Many organisations begin discussing specialist AI in a very short way: choose a stronger model and use it for an industry. But work with serious consequences does not become dependable merely because a model answers more fluently. A useful system must know which sources to use, which rules to apply, where to stop, and who should make the final judgement. OpenAI’s new Astra for Law announcement is worth watching not only for law, but for any organisation seeking to introduce AI into professional work.
What is new?
On 17 September 2026, OpenAI introduced Astra for Law, a foundation for law firms and legal-technology companies to build products and workflows around their expertise. According to OpenAI, the system combines GPT-6 Astra with instructions, tools, and context designed for legal work; a legal search index can retrieve US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily.
OpenAI also says it tested this configuration on 200 US legal-research questions from Vals AI’s private validation set. At the highest reasoning effort, Astra for Law passed the overall correctness check on 54.0%, compared with 38.7% for GPT-6 Astra using web search alone, which is a 40% relative improvement as calculated by the company. This is a provider-reported result in a product article; it does not replace each organisation’s own evaluation on its work, data sources, and risk level.
The announcement also describes control layers for confidential work, including a controlled-access programme for eligible law firms, Zero Data Retention on the API for those users, and collaboration on information permissions, ethical walls, client instructions, and firm oversight. Together with the announced 26 partner plugins, the main message is clear: the model is only one part of a specialist-work configuration.
Source: OpenAI, “Introducing Astra for Law”, 17 September 2026.
The important point: specialist AI is a work system
The management interpretation is this: for professional work, the better question is not “which model is most intelligent?” but “which sources may this system rely on, and what verifiable result must it produce?” A specialist answer commonly needs at least four layers: authorised sources; instructions for applying sources to the situation; tools or workflow to retrieve the right context; and a review step by the person responsible.
This explains why an AI assistant with access to internal documents but without permissions can still create risk, even when its answer is persuasive. Conversely, a system that begins with a narrower source scope, shows its basis, flags uncertainty, and hands work over at the right point can be far more useful. The value is not in making AI speak like an expert; it is in helping an expert see the basis, exceptions, and points that require judgement.
Four decisions to lock before a trial
| Decision | Question to answer | Risk if ignored |
|---|---|---|
| Sources | Which documents may the AI search, which are excluded, and how are sources updated? | Answers rely on outdated, unauthoritative, or contextually unsuitable material. |
| Work scope | May the AI summarise, compare, draft, or propose a conclusion? | Users assign a decision beyond the evidence. |
| Review | Which results need citations, uncertainty warnings, or human approval? | Errors are hard to notice because an answer looks too confident. |
| Accountability | Who updates sources, changes rules, and handles user feedback? | The system slowly degrades without a clearly responsible owner. |
For example, a business may want AI to help its sales team answer product questions. The first step should not be to let the system answer every customer question. Begin with a narrower scope: retrieve approved policy, summarise service-package conditions, and draft an answer for an adviser to review. Each answer should identify the internal source used, state what needs confirmation, and provide a route to the responsible person when a question involves special pricing, legal commitments, or a situation outside policy.
Measure quality through work, not impression
Evaluating a specialist system only by the feeling that it “answers well” can easily create an illusion. An organisation should create a set of realistic cases with sensitive information removed, including common questions, exception questions, and questions the system must know to decline or transfer. Then assess: did it use the right sources; did it miss an important condition; how long did review take; and which errors repeat in a pattern?
There is no need to wait for a perfect system before a trial. But the trial’s scope must match the ability to review it. An assistant that helps find and compare documents can be deployed narrowly first; a system that recommends decisions for customers, personnel, or legal matters needs tighter control. Each expansion of scope should be treated as a new assumption to test, not as an automatic reward because the earlier model answered well.
Do not buy a promise; design a capability
Astra for Law is new information from one provider and remains limited by access group, geography, product, and deployment stage. Still, the management lesson does not depend on whether an organisation uses this product. Specialist AI creates more durable value when it is designed as a capability: authoritative sources, clear rules, suitable tools, accountable people, and a mechanism for learning from errors.
Seen this way, the AI discussion moves from “which model should we buy?” to “which work needs a better evidence and review path?” That is often a more practical starting point for a trial that is both useful and controllable.
