The seventh revolution in law.
Legal work changes when something expensive becomes cheap. It has happened six times in two and a half thousand years, from the day Rome stopped keeping its law a priestly secret, and each time lawyers ended up more capable rather than less necessary. The seventh is happening now: for the first time what gets automated is not the finding but the thinking. It lands first on the expert-witness fight.
Seven revolutions, two patterns.
Software has compute per dollar. Legal work has an equivalent, and it has fallen at every revolution: what a single authoritative answer costs, in how far you went to reach the text and how long it took to get an answer out of it. The first has hit its floor. The second has not.
Distance stops falling in 2012 and cannot fall again. Time has not stopped: the day that yielded one answer in the 1870s yields hundreds now. Each dot is an estimate and each whisker the range it was reasoned within. Hover an era for how its two numbers were arrived at.
The second pattern is the spacing: the last three fall inside a single working lifetime.
Nine hundred and seventy-nine years separated the first two. Thirteen separated the last two.
Held by a priestly class, then cut into tablets and posted in the Forum under plebeian pressure. The rules did not change. Who could know them did.
A thousand years of scattered juristic writing compressed into one organized body, and the spine of the civil law ever since. Compilation is a technology.
Within a generation of Gutenberg, statutes and year books were in print, reaching every practitioner instead of whoever owned the manuscript.
West's National Reporter System and its key-number index made scattered decisions searchable. Finding the right precedent stopped being luck.
LexisNexis, then Westlaw. A week in a reading room became an afternoon at a terminal.
Predictive coding, approved in Da Silva Moore, moved first-pass review from rooms of associates to classifiers.
Frontier models crossed PhD-level reasoning: 92 to 94 percent on GPQA Diamond against 65 to 70 for in-field PhD holders. Every earlier revolution made the material easier to find. This is the first that does the analysis.
Every one made lawyers more capable, not less necessary. The work moved up a level each time, and the firms that moved first set the terms.
The expert-witness market has not had its turn. A first strategic read still takes weeks and a specialist’s hours. That is the part a machine can now do. Signing the report and taking the stand stay where they are.
Having it has never been the same as using it.
Each of those arrived long before it was general, and the gap is the part firms live through. The first two changed who was permitted to reach the law, and landed over generations. Everything after changed what reaching it cost, and landed in decades. The lag is now shorter than a single matter.
Solid bars reached a majority; hatched bars are estimates. The fall is real, with two interruptions: one that lasted five centuries, and one happening now. Hover a revolution for what happened in that gap.
Twice it stopped collapsing, and the second time is now. Predictive coding, defensible since 2012, is still named in 22 percent of responses on how electronic evidence gets reviewed; three quarters of litigators say the reason is that they do not know the technology. Machine reasoning moves faster: 30 percent report their offices using AI tools, 45 percent expect it mainstream within three years. Roughly a third of the profession has this. The other two thirds include the lawyers on the far side of the table.
The experts are already using it.
In August 2026 a Harris County jury returned $61.5M against 3M over the Watson Grinding explosion. The defense engineering expert, billing $475 an hour, had produced a report both sides put at 85 to 90 percent ChatGPT. Per trial reporting, discovery turned up roughly 350 pages of his prompts, one asking the model to show how 3M is 0% at fault.
The written rulings are more careful. A Stanford expert on AI and misinformation had his declaration struck for two fabricated citations. A valuation expert’s 172-page report, produced in 72 hours, was excluded as “written by artificial intelligence at the instruction of” the expert. A damages expert was excluded for relying on an analytics tool whose algorithm he could not explain. By September 2026 one academic tracker counted thousands of AI-hallucination cases across all filings.
The line courts are drawing is judgment against substitution, not AI against no AI. The opinion that struck the Stanford declaration did not fault the expert for using the technology to research, and a later court declined to exclude an engineer who wrote his report from experience and used a model to check it.
This is the seventh revolution arriving in this market: typed into a chat window at $475 an hour with nothing checking the output. Every design decision on this site, the verified quotes, the dropped citations, the refusal to produce anything that gets filed, answers that.
Where this goes.
Analytical capability that was scarce and expensive is becoming abundant and cheap, and it is not stopping at the level it reached in 2025. Three curves carry it: whether a machine can do expert-grade analysis, how much human work it can replace in one go, and what that work costs.
Graduate-level science that in-field PhD holders answer at just under seventy percent. Models went from thirty-nine to ninety-four in two years.
Epoch AI; baseline from the o1 recruitment study.
Four seconds of human-equivalent work in 2019, six minutes by GPT-4, fourteen and a half hours by Claude Opus 4.6: doubling every seven months or so. The dashed run carries that doubling to 2030; arithmetic, not a forecast.
METR Time Horizons 1.1, at 50 percent success, on software and research tasks rather than legal work. Log scale. METR reports its suite cannot measure reliably above sixteen hours.
From our own inference costs, with a range for what a mature run adds. Even at our ceiling a run is under a hundredth of an expert hour, and a firm pays $50 whether it makes eight model calls or eighty.
Floor measured from our model-call ledger, rounded up. Ceiling reasoned. Expert hour from SEAK’s 2024 fee survey, carried forward at the 4.2 percent a year it measured since 2021. Log scale.
Watch the horizon. A run today takes a few minutes, far below what today’s models already replace. The next layer, now in development, is a settlement engine: a matter argued out end to end across a thousand scenarios, agent against agent, your experts against the other side’s. It returns one strategic recommendation, a distribution of outcomes instead of one read. The top percentiles show where to press and what to build strategy from; the bottom ones show what to avoid before settlement. Work of that length becomes ordinary in the next few doublings.
None of this is new. It has been forecast for three quarters of a century, with dates that can now be marked.
In conversation recalled by Stanisław Ulam, described accelerating progress approaching “some essential singularity beyond which human affairs could not continue.”
The idea arrived before the field did.
Predicted that within fifty years an average interrogator would identify a machine correctly no more than seventy percent of the time after five minutes.
A 2025 study found GPT-4.5 judged human in seventy-three percent of five-minute exchanges.
Predicted a computer would take the world chess title before 1998.
Deep Blue beat Kasparov in 1997, a year inside the window.
In The Future of Law, predicted email would become how lawyers and clients mainly communicate.
It is how the profession runs.
None of that replaces lawyers. It moves the constraint. When analysis stops being rationed the binding constraint is judgment: which argument to run, which expert to retain, which matter to take. Those decisions improve when the analysis under them is complete rather than affordable.
Most matters buy no analysis at all.
The obvious objection is that a market cannot survive a price cut of three orders of magnitude. The counter-argument is the Jevons paradox: a falling unit cost expands what is worth doing faster than the saving shrinks the bill. The IRS counted tax software’s share tripling to a quarter of returns by 2003 while paid-preparer use rose to 62 percent. Both grew. What shrank was the number of people getting no help at all.
A kilogram to orbit cost about $54,500 on the Space Shuttle and about $2,720 on a Falcon 9, and the launch market grew rather than shrank: Starlink alone now approaches $12 billion a year.
That is not a law. Total spend falls where demand is capped outside the price, which is why cheap booking shrank travel agency. The number of matters is capped. The number of analyses per matter is not, and today it is usually zero. At $25,000 a read most matters get none. At $50 it runs at intake, against every expert the other side might retain, and again when the served report moves the facts. None of that comes out of an existing budget, because none of it happens today.
Market sizes are usually asserted rather than built. Ours has its derivation attached. For scale: IBISWorld puts the US law-firm industry at $405B in 2026, and litigation is a little under thirty percent of legal-services spending, so US litigation runs on the order of $116B a year. That is legal fees, not expert spend, and everything below sits inside it.
Damages, loss causation, market efficiency and accounting. Securities and antitrust are $1B to $2B of it.
Reasoned from CRA at $752M and FTI Economic Consulting at $850M, both global, and from twenty securities fee petitions: median $209,000 a case.
Medical, forensic and engineering testimony across tort litigation. The plaintiff half is estimated, not measured.
NAIC defense and cost containment expense of $26.4B (2023), a peer-reviewed 13 percent expert share, and two builds converging at $6B to $9B.
Firms sourcing and supplying experts, mostly inside the figure below rather than beside it.
IBISWorld OD4885, 2025.
Court reporting, IME networks, expert coordination.
Summed from parts, with overlap between census categories. Court reporting alone is about $3.2B.
Outside both: the firm's own hours, e-discovery, and the technical experts in patent disputes, which are a few hundred million rather than billions.
What matters most is in neither: the reads nobody buys at $25,000, which no survey counts because they never happen.
Where this ends is a statement of intent, so it lives on its own page: the fourth institution.