Evidence at a glance
One Case, Two Sets of Unsettling Numbers
On October 28, 2026, OpenAI published a case study about Legora, a legal AI company founded in Stockholm in 2023 whose CEO and co-founder Max Junestrand describes it as building 'the agentic operating system for legal work.' The case offers several figures: NDA review consumed about 20% of Legora's own legal team's capacity before, and now under 1%; an unnamed Fortune 100 customer cut contract turnaround by 90%, freed 9.5 hours per lawyer per week, and saved over $20 million in external litigation and due diligence spending.
These numbers are worth a technology leader's attention not because they are high but because they point to a different automation narrative: the model does not hand down legal conclusions, it expands work that was previously sampled only because of limited staffing into full coverage, and then returns lawyers to the judgment step. Understanding that distinction matters far more than remembering 95%.
What also deserves candor is that all these figures were published by OpenAI, with no disclosed customer name, measurement period, calculation method or independent audit. They show what Legora is doing, not that it works universally across legal teams.
Rules Before Models: Why NDA Review Can Drop Below 1%
The approach taken by Legora General Counsel Emma Kadri Bergström is worth unpacking. Her team first analyzed the NDAs the company had historically signed, built a review playbook from that, and let the system take a first pass at incoming NDAs, flagging terms outside acceptable parameters. The key here is not how smart the model is but the conversion of accumulated contract judgment into executable screening criteria, so the machine handles the mechanical part of checking each term against parameters.
NDAs happen to be an ideal target for this kind of redesign: highly templated, limited in clause types, with relatively clear criteria. The material says reviewing these 'boring documents' previously consumed about 20% of the legal team's capacity and now sits under 1%. The more accurate reading is that this capacity was freed up, not that the entire legal team became 95% more efficient.
That distinction matters for any team preparing to copy the path. If your contract types are highly customized and clause boundaries are fuzzy, converting historical contracts into rules will itself absorb significant lawyer time, and the payoff will be far slower to appear.
The Engineering Side of Bulk Review: Parallel Calls and Returning Results
Legora's Tabular Review is another path that leans more on engineering. Per OpenAI's case, it can pose multiple questions across large sets of contracts simultaneously, extract clause information, and analyze thousands of contracts within minutes. A Microsoft customer case from May 2025 offered earlier implementation clues: multi-agent design, retrieval-augmented generation and parallel model API calls, with 'tens of thousands' of parallel calls claimed at the time.
The two sources use different framings and cannot be merged into a single measured metric. But together they show one thing: the bottleneck in bulk legal review is not only model reasoning quality but scheduling, concurrency, result aggregation and traceability. When thousands of contracts are asked the same handful of questions, how outputs are structured and how a person quickly locates the source text is the core product work.
This also explains why Legora stresses 'operating system' rather than 'model.' Model capability is upstream supply; the real barrier is embedding model output into the document tools and research workflows lawyers already use. The Microsoft case mentions Word and SharePoint integration, while OpenAI's plugin description says users can invoke Legora's legal research from ChatGPT, get linked answers, and return to Legora to view citations, sources and audit information.
Who Holds the Gate: Agents Expand Scope, Lawyers Keep the Explanation
The M&A due diligence flow described in the material runs like this: agents review contracts and extract information first, then lawyers verify and explain the output. That ordering is not ceremonial; it is the precondition under which legal AI products are currently acceptable. Unusual clauses in a contract often require judgment against deal context; a model can flag what deviates from standard, but it cannot decide for the client whether that deviation is acceptable.
So claims like 'higher quality' and 'broader analysis' are, in the public material, interviewee assessments or product claims, unaccompanied by accuracy rates, error rates or controlled experiments. When a technology leader reads such phrases, they should treat them as product direction rather than proven general effect.
Traceable sourcing is not an add-on in this setting either. Linked answers, retrievable citations and audit information are the conditions under which a lawyer is willing to fold machine output into their workflow and sign off on it. Without that layer, bulk review only expands the volume of work awaiting verification.
Keep Two Pipelines Separate: Internal Coding Agents Are Not Legal Quality
The case also contains a set of figures easily misread. CTO Jacob Lauritzen says the engineering team can explore 20 solutions at once with coding agents and pick the best within days; when diagnosing issues, coding agents can search millions of lines of logs and sometimes move from bug report to working fix in about five minutes. These are executives' descriptions of internal cases, not general guarantees.
This efficiency and the legal product's effects are two different pipelines. The former accelerates experimentation and fixes, changing software delivery speed; the latter expands review scope, changing how much work lawyers can cover. Putting 'five-minute fixes' and 'faster contract review' into one efficiency narrative blurs what actually needs verification.
For legal or engineering teams preparing to adopt similar tools, the safer approach is to measure separately: on the engineering side, idea-to-prototype cycle and incident recovery time; on the legal side, review duration, coverage and human verification quality.
Boundaries and an Actionable Judgment
One layer of background remains unclear. The Microsoft case from 2025 said Legora used Azure OpenAI services and listed models such as GPT-4o and o1; OpenAI's 2026 case calls OpenAI a key model provider but does not specify current models, how they are divided across tasks, or the deployment architecture. Public material does not say whether Azure still hosts model calls, whether other models are used alongside, or whether OpenAI is the sole supplier. This is not a conflict, but it reminds readers that the case is a supplier narrative with incomplete stack detail.
For a seasoned legal team, the more practical starting point is to pick contract review or due diligence work that is rule-bound, high-volume and previously sampled only because of limited capacity, verify whether AI can broaden coverage, then shift lawyer time toward exception judgment and deal strategy. Measure review duration, coverage and human verification quality separately; do not substitute falling NDA hours for overall legal efficiency.
And 'agentic operating system' remains a positioning claim, not a public technical specification. Permission controls, automation boundaries, and which steps require lawyer confirmation are the details that determine whether the platform can be trusted, and the material does not cover them. Until those questions have answers, any verdict that legal AI is reshaping the profession should be held open.