Legal Research in Minutes: What Specialized Law-Firm AI Means for Your Contract Review

On September 17 and 18, 2026, OpenAI unveiled "Astra for Law," a version of its GPT-6 Astra model configured specifically for legal professionals. The timing is no coincidence: competition for the "right" AI for law firms and legal departments has sharpened considerably in recent months, and with Astra for Law, the world's largest AI provider is now entering directly into a field previously dominated by specialized providers like Harvey or Legora. For executives and in-house counsel at German mid-market companies, this raises a concrete question: what does it mean for your own contract review, and where are the limits you should know before feeding sensitive client or contract data into such a system?

This article breaks down the announcement, shows which figures are solid and which should be treated with caution, and translates the development into concrete implications for companies that don't treat contracts as their core business but as a necessary part of their day-to-day operations. Because amid all the excitement over a new OpenAI product, one thing is easily overlooked: the real question for your company isn't "when will we get Astra for Law?" but "which of our recurring contract reviews can already be meaningfully supported by AI today, and under what data-protection conditions are we even allowed to do that?"

What Astra for Law actually is

According to OpenAI, Astra for Law isn't a new model generation, but a version of the existing GPT-6 Astra model configured for legal work. Three building blocks set it apart from a standard ChatGPT account: a dedicated legal search index, tailored instructions for legal analysis and legal writing, and enterprise controls such as ethical walls, granular access rights and client-specific requirements.

According to OpenAI, the legal search index covers US case law, statutes, regulations, court rules and administrative decisions. To build this index, OpenAI is partnering with, among others, the nonprofit Free Law Project and its CourtListener database. According to available sources, the index covers more than 230 million URLs of US legal sources – an impressive figure that, however, comes from a single secondary source and could not be independently verified.

At launch, OpenAI is integrating 26 partner plugins, including names like Thomson Reuters, Intapp, Harvey, Legora, DeepJudge and iManage. OpenAI President Greg Brockman added on X that an additional 47 community plugins would also be available for ChatGPT. Access itself is initially limited: it runs exclusively through a "Trusted Access Program" in ChatGPT and Codex, with API access to follow. None of the sources researched mentioned specific pricing or an explicit rollout for Europe or Germany – an important point for context, because it means Astra for Law simply isn't available yet for German mid-market companies.

Also notable is the enterprise-controls layer that OpenAI is putting front and center. Ethical walls between clients, granular access rights for individual teams, and contractually agreed client requirements have been standard practice at large law firms for years, because a conflict of interest between two clients of the same firm can carry serious professional-conduct consequences. That OpenAI is explicitly marketing these controls as a product feature shows how seriously the provider is taking the structural requirements of the legal industry – and how far these requirements go beyond what a standard chatbot account typically offers.

The numbers behind the hype – and how solid they really are

OpenAI itself cites an overall accuracy rate of 54.0 percent for Astra for Law on a legal-research benchmark, compared with 38.7 percent for the standard GPT-6 Astra model using plain web search. That's a relative improvement of roughly 40 percent – a fairly strong signal that a specialized legal search index makes a real difference over general web search. Important for context: this figure comes from a single source (LawSites, citing OpenAI) and hasn't been confirmed by a second, independent party.

The market size for "legal AI" also can't be boiled down to a single number. Depending on scope and research provider, estimates for 2026 range between roughly $3.7 billion and $5.6 billion, with projected annual growth rates of between about 22 and 31 percent into the 2030s. This spread of more than double shows that these are rough estimates from various analyst firms, not a fixed metric. You should treat figures suggesting that 79 percent of legal professionals already use AI in 2026 (up from 19 percent in 2023), or that 52 percent of Fortune 500 companies were already using AI-assisted contract review at scale in 2025, with similar caution – both numbers come from secondary sources without clearly documented methodology.

A single vendor press release claiming "94 percent accuracy in 26 seconds" for AI contract review is a marketing promise, not an independently verified metric. Such figures belong in any assessment with their source clearly noted – not as proof, but as an example of how widely the reliability of claims in this market can vary.

The competition: Harvey, Legora and Robin AI

OpenAI isn't entering an empty field with Astra for Law. Harvey, one of the best-known legal AI providers and also an Astra for Law partner, was valued at around $11 billion following a $200 million funding round in March 2026. Legora, also a partner in the new program, raised a $550 million round in March 2026, topped up with a $50 million extension in April 2026, bringing it to a valuation of roughly $5.5 billion to $5.6 billion on total funding of around $866 million. These figures show how much capital is currently flowing into specialized legal AI – and that Astra for Law is more of a complement than a replacement for these providers, since both are simultaneously partner plugins in the new OpenAI program.

Robin AI positions itself differently: as a document- and contract-focused platform with a stated focus on mid-sized law firms and corporate legal departments. For the German mid-market, that's the more relevant point of comparison, because it shows that there are tools specifically built for contract review that don't target US big-law firms, but exactly the scale at which many mid-market companies operate with a small in-house legal team.

It's worth looking at how differently these providers are positioned for your own assessment. Astra for Law bets on breadth: a huge US legal-sources index, a large partner ecosystem, and OpenAI's market power, but for now only limited, curated access. Harvey and Legora bet on depth within the big-law world, with correspondingly high valuations and correspondingly high licensing costs. Robin AI and comparable providers bet on focus: a narrower product for contract review, designed to be accessible to smaller legal departments too. None of these three approaches is inherently "the right one" – what matters is your company's actual contract volume, language and data-protection requirements.

What this specifically means for the German mid-market

Here's the key point for your assessment: according to the announcement, Astra for Law is primarily aimed at large law firms, legal-tech providers and US legal sources. No rollout for the EU or German-language legal databases has been confirmed. The direct benefit for German mid-market legal departments is therefore currently unclear – you should treat that as an open question, not a foregone conclusion.

Even so, it's worth looking at the underlying patterns that apply to your company too. Industry sources see the practical benefit of specialized legal AI mainly in the roughly 80 percent of standard contracts with a high degree of repetition: NDAs, data processing agreements, framework agreements and terms and conditions. This is exactly where specialized, retrieval-augmented tools can noticeably reduce research and review effort, while complex individual contracts continue to require legal review by a lawyer. The German Legal Tech Monitor 2026, for which the Legal Tech Verband Deutschland surveyed nearly 360 companies, offers two figures worth knowing here: 57 percent of surveyed in-house counsel are increasingly handling tasks themselves with AI that used to be outsourced to law firms. And 73 percent would consolidate their mandates with law firms that demonstrably cut costs through AI.

Together, these two figures create a double lever for mid-market companies. First: with your own AI support tailored to standard contracts, you can reduce external legal fees without replacing your law firm for complex cases. Second: when selecting external law firms, it's increasingly worth asking how consistently they use legal AI productively themselves – a cost lever via law-firm selection, not just your own tooling.

In practice, this means for internal preparation: first get an honest picture of how many of your annual contract reviews are actually standard cases – NDAs with new suppliers, recurring data processing agreements with SaaS providers, framework agreements with known contract partners. Experience shows that this share in many mid-market companies falls right in the range industry sources cite for the market as a whole: well over half, often around four out of five cases. For exactly these cases, an AI-assisted first-pass review pays off – one that flags typical risk clauses, marks missing standard language, and provides an initial assessment before a contract ever reaches a human. The remaining, genuinely individual contracts – such as complex cooperation, licensing or M&A agreements – remain a job for the legal department or outside counsel, but now with more time, because routine cases no longer tie up the same effort.

Data protection and professional-conduct rules: the real stumbling block

Before feeding contract or client data into any AI system, you should take the legal side seriously. German legal experts warn that entering such data into US cloud AI services like ChatGPT could potentially touch on attorneys' duty of confidentiality under Section 43a of Germany's Federal Lawyers' Act (BRAO), with possible implications under Sections 203 and 138 of the German Criminal Code (StGB). A definitive court ruling on this question is still pending according to available sources – you should treat this as an open legal question, not a settled one. Using such a system as a data processor also requires a data processing agreement under Article 28 GDPR, and data transfers to the US remain contested from a data-protection standpoint following the Schrems II ruling.

In practice, that means: before putting any contract-AI tool into productive use – whether from OpenAI, one of the competitors named above, or another provider – you should clarify at least three points. First, whether a valid data processing agreement under Article 28 GDPR is in place and which subprocessors it names. Second, whether and how personal or client-related data leaves the EU, and on what legal basis that's permissible under the principles established by Schrems II. Third, whether your own professional or industry code of conduct – for example as in-house counsel, a tax advisor, or in a regulated industry – imposes additional confidentiality requirements beyond general data-protection compliance. None of these points can be answered with a blanket "yes, that's settled" – that goes for Astra for Law just as much as for any other tool that processes contract content.

As a privacy feature, OpenAI cites "zero data retention" via the API for eligible law firms, as well as a default exclusion from human review when used via ChatGPT Enterprise. For controls around information barriers, ethical walls and client requirements, OpenAI says it's working with the law firm Latham & Watkins. These are useful building blocks – but they don't replace your own legal assessment of whether and how such a system can be used in a data-protection-compliant way at your company.

Add to that the regulatory dimension at the EU level: obligations for general-purpose AI under Articles 51 through 56 of the EU AI Act have applied since August 2025. According to one industry source, high-risk-classified AI systems in the area of access to justice will need complete compliance documentation and registration in the EU AI database by August 2, 2026. If you're deploying or evaluating AI systems in the legal field, this assessment belongs in your review from the start, not as an afterthought once compliance becomes urgent.

Specialized legal AI versus generic chatbots

One reason specialized, retrieval-augmented legal research tools are considered significantly more reliable than generic chatbots in trade publications lies in how they work: they pull from primary sources, verify citations, support separation of clients and files, and keep audit logs. A widely cited 2025 Stanford benchmark found differences to match: on highly specific legal questions, generic models reportedly hallucinated in 69 to 88 percent of cases, while specialized tools like Lexis+ AI and Westlaw AI-Assisted Research came in well below that, with hallucination rates of 17 and 33 percent respectively; GPT-4 was reported at 43 percent in the same study. For generic document summarization – a less demanding task – hallucination rates dropped to 1 to 3 percent. These figures come from secondary citations and should be treated with corresponding caution, but they point to a plausible pattern.

In practice for your company, that means: a specialized tool with access to verified primary sources and a traceable audit trail is clearly preferable for legally binding statements – such as on contract risks or the review of individual clauses. A general-purpose chatbot is better suited for initial orientation or rough research, not for the final legal assessment of a contract. This distinction should inform any internal policy on AI use in contract review, regardless of which specific tool you ultimately deploy.

Astra for Law is thus less a ready-made solution for the German mid-market than a clear signal: the major AI providers are investing specifically in legal specialization, because generic models have noticeable weaknesses on legal questions. For your company, that doesn't mean waiting for a US announcement, but clarifying now which of your own recurring contract reviews can already be meaningfully supported by specialized, data-protection-compliant AI agents available today – and where legal review by a lawyer remains indispensable.