Tech • AI • Robotics • Game
Harvey’s Margins, Are Insects Worth More Than Humans?, Meta Tests Human Help for Muse | Diet TBPN
7/10TL;DR
Harvey faced a brief collapse in gross margins as legal AI usage surged, but recovered within a quarter by changing model routing, fine-tuning lower-cost systems and beginning a transition away from pricing that no longer matched agent-heavy workloads.
KEY POINTS
Negative margins triggered scrutiny
Bloomberg reported that legal AI company Harvey saw gross margins fall from about 50% to negative 50% in June. The drop drew attention because it implied the company was effectively selling more AI output than it could profitably support. In software, especially enterprise software, such economics are usually treated as a warning sign.
Usage, not demand weakness, drove the problem
The margin squeeze was tied to a sharp rise in customer usage rather than a slowdown in sales. Token consumption reportedly increased 20-fold this year as legal work shifted from simpler prompts to reasoning-heavy and agentic workflows. A single legal review task now can trigger broad searches across contracts, firm records and internal policies, producing far more model calls than older seat-based pricing anticipated.
Harvey says margins are positive again
Gabe, one of Harvey’s founders, said the company restored gross margins from negative 50% to positive within a single quarter even as usage kept doubling month over month. He said Harvey chose not to force customers immediately onto consumption pricing and instead continued serving stronger models while building cost controls. Those controls included routing requests to different models, post-training and spend-management tools.
Open-weight models are now part of the stack
Harvey has also fine-tuned an open-weight model, branded Harvey Tenant, as part of its effort to lower inference costs. That move cuts against the view held by some rivals that application companies should rely mainly on rapidly improving frontier models from providers such as OpenAI and Anthropic. The debate reflects a broader split across AI software: whether specialization through post-training can create durable advantages as base models get cheaper.
The burn rate looked painful but not existential
At an estimated $400 million annual recurring revenue run rate, Harvey would be generating roughly $33 million a month in revenue. A negative 50% gross margin month would imply roughly $16 million in gross losses for that period. That is substantial, but Harvey has reportedly raised about $500 million, making the episode serious without necessarily becoming a liquidity crisis.
Investors likely saw the issue before committing capital
Harvey’s recent Sequoia-led financing valued the company at about $15.5 billion. Given the timing, it is widely assumed that major investors had visibility into the June margin pressure before the round was announced. That suggests backers viewed the problem as a pricing and infrastructure mismatch during a product transition, not as evidence that demand or retention was failing.
The long-term benchmark remains much higher
The economics Harvey is trying to reach are closer to established legal software and services businesses. Public e-discovery company DISCO has posted around 75% GAAP gross margins, while Thomson Reuters has reported nearly 50% adjusted EBITDA margins in its legal professional software segment. Large law firms themselves are also highly profitable, with PwC estimates placing net profit margins for top firms at about 41%.
Cheaper model supply may ease pressure
Harvey’s recovery comes as the AI model market keeps getting more competitive. New releases, including Opus 5.5, Grok 4.7, GPT-6 Soul and Luna, underscore how quickly quality is improving while prices fall. One notable detail is that Grok 4.7 performed strongly on a Harvey legal agent benchmark focused on cost-sensitive legal tasks, suggesting Harvey and similar firms may gain additional leverage by shifting workloads across vendors.
The broader lesson reaches beyond legal AI
Harvey’s experience highlights a common challenge for AI application companies: rapid product improvement can break legacy pricing faster than finance teams can react. Seat-based subscriptions work poorly when users suddenly move from lightweight prompts to autonomous agents consuming vastly more compute. The companies that respond fastest with routing, pricing changes and tighter infrastructure may emerge stronger from the transition.
CONCLUSION
Harvey’s margin shock exposed how quickly AI economics can turn when customer workflows become more agentic. The company’s rebound suggests the main battle is not demand, but whether legal AI providers can align pricing and model strategy before usage growth outruns profitability again.
Ask a question