Canopy, whose software works out what data was exposed in a breach and who was affected. Consilio bought Australian e-discovery provider Icourts, and Clio bought Learned Hand, an AI workspace for judges.
Why it mattered: if one of these providers runs work for you, expect account and pricing changes. Use our 30-day vendor acquisition checklist.
Why it mattered: adoption is nearly universal. Governance and value measurement are not. That 86% vs. 17% gap is the slide to put in front of your CFO.
September's lesson: the AI rules are no longer suggestions. Build your policy as if a judge will read it, because one might.
More from the 2026 Legal Tech Rewind: ← August 2026 · The full 2026 Rewind: 9 lessons
This article is general information, not legal advice.