Legal was supposed to be the industry that AI couldn't crack.
Slow procurement. Conservative buyers. A profession built on human judgment. For decades, software companies struggled to sell anything to lawyers beyond email and document management. Enterprise sales cycles stretched for years. Deals died in legal review committees. Every tool required a partnership with a firm before it could get in front of anyone who mattered.
Now legal is generating $500M in annualized revenue from AI-native startups alone. That makes it the #2 vertical in enterprise AI, sitting just behind coding.
If you had predicted that five years ago, you would have been laughed out of the room.
The Hard Data
a16z published hard data on where enterprise AI is actually working. Not survey sentiment. Not self-reported adoption numbers massaged through an analyst report. Revenue from real enterprise deployments.
The figures tell a clear story.
Coding leads at $3B. Legal sits at $500M. Support at $400M. Medical and health administration at $350M. Search at $250M. Writing, real estate, financial analysis, nursing, and accounting all sit below $150M.
Coding dominates by nearly an order of magnitude, and that makes intuitive sense. Code is text-based, output is verifiable, feedback loops are tight, and developers have always been willing to adopt new tools quickly. There is also an enormous addressable market. Every company on earth employs software developers.
But the legal number is the one that should make every other industry pay attention. Because legal is not supposed to look like this.
Why Legal Was Supposed to Be Different
The conventional wisdom around legal technology held for good reason.
Lawyers are trained to be skeptical of new tools. The profession runs on precedent, and precedent is inherently conservative. A mistake in a contract is not a bug you patch in the next sprint — it is a liability with a client name attached to it.
Procurement in law firms moves through layers of approval. Risk and compliance sign off. IT reviews infrastructure. Partners who generate revenue hold veto power. Selling into a firm meant a relationship cycle measured in years, not quarters.
And the work itself seemed uniquely resistant to automation. Reading thousands of pages of documents, reasoning through complex arguments, understanding the commercial context behind a clause, drafting a response that reflects a client's specific posture — that work required the kind of judgment that made people believe it could only be done by humans.
What Actually Changed
Large language models made workflows possible that never existed before.
Previous generations of legal tech worked around the constraint of software not being able to read and reason at scale. It indexed documents. It flagged keywords. It surfaced relevant precedents. It automated the periphery — billing, scheduling, simple document assembly.
The core work, the stuff that consumed most of the billable hours, remained human.
LLMs changed the core. Now the software can read the thousands of pages. It can reason through complex arguments. It can hold the context of a deal and draft a response that reflects actual commercial judgment. Not perfectly, not without review, but at a level of quality that makes it genuinely useful as a first pass rather than a toy.
That shift unlocked a category of workflow automation that had been impossible to build before.
The Companies That Proved It
Harvey hit roughly $200M ARR within three years of founding. Eve crossed a $1B valuation with 450+ enterprise customers. These are not pilots sitting in an innovation lab waiting for a champion to sponsor them through the next budget cycle. These are production deployments generating real revenue in an industry that made software vendors miserable for decades.
What both companies understood is that the value proposition in legal is not "save lawyers from doing things they hate." Lawyers are mostly fine with the work. The value proposition is "let the same number of lawyers do more of the work that actually matters to clients."
When you can generate a high-quality first draft in minutes instead of hours, you can take on more matters with the same team. When you can run a contract review across hundreds of documents overnight, you can surface risk that would have required a dedicated due diligence team. The leverage story is the economic story.
The Benchmark That Changes the Framing
The revenue numbers are significant. But the second chart in the a16z report is the one that reframes what is actually happening.
It overlays revenue data against GDPval, OpenAI's benchmark that measures how often models outperform human experts on economically valuable tasks. The results explain why legal is not a special case. It is a preview.
In just four months, model win rates against human accountants jumped 18 percentage points. Against industrial engineers, 27 points. Real estate sales agents, 24 points. Software developers, 12 points. Even performance in police and detective work improved by 27 points.
These are not incremental improvements. These are shifts large enough to change what it means for a model to be "useful enough to deploy" in each of these fields. A model that goes from losing 70% of the time to losing 43% of the time has crossed a threshold. It becomes worth integrating into a workflow even if it still requires human review.
The capability curve is not moving slowly. It is not pausing to wait for a particular profession to feel ready. It continues regardless of procurement timelines.
The Real Implication
29% of the Fortune 500 are already live, paying customers of AI applications. That number was not imaginable three years ago.
The frame that most people are working with — that industries will have years to prepare, that adoption curves will be gradual, that the change will come slowly enough to be managed through normal organizational processes — may already be wrong.
Legal was supposed to be the test of that frame. The argument was that if any industry could resist the pace of AI adoption, it would be the one built most deliberately on human judgment, professional relationships, and institutional conservatism.
Legal did not resist. It became the #2 vertical in enterprise AI.
The timeline for every other industry just got shorter. Not because the technology is moving faster than expected — it is — but because the evidence that change happens even in resistant industries is now in the data.
The slowest industry to change changed quickly. Every industry sitting at under $150M in AI revenue is looking at what legal did in three years and recalibrating their timeline accordingly.
The capability curve does not wait for procurement cycles. It compounds while they run.