Evidence at a glance
From Deep Review of a Few Documents to Broader Coverage
An OpenAI case study published on October 28, 2026, introduces Legora, a company founded in Stockholm in 2023 that is building an AI-agent platform for everyday legal work, including contract review and due diligence. Its target is not legal judgment itself, but the practical problem that teams lack the time and budget to examine every document in depth.
That distinction shapes how to read Legora’s claim. Legal teams rarely lack documents to process; they lack the capacity to screen them one by one. If agents can broaden the first pass, lawyers can focus on unusual clauses and risks that require interpretation. The potential change is therefore not just speed, but how much material gets reviewed at all.
The “Operating System” Is Better Understood as Workflow Orchestration
Legora calls its product an “agentic operating system for legal work,” but the public materials more concretely support viewing it as a workflow orchestration layer, rather than a single model that replaces lawyers. A lawyer describes a goal and provides materials; Workflows can then plan steps and call tools and data sources. One workflow might combine tabular review, web search, and citation checks in legal databases before summarizing risks and drafting a report.
The underlying tasks are varied. A 2025 Microsoft customer case says Tabular Review can use models in parallel to extract clauses or answer questions across large document sets, while AI Agents use retrieval-augmented generation to find information across files; Legora also offers Word integration. Its ChatGPT plugin can return research results that link back to citations, sources, and audit information in the platform. For technical leaders, the key design question is not whether an agent can produce an answer, but what evidence each step uses, whether the result can be traced to that evidence, and how lawyers review it.
The Case-Study Numbers Show Potential, Not General Proof
Legora’s figures are worth examining, but only with their scope attached. The company’s general counsel estimates that NDA review once consumed about 20% of her team’s capacity. After the team created review rules and had the system take the first pass, that share fell below 1%. This is an internal example from Legora’s own legal team, not an average result that other firms or companies can assume they will reproduce.
A second set of figures concerns an unnamed Fortune 100 customer. Legora says the customer cut contract turnaround time by 90%, returned 9.5 hours per lawyer per week for higher-value work, and saved more than $20 million in external litigation and due-diligence costs. The public case does not disclose its baseline, calculation methods, or an independent audit. The numbers indicate possible scale, but cannot establish which part of the workflow drove the savings or predict returns elsewhere. They are more useful as hypotheses to test than as a procurement promise.
Faster Internal Engineering Does Not Prove Better Legal Analysis
Legora also uses AI within its own engineering team. Its CTO says coding agents let the team explore multiple approaches in parallel and search millions of lines of logs while investigating problems; sometimes, a reported issue can lead to a working fix in about five minutes. The company says this helps product teams receive feedback sooner and lets them test ideas with customers earlier.
That efficiency chain matters, but it should not be confused with the quality of legal analysis. OpenAI says Legora’s internal teams use ChatGPT Work, Codex, and dots. Codex is described here as an engineering tool, not as the legal analysis engine offered to lawyers. Its effect is indirect: shorter cycles for experimentation and debugging may help resolve customer issues sooner. Whether legal workflows are accurate and their citations reliable remains a separate question that needs its own evaluation.
Audit the Evidence Trail Before Expanding Automation
Architecture decisions should also avoid reading “powered by OpenAI” as a complete account of a single-model or single-cloud stack. OpenAI describes itself as an important model provider for Legora, while a 2025 Microsoft customer case records the use of Azure infrastructure and Azure OpenAI Service. The materials are from different dates and may cover different parts of the product; they do not reveal which models and cloud services each current workflow uses.
For teams considering legal agents, a practical approach is to start with tasks governed by clear rules and where mistakes are easy to detect, such as screening NDAs against an agreed checklist. Then test exception detection, citation links, audit records, and human review within the existing process. Expansion into more complex work such as due diligence should follow only when false positives, missed issues, and review costs can be measured, and responsibility remains with the appropriate professionals. Bulk processing can widen a lawyer’s field of view; organizing material is not the same as making a legal judgment.