Legora’s legal AI agents report big time savings, with lawyers still reviewing the work

Introduction: Legora’s AI‑agent breakthrough

Legora is using AI agents to move legal work through review faster and extend the amount of material lawyers can examine. Its own legal team reports a reduction of more than 95% in time spent reviewing nondisclosure agreements. Separately, CEO and co-founder Max Junestrand cites a Fortune 100 customer that reduced contract turnaround time by 90% and recovered 9.5 hours per lawyer each week.

Those figures appear in an OpenAI customer account published on October 28, 2026. They are reported outcomes from Legora and its leadership, rather than independently verified benchmarks. The distinction matters: the NDA result concerns Legora’s internal team, while the contract and weekly time figures concern one unnamed customer. They do not establish what every legal department should expect.

Founded in Stockholm in 2023, Legora describes its ambition as an “agentic operating system for legal work.” The practical idea is to embed AI in recurring legal workflows: taking an initial pass at agreements, surfacing exceptions and reviewing larger collections of documents before lawyers assess the results.

The strongest evidence in the account concerns that division of work. Agents handle an initial layer of analysis; lawyers review, interpret and exercise judgment. The opportunity is therefore broader than reducing the time spent on a familiar task. It includes examining documents that limited capacity might previously have left outside a detailed review.

How the agentic operating system works

OpenAI is a key model provider for Legora, powering workflows that include agentic legal work and large-scale document review. The account traces the relationship to GPT‑3.5, whose release made legal use cases the founders had been exploring possible. It does not identify the specific models used for each current workflow or provide a detailed account of the platform’s technical architecture.

The clearest operational example is Legora’s own NDA process. General Counsel Emma Kadri Bergström’s team examined agreements it had signed historically and used them to build a playbook. Legora then takes a first pass at incoming NDAs and highlights terms outside acceptable parameters.

That sequence explains what makes the workflow useful. The team supplies an established view of acceptable terms, and the system directs attention to deviations. The reported outcome is grounded in a defined review process, rather than an unspecified ability to answer legal questions. The account does not describe how the playbook is encoded or provide evidence that the workflow independently approves agreements.

For larger document sets, Legora’s Tabular Review workflow can, according to the account, analyse thousands of contracts across multiple questions in minutes. Alex Fortescue-Webb, Global Head of Legal Engineering, presents this as a way to broaden due diligence beyond the most valuable contracts. Agents surface information across the deal set, with lawyers reviewing and interpreting the output.

The account also says legal teams can access Legora’s capabilities within ChatGPT conversations through its plugin. That provides another access point, but the supplied evidence does not establish whether every workflow described is available through that interface.

Taken together, these examples give the operating-system language a concrete meaning: AI is being placed inside repeated legal processes, with human review remaining part of the work. They do not establish a fully autonomous legal practice or disclose enough implementation detail to reproduce the system.

Quantifiable efficiency gains (NDAs, contracts, lawyer time)

The headline numbers measure different things. Keeping their scope visible is essential to understanding what Legora has reported.

The NDA example supplies the most concrete before-and-after explanation. Kadri Bergström estimates that these agreements previously consumed around one-fifth of her team’s capacity. With the playbook workflow, the share is under 1%. This is a measure of legal-team effort, not a claim that every NDA completes its entire approval process more than 95% faster.

The context helps explain why that capacity matters. Kadri Bergström says Legora expanded from three offices to 18, completed three financing rounds and made five acquisitions in the preceding year. Her team had to support that growth while remaining close to the business. Reclaiming review capacity could help it absorb other work; the account does not quantify how the recovered time was distributed.

The weekly lawyer-time figure needs similar care. Although it appears prominently in the source’s headline statistics, the detailed account places it within Junestrand’s description of one Fortune 100 customer. It should therefore be read as a customer-specific result, not a platform-wide average or a promise to each user.

These are substantial reported gains. Their evidentiary limits are also clear: the supplied material offers no independent audit, comparative study or detailed measurement protocol. It supports describing successful reported workflows, while leaving their repeatability across other teams unresolved.

Internal impact: engineering speed and bug‑fix turnaround

Legora’s use of AI also extends to building its product. The company has expanded its internal use of ChatGPT Work and Codex, and CTO Jacob Lauritzen describes coding agents as changing both experimentation and troubleshooting.

For development, Lauritzen says engineers can explore 20 solutions at the same time and choose an approach within days. The significance is the ability to investigate alternatives concurrently before committing to a direction. This is a reported change in how the team explores a problem; it is not evidence of a twentyfold increase in engineering productivity.

Troubleshooting follows a similar pattern. Previously, several engineers might spend hours investigating different explanations for an issue. Coding agents can search millions of lines of logs and pursue possible causes simultaneously. The account says this sometimes takes the team from bug report to working fix in around five minutes.

“Sometimes” is a material qualification. Five minutes is an example of what has occurred, not an average resolution time or a commitment covering all defects. The account supplies no distribution of incident severity, typical repair duration or subsequent validation time.

The connection to customers is plausible and is explicitly made in the source: faster investigation can shorten the path from feedback to a product change, while parallel prototyping can bring ideas to customers earlier. Lauritzen credits this approach with helping the team build a better product faster.

These engineering examples should remain distinct from the legal outcomes. Coding agents are described as tools used by Legora’s developers. The evidence does not say that those same agents perform NDA review, or establish that engineering speed caused the reported contract-turnaround improvement.

Customer case study: Fortune 100 contract acceleration

The Fortune 100 example is the account’s clearest claim about external customer impact. Junestrand reports three outcomes for the same organisation: contract turnaround time reduced by 90%, more than $20 million saved on external litigation and diligence, and 9.5 hours returned per lawyer each week.

The source’s headline describes contract turnaround as “90% faster,” while its detailed customer passage says turnaround time was reduced by 90%. Those phrases are not mathematically interchangeable. This article uses the more explicit description in the customer passage and preserves the headline wording as a wording difference in the source. Without underlying measurements, the result cannot be recalculated or independently reconciled.

Nor does the account identify the customer, its original turnaround time, contract mix, implementation period or legal-team size. A percentage alone cannot establish how many days or hours were saved on a typical contract. The dollar figure likewise cannot be converted into an annual saving or return on investment without a measurement period, implementation costs and a fuller breakdown.

The account offers a possible explanation for the broader value of legal AI through its discussion of due diligence. Fortescue-Webb describes how time and cost constraints can limit detailed manual review to a subset of valuable contracts. Broader analysis can surface risks in contracts outside that subset for a lawyer to assess.

That mechanism is relevant to the case, but the source does not specify which workflows produced each Fortune 100 result. It would be premature to attribute the full saving to Tabular Review, the NDA playbook or any particular feature.

For another legal team, the useful questions follow directly from those gaps: which work changed, how much human effort remained, how turnaround was measured and whether broader coverage improved the resulting analysis. Legora’s reported customer results make that evaluation worth examining; they do not supply its answer for every organisation.

Future roadmap: longer‑running agents and real‑time collaboration

Lauritzen describes two directions for Legora’s future: agents handling longer-running tasks with more context and autonomy, and faster models supporting more immediate collaboration with lawyers. These are roadmap ambitions, rather than documented capabilities at the scale he imagines.

For longer-running work, his example is due diligence spanning millions of pages, processed overnight so a lawyer can review the analysis the next morning. This extends the current emphasis on breadth: agents would work through much larger collections before presenting results for human assessment. The account does not provide a release date or a demonstrated benchmark for that scenario.

Real-time collaboration addresses a different working pattern. Rather than sending a task and waiting for a completed answer, a lawyer could work back and forth with AI on a contract, analysis or piece of advice. The proposed benefit is a more continuous interaction between human judgment and model output.

The two directions need not be treated as a single performance claim. Overnight diligence depends on sustained work across extensive material; interactive collaboration emphasises responsiveness during an ongoing exchange. The source describes both as opportunities enabled by improving models, without specifying implementation requirements or availability.

Lauritzen says model improvements lead him to reconsider the roadmap every two weeks. That makes adaptability part of Legora’s product strategy, but also limits how firmly readers can interpret the future scenarios. They describe where its leadership wants to go, not a fixed delivery schedule.

Conclusion: Implications for the legal industry

Legora’s account suggests that the value of legal agents can lie in changing which work receives attention. Its NDA playbook reduces effort on a recurring task; broader document review can bring more contracts into scope; coding agents let its engineers investigate more alternatives. Across these examples, additional capacity is the connecting idea.

The leadership lessons in the source also point beyond model capability. OpenAI’s account recommends starting with the problem, building internal advocates, and incorporating evaluation, testing and transparency from the outset. Those recommendations fit the documented NDA workflow, where an established legal playbook gives the automation a defined purpose.

The evidence remains a vendor-published account with attributed results, an unnamed customer and limited measurement detail. It does not establish universal savings or independently measured improvements in legal quality. Its strongest contribution is a concrete description of how agents can take on initial review while lawyers retain interpretation and judgment. Whether that arrangement produces similar gains elsewhere depends on evidence from the workflows and teams that adopt it.