How Legora is transforming legal work with OpenAI

OpenAI's models have been the foundational engine enabling Legora to build an AI platform that transforms legal work by dramatically increasing both the spee

Abstract

This synthesis examines a practitioner account of foundation-model dependency and workflow transformation at Legora, a legal AI platform, framed around access to OpenAI's GPT-3.5 as a founding catalyst. The account's central thesis is that successive OpenAI models have functioned as a continuously upgraded capability engine beneath a stable application layer, enabling transformation along two axes: the speed of existing legal tasks and the breadth of tasks that become economically viable. A representative metric - 10,000 contracts evaluated against 10 queries each within minutes - illustrates a shift from selective, coverage-based document review toward exhaustive review. The analysis extracts implications for application-layer platform design, agentic workflows with human review, and organizational scaling under AI-enabled productivity, while noting the account's limitations as a single-vendor narrative.

1. Introduction

The commoditization of general-purpose language models has reshaped the economics of building vertical software. Where domain-specific natural language processing once required bespoke model development, application builders can now treat frontier models as an interchangeable capability substrate, differentiating instead through workflow design and domain integration. Legal practice offers a particularly instructive case: much of the work consists of reading, comparing, and summarizing large volumes of semi-structured text under severe time constraints, making it sensitive to improvements in model throughput and reasoning quality.

Several terms require definition before proceeding. An application-layer platform refers to software whose differentiation lies in workflow design, interface, and domain integration rather than in the underlying model weights. An agentic workflow denotes a process in which a model-driven agent autonomously initiates and executes a multi-step task, surfacing results for subsequent human adjudication rather than requiring step-by-step human direction. Coverage-based triage describes the longstanding practice of reviewing only a subset of available documents - typically those judged highest-value - because exhaustive manual review is infeasible within available time and budget.

The central claim examined here is that foundation models, treated as a substitutable "engine," have enabled an application-layer legal platform to transform practice along two distinct dimensions: task speed and task breadth, with downstream effects on firm-level scaling and individual productivity. Section 2 establishes the conceptual framing used in the account. Section 3 analyzes the reported transformation across origin, workflow impact, and organizational growth. Section 4 extracts technical findings with implementation considerations. Sections 5 and 6 discuss broader implications and limitations.

2. Background and Related Work

The account organizes its claims around two linked frameworks. The first is the AI-embedded legal workflow platform, exemplified by Legora, whose stated design goal is to make AI available within every task in a lawyer's working day rather than function as a discrete, separately invoked tool. This framing positions the product as workflow infrastructure rather than as a standalone assistant, implying that integration surface area - not raw model access - constitutes the platform's defensible value.

The second is the agent-based due diligence workflow, in which contract review tasks embedded within a larger diligence exercise are delegated to autonomous agents, with human practitioners positioned downstream as reviewers of agent output rather than as the primary readers of source documents.

The relationship between these frameworks and the underlying models is made explicit through a mechanical analogy offered in the source account:

"I like to think of Legora as a really fast Formula One car. But that car needs an engine."

This analogy carries an architectural implication worth isolating analytically: it treats the model as a replaceable high-performance component within a chassis that retains independent engineering value. The platform's roadmap is thus coupled to, but not fully determined by, the pace of improvement in the underlying OpenAI models, a dependency the account states directly: "Legora wouldn't be here today without OpenAI."

3. Core Analysis

3.1 Origin and Model Dependency

The account traces the founding of Legora to a discrete capability threshold rather than a gradual accumulation of interest or opportunity. Access to GPT-3.5 is identified as the specific trigger that prompted the founder to leave university and start the company: "It was the access to GPT-3.5 that made me drop out of Uni and that made me start this company." This detail is analytically significant because it locates the founding event at a model-release boundary rather than at a market-research or customer-discovery milestone, suggesting that the perceived step-change in model capability was itself treated as sufficient evidence of a viable product opportunity. Subsequent model releases are described as continuing this pattern, each functioning as an "engine" upgrade that increases the platform's speed and capability without requiring a change to the surrounding application.

3.2 Speed and Breadth of Contract Review

The most concrete transformation described concerns contract review practice. Prior to AI-assisted workflows, resource constraints forced firms into coverage-based triage: "Previously you'd have situations, for example, where only the 10% most valuable contracts would be reviewed." The remaining 90% went unreviewed not because they lacked legal relevance, but because manual review capacity was the binding constraint. The account reports that this constraint has been substantially relaxed, with the remaining 90% now reviewable "almost instantly using AI."

A specific throughput figure anchors this claim: 10,000 contracts, each evaluated against 10 queries, can be processed "within a couple of minutes." This figure is useful as an order-of-magnitude indicator of query-level parallelism and per-document latency, though the account does not specify the review depth per query, accuracy benchmarks, or the human verification overhead required afterward. The account further describes a full due diligence process run through an agent-based due diligence workflow, in which agents autonomously initiate and conduct contract reviews, with human practitioners reviewing agent output rather than conducting first-pass reads. This reflects a broader pattern in which the account distinguishes two separate effects of AI adoption: increased speed on tasks already being performed, and expansion into tasks previously excluded by capacity constraints.

3.3 Organizational Growth and Productivity Claims

The account situates these workflow changes within a period of rapid organizational scaling, reporting growth from three offices to eighteen within twelve months, alongside three financing rounds and five acquisitions in the same period. Legora is characterized as a "secret weapon" that allowed a comparatively small team to "keep pace with the company and drive the projects forward." This claim links the document-level throughput gains described in Section 3.2 to firm-level outcomes, though the account does not isolate the platform's contribution from other factors (capital availability, market timing, or team expansion) that plausibly also contributed to this growth trajectory.

4. Technical Insights

Several implementation-relevant observations can be drawn from the account, with appropriate caveats about the limited technical detail provided:

• Query batching at scale: The reported ability to apply 10 queries across 10,000 contracts within minutes implies a batched or parallelized inference architecture rather than sequential per-document processing. This is consistent with application layers built on top of model APIs that support high-concurrency request patterns, though the account provides no detail on model choice per query, context window management, or error-handling for malformed source documents.
• Agentic task delegation with human-in-the-loop review: The due diligence workflow described positions agents as initiators of multi-step review tasks, with humans reviewing output rather than directing each step. This is a design pattern distinct from conversational assistant tools, as it shifts the human role from primary reader to auditor, with implications for how review quality is measured and how liability for missed issues is assigned.
• Model-as-replaceable-component architecture: The "engine" framing implies the platform's application layer (interface, workflow logic, domain-specific prompting or fine-tuning) is engineered to tolerate underlying model substitution across successive OpenAI releases. This has a direct trade-off: it reduces switching costs if model providers change but concentrates dependency risk, since the account explicitly states the company "wouldn't be here" without continued access to the provider's models.
• Shift from triage to exhaustive coverage: Extending review from the top 10% of contracts to the full set changes the review population, which may affect what quality assurance or sampling methodology firms need going forward, since exhaustive AI-assisted review introduces a different error profile (systematic model blind spots) than selective human review (reviewer fatigue or inconsistency).

5. Discussion

The account's throughput and breadth claims point to a broader industry pattern in which AI adoption is reframed not merely as task automation but as a change in the addressable scope of professional work. The shift from reviewing 10% to effectively 100% of contracts is a meaningful claim because it implies that previously unreviewed legal risk is now visible, a change with implications for risk management practice beyond simple time savings. This suggests that productivity gains in knowledge work sectors may manifest less as headcount reduction and more as expanded task coverage, consistent with findings in the account that Legora's growth has been accompanied by increased, not decreased, organizational hiring and expansion.

A notable gap in the account is the absence of discussion of error rates, false negatives in contract review, or methodology for validating AI-identified issues against human-expert ground truth. Given that legal due diligence carries direct financial and liability consequences, the absence of such detail limits the extent to which the throughput figures can be treated as evidence of comparable or superior review quality, rather than simply increased speed and coverage.

The dependency framing in Section 3.1 also raises a structural question relevant to vertical AI platforms generally: the extent to which application-layer differentiation persists as underlying models improve and potentially subsume workflow functionality that previously required bespoke engineering. The account does not address this risk directly, instead emphasizing the platform's role as integration infrastructure rather than as a model wrapper.

6. Conclusion

This synthesis extracts from a single practitioner account two durable claims about AI-enabled transformation in legal practice: first, that foundation models can serve as a continuously upgraded "engine" beneath a stable application layer, with the founding and subsequent trajectory of Legora directly tied to specific OpenAI model releases; second, that the resulting productivity gains manifest along two distinct axes - increased speed on existing tasks and expanded breadth into previously excluded tasks - with the shift from coverage-based triage to near-exhaustive contract review offered as the clearest illustrative case.

Practically, the account suggests that legal and other document-intensive professional services firms evaluating AI adoption should distinguish between speed gains on current workloads and breadth gains that expand the scope of work undertaken, as these have different implications for staffing, risk exposure, and quality assurance design. Future investigation would benefit from independent validation of review accuracy at the reported scale, and from comparative analysis of how application-layer platforms manage dependency risk as underlying foundation models continue to evolve.

Sources

• How Legora is transforming legal work with OpenAI - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub