Building the future of legal work with AI | Legora CEO Max Junestrand

Legora argues that the legal industry is fundamentally supply-constrained and that AI (enabled by models like GPT-3.5) can unlock vastly more legal capacity

Abstract

This synthesis examines a thesis advanced by the legal technology firm Legora: that the legal services market is structurally supply constrained rather than demand constrained, and that large language models (LLMs) relax this constraint by expanding the volume of legal work performable at acceptable cost, rather than simply substituting for existing labor. The account is reconstructed from a founder-level discussion of the company's origins, product philosophy, and deployment outcomes. Three findings are salient. First, the viability of the firm's earliest use cases shifted discontinuously with model capability, moving from intractable with BERT to workable with GPT-3.5. Second, reported enterprise outcomes at a Fortune 100 client include a 90% reduction in contract turnaround time, over $20 million in avoided litigation and diligence costs, and 9.5 hours returned per lawyer per week. Third, trust is characterized as the binding constraint on adoption. Implications concern capability-threshold product planning and verification design in high-liability domains.

1. Introduction

Discussions of artificial intelligence (AI) in professional services typically center on substitution - which tasks a model can perform in place of a human, and at what cost saving. The account synthesized here advances a different framing: the legal sector is supply constrained, meaning demand for legal services significantly exceeds the supply of qualified capacity to meet it. Under this framing, the primary effect of AI adoption is an expansion of total legal work performed - contracts reviewed, disputes litigated, assets diligenced - rather than the displacement of work already underway.

Several terms recur and warrant definition. Diligence denotes the systematic review of an asset, entity, or document corpus prior to a transaction such as an initial public offering (IPO) or a merger and acquisition (M&A) event. Contract turnaround time measures the interval between receipt of a draft agreement and return of a reviewed, negotiated, or executed version. AI-native describes an organization whose product architecture and development practice assume continuous improvement in underlying model capability, rather than treating model capability as a fixed baseline to be engineered around.

The central question this analysis addresses is whether capability improvements in general-purpose LLMs translate into structural change in a professional domain characterized by high liability, high verification cost, and entrenched billing conventions (notably the billable hour). Section 2 establishes relevant context. Section 3 analyzes the supply constraint, the observed capability threshold between model generations, reported deployment outcomes, and the role of trust. Section 4 extracts technical and implementation findings. Section 5 situates these within broader industry trends and notes evidentiary limitations. Section 6 concludes.

2. Background and Related Work

Two frameworks organize the account. The first is the operating system positioning: Legora describes itself as the operating system for legal work, with the stated ambition that every piece of legal work eventually run on the platform. This framing implies a general substrate over which heterogeneous legal tasks execute, distinct from narrow, task-specific tooling such as a single-purpose document classifier.

The second is the Formula 1 analogy, which decouples the product layer from the underlying model layer: the product is described as "a really fast Formula 1 car," while successive model releases from OpenAI are described as new "engines" that make the car faster without requiring the car itself to be rebuilt. This framing is paired with an explicit product philosophy of "building sandcastles" - treating current implementations as provisional, since the underlying technology (the engine) shifts quickly enough that permanence is neither achievable nor a sound design goal.

3. Core Analysis

3.1 The Supply Constraint as Market Structure

The account characterizes legal services as fundamentally undersupplied relative to demand: not every contract is reviewed, not every injustice reaches litigation or courts, and diligence on assets is often performed reactively (after an event) rather than proactively, because capacity does not exist to perform it beforehand. This is presented as a structural condition of the market rather than a temporary bottleneck, which matters for how AI's effect should be interpreted. If the constraint is genuinely on the supply side, then capacity added through AI tooling is more likely to be absorbed into previously unaddressed demand than to produce net headcount reduction among incumbents - a distinction with direct bearing on adoption incentives for law firms and in-house legal departments.

3.2 A Capability Threshold Between Model Generations

A specific and falsifiable technical claim in the account is that Legora's founding use cases were explored first against BERT, an early Google model, and were judged "too difficult to solve" at that time. The same class of use cases became workable only once access to GPT-3.5 was obtained. This is described as a discontinuous shift rather than a gradual one: the founder left university upon gaining GPT-3.5 access, characterizing it as "my generation's internet moment," and the account states plainly, "with no OpenAI, there's probably no Legora." The implication is that certain classes of legal-document reasoning sit above a capability threshold that intermediate-generation language models did not clear, and that product viability in this domain was gated by exogenous model progress rather than by engineering effort alone.

3.3 Reported Deployment Outcomes

The account reports outcomes from a single named category of client (a Fortune 100 company) rather than an aggregate or benchmark figure: a 90% reduction in contract turnaround time, more than $20 million saved on external litigation and diligence costs, and 9.5 hours per lawyer per week returned for reallocation to higher-value tasks. These figures function as evidence that AI-enabled capacity expansion, rather than the supply constraint remaining binding, is observable concretely at enterprise scale once adopted. No baseline methodology, measurement period, or independent verification is described in the source material, a limitation addressed further in Section 5.

3.4 Trust as the Binding Constraint

The account identifies trust as "perhaps the most important currency when it comes to law, even more important than the billable hour," and states that customers must never feel they are second-guessing or compromising on trust when using the product. This positions trust not as a soft reputational consideration but as the primary adoption gate in a domain where output errors carry direct liability exposure. Unlike the capability threshold (Section 3.2), which is a property of the model, trust is framed as a property of the product and its surrounding verification affordances - implying that model capability alone is necessary but not sufficient for adoption in legal workflows.

4. Technical Insights

Several implementation-relevant findings emerge. First, the shift from BERT to GPT-3.5 suggests that legal document reasoning tasks of the kind Legora targets require capabilities - plausibly including longer context handling, stronger few-shot generalization, or more robust instruction-following - that were absent or too weak in earlier encoder-only architectures to be commercially viable. Second, the "sandcastles" philosophy implies a specific engineering trade-off: building thin, frequently-replaced orchestration layers around a frontier model rather than investing heavily in model-specific fine-tuning or permanent custom infrastructure, since the latter risks obsolescence as model generations turn over. This has direct implications for build-vs-buy decisions in applied AI products - architectural permanence is treated as a liability, not an asset, when the underlying capability layer is improving quickly. Third, the reported outcome metrics (90% turnaround reduction, 9.5 hours/week returned) suggest the product targets high-volume, moderate-complexity review tasks (contract review, diligence document review) where throughput gains compound, rather than low-volume, highly bespoke legal reasoning. Fourth, the emphasis on trust implies that production deployment in this domain likely requires explicit verification or citation mechanisms, though the source material does not specify technical detail on how trust is operationalized within the product itself - a notable gap given its stated centrality.

5. Discussion

The supply-constraint framing, if accurate, has implications beyond Legora specifically: it suggests that AI-driven productivity gains in professional services with structural undersupply (law, and plausibly medicine or specialized engineering review) are more likely to manifest as expanded service volume than as labor displacement, at least in the near term. This is consistent with the account's framing of outcomes in terms of hours "returned... for higher-value tasks" rather than roles eliminated.

A notable gap is evidentiary: the reported client metrics derive from a single enterprise account, as communicated by the vendor, without independent benchmarking, disclosed measurement methodology, or comparison against alternative tooling. This is a general limitation of founder-narrated case studies and should temper the confidence with which the 90% and $20 million figures are treated as representative rather than illustrative.

The capability-threshold finding (Section 3.2) connects to a broader pattern observed across applied-AI startups founded in the GPT-3.5-to-GPT-4 era: product viability frequently tracked exogenous model releases rather than incremental in-house engineering, a dynamic with implications for how founders and investors assess defensibility in AI-native products built atop third-party frontier models.

6. Conclusion

This account supports two connected claims: that a structurally undersupplied professional market can convert model capability gains directly into expanded service delivery, and that in liability-sensitive domains, model capability alone does not secure adoption - trust in output remains the binding constraint. Practically, this suggests that organizations building AI products for high-liability professional domains should treat capability thresholds as discrete, model-generation-dependent events to plan around, and should invest commensurately in trust and verification affordances rather than assuming capability improvements alone will drive adoption. Future inquiry should seek independently verified outcome data and greater technical detail on how trust is operationalized at the product layer.

Sources

• Building the future of legal work with AI | Legora CEO Max Junestrand - 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