AI capability is years ahead of the institutions meant to use it. The business of the next decade is closing that gap.
Over the past few months, I have sat with partners at some of the largest law firms in the country and asked them to show me what they had actually tried to do with AI.
The pattern was remarkably consistent. A sophisticated lawyer, with twenty or thirty years of experience, would upload a document and ask the model to “review this agreement and flag issues.” The model would return a competent, generic, mostly useless answer. The lawyer would nod, because the answer confirmed the suspicion he had walked in with. Interesting tool. Fine for summaries. Not ready for real work.
This is the natural first instinct, since the box looks like a search bar that invites a one- to three-sentence query.
But then we rebuilt the instruction.
We did not add magic words. We did what a senior lawyer would do before handing the assignment to a strong associate. We explained the background on the client, the posture, the business objective, the counterparty dynamic, the provisions that usually matter, the issues that look legal but are really commercial, the arguments not to make, the level of confidence the lawyer was willing to stand behind, the format the client would actually read, and the checks the AI had to run before the answer came back.
Same model. Same document. Different instruction.
The output would change so much that the mood in the room often changed with it.
That is the gap everyone is missing. The model was not too weak. The institution had not learned how to absorb it.
For the past couple of months I have been working quietly with two of the largest and oldest law firms in the United States, helping them absorb AI into the daily work of their practice groups. These are firms with every structural reason to move slowly: enormous profits, powerful internal constituencies, deeply embedded workflows, clients who still pay the bills. And even there, the serious conversation has shifted.
The question is no longer whether lawyers can use AI to summarize documents. It is how to rebuild actual legal work around frontier models.
Two clocks are running, and they have fallen out of sync.
The first measures the progress of the technology. It ticks forward every few weeks: a smarter model, a longer context window, a better agent, a system that can take a messy file set and return work product that used to require a team. The second tracks the institutions meant to use the technology, and it moves the way institutions always move: through committees, approvals, pilots, policies, trainings, steering groups, and the quiet hope that nothing fundamental has to change before the next compensation cycle.
The distance between those clocks is the most important fact in business right now.
The public argument about AI mostly misses this gap because it is almost entirely an argument about the first clock. One side thinks AI is about to swallow the economy whole. The other thinks it is just overhyped, overly expensive autocomplete. Both arguments are too machine-centered. The story that matters more in mid-2026 is everything around the machine: the incentives, the habits, the pricing, the human work of changing how an organization does its job.
The bottleneck has moved. It is not the intelligence anymore. It is the absorption of it.
The companies building the technology raised enormous sums against the promise that it will remake the economy (and fast), and now they have to show the remaking is real. The companies meant to use it face customers demanding the savings everyone keeps promising and new “AI-native” rivals who are starting to take their work. Both sides need the same thing, and it is in desperately short supply: real capability, absorbed into how white-collar work gets done.
That absorption is the greatest business opportunity in professional services.
The Fast Clock
In the time it takes a large firm to schedule a committee meeting about AI, two generations of new frontier models ship. Each one feels incremental to the firm because it arrives inside the same chat box as the last one. The interface barely changes, so people miss the scale of what changed underneath.
A lawyer in 2016 would have understood the current AI frontier as science fiction. A model can read a record, split a hard problem into subparts, work those subparts in parallel, search through a file set, manipulate documents, write code, run that code, check citations, and return finished work product without any human involvement beyond the initial prompt. Show that to a lawyer ten years ago and the demonstration would have ended in an emergency meeting of the executive committee. Show it to a lawyer today and he asks whether the firm’s IT department has approved the tool.
The software industry is the easiest place to see the fast clock in action because code either runs or it does not. Inside Anthropic, Claude now writes more than eighty percent of the code that ships to production, and the median researcher, polled in March, put his output at four times what it would be without AI.¹ Clive Thompson interviewed some seventy engineers across Google, Amazon, Microsoft, and Apple and found the same shape everywhere: the senior person writes less, directs more, and ships far more than before.² The unit of work has moved from production to orchestration. The human is still responsible, but the human is no longer typing every (or almost any) line of code by hand.
But law has no compiler. A contract that is wrong does not crash. It sits in a drawer, apparently fine, until the day a counterparty exercises a consent right no one had adequately considered, or an indemnity clause creates uncapped liability for an unsuspecting client. That makes legal AI harder to evaluate than coding AI, but I can tell you firsthand it is no less powerful. ³
My engineer friends run six months to a year ahead of the rest of white-collar work in how seriously they use these tools, and much of what I now teach lawyers I learned by watching them. In my own practice and in my consulting work, I’ve watched the shift the tech industry just lived through begin in small pockets of the legal profession. A litigator turns a day of research into twenty minutes. A deal team compresses a week of document review into an afternoon. A solo lawyer takes on work that used to require a floor of associates beneath her.
Some of these lawyers sit inside the largest firms in the world, building things their own partners have not noticed and would not believe. Many are in smaller practices with no committee to ask: solos tearing their workflows down to the studs, boutiques built around these tools from the beginning, lawyers who can change the work because they do not need permission from the institution the work threatens.
The fast clock is not waiting for the slow one.
The Slow Clock
Walk the halls of a standard AmLaw 50 firm and you will not, for the most part, find lawyers running their practices through frontier models.
You will find expensive legal-AI subscriptions. Approved tools. Vendor trainings. Responsible-use policies. Innovation awards. Partner-retreat panels where everyone agrees that AI is important and nobody says exactly which workflows should change.
Ask lawyers inside a big firm what they are using AI for today, and you will mostly find them using some of the most powerful technology ever built to clean up time entries, summarize documents nobody plans to read, and draft emails scheduling their next meeting. Trivial uses of a serious tool.
The capabilities that would matter, the ones the models have grown into, go untried: substantive delegation, briefing the model the way you would a good associate, context and standards and judgment calls spelled out, and getting elite work product back that would have taken days.
The usage is timid even where the capability is not.
The Incentive Trap
The slowness is understandable, which is not the same as defensible.
A large law firm’s profits rest on two pillars: the billable hour, which charges for time, and leverage, which stacks junior lawyers beneath every partner and bills their hours at a markup. AI threatens both. Every hour it saves is an hour that cannot be billed the old way. The work it does best, first-pass drafting, diligence, document review, cite-checking, summarization, comparison, formatting, is exactly the work the BigLaw pyramid exists to sell.
So the rational partner experiments privately. The rational firm moves slowly. Both are protecting something real.
That is the innovator’s dilemma in its cleanest form. The firms with the most to gain from the rebuild are the ones whose current economics make the rebuild most painful. They wait, and the waiting is rational until it is fatal.
The people who could force the change often have the least reason to. A law firm pays out its profits every year. A partner’s draw is a share of what the firm earned this year, not a claim on the next ten. A public-company CEO who transforms his business is paid in stock, which prices in future earnings the moment the market believes the story. A managing partner (sometimes making ten to twenty-five million dollars in salary a year, five years from the end of a long career) who transforms his firm gets disruption now, a compensation fight now, lower billable volume now, and a payoff that may land after he is gone. Running out the clock pays him. Fixing it pays his successors.
The slow clock also runs on fear.
First there is the asymmetric fear of becoming the cautionary tale. The partner who quietly rebuilds a workflow gets a polite nod. The partner whose AI filing cites fake cases gets a headline that follows him for the rest of his career. Sullivan & Cromwell learned that this spring, when an emergency motion in a bankruptcy case went out with a host of AI-generated citation errors in it.⁴ S&C is no one’s idea of a careless firm. Which is the point. Prestige does not prevent this failure. Process does.
There is also the quieter fear, the one every lawyer has scrolled past a hundred headlines about, that AI is coming to replace them altogether. The fear is not irrational, given that lawyers keep hearing this narrative from the people building the technology. Dario Amodei, Anthropic’s chief executive, went on the record last year warning that AI could wipe out half of all entry-level white-collar jobs, including in law, within five years.⁵ I think he is wrong about lawyers, and I will come back to why. But a partner does not have to believe the prediction to feel its pull. Seen from inside a big firm, every serious use of the model can look like a rehearsal for your own replacement: teach the machine the work and you have taught it your job.
So firms, by and large, retreat into AI theater. A task force. A policy. A pilot. A vendor. A speech about “responsible innovation.” More than once in the past few months I’ve had the opportunity to present on panels alongside leaders of top law firms that call their AI programs “best in class” and then have no coherent answer to the only questions that matter: which workflows changed, how much faster did they get, what improved for the client, and what does the firm now do differently on a live matter?
Generality is always the tell. A firm that rebuilt a workflow would talk about the workflow.
The Driveshaft
This has all happened before. When electricity replaced steam in the factory, the factory owners did the obvious thing: they pulled out the steam engine, dropped an electric motor in its place, and ran the machines off the same long central driveshaft. For nearly thirty years, plants kept that layout, as if the power still came from a furnace in the basement.
The productivity gains economists kept waiting for arrived only when, a generation later, factory owners tore up the floor and rebuilt the assembly line around the new power source, putting a small motor on each machine and letting the line follow the task instead of the shaft.
The problem was not that electricity was not overhyped; it was that a general-purpose technology pays off only when someone redesigns the work around it, and the redesign can run a generation behind the invention.⁶
AI is at that stage now. The new motor is, at best, bolted into the old driveshaft, and the floor underneath it is still the one built for steam.
Coca-Cola, Not General Electric
When mechanical refrigeration became cheap and reliable in the early twentieth century, the obvious bet was the companies building the refrigeration machines: General Electric, Westinghouse, Frigidaire. But the biggest winner was none of them.
It was Coca-Cola, a regional fountain syrup company out of Atlanta, which under Robert Woodruff set out in the 1920s to put its product, in his phrase, within arm’s reach of desire, in every town on earth.⁷
Coca-Cola never built a refrigerator. It understood, earlier and more completely than anyone else, what cheap cold made possible, and rebuilt itself around that understanding until a cold Coke became a permanent fixture of human life.
The frontier labs are the General Electric of this moment. The thing they make, raw intelligence, is getting cheaper at a rate with few precedents; set against the human hours a unit of it replaces, it is close to a rounding error per task.
But the Coca-Cola fortune will go to whoever works out, before anyone else, what the “cold” is for, and builds something on top of it that was not possible at any price the year before. That lane is wide open right now, in every industry at once.
Kirkland’s Bet
Kirkland & Ellis announced in May that it would spend $500 million over three or four years building its own AI platform.⁸
That number (which garnered all the headlines) matters less than what it reveals. The highest-grossing law firm in the world has concluded that renting the same tools (e.g. Harvey, Legora, etc.) as everyone else cannot protect what it has built. Hard to argue. A subscription available to every firm cannot be what sets one firm apart, and the change underway in the practice of law is too large to meet with a license key.
Kirkland is also more exposed than most, and the exposure comes from the same place as the profits. Last year the firm cleared $10.5 billion in revenue and $11.1 million in profits per equity partner, both records.⁹ Those profits rest disproportionately on private equity, the wrong client base to have when production gets cheap. Sponsors run the same deal structures dozens of times a year, track legal spend to the basis point, and have begun asking why work a machine can draft still bills at associate rates. Work that repeats is the work an AI model learns fastest. Even Blackstone, the marquee relationship, has started paying the firm less.¹⁰
Private equity is squeezing from the other side too. Blackstone and Bain Capital money now sits behind Norm Law, an AI-native legal platform that recruited the former chair of Sidley Austin’s executive committee as its chairman.¹¹ The industry that made Kirkland the most profitable law firm in history has started funding its challengers. Kirkland can read its own market. The first product out of the half-billion-dollar program arrived a week after the announcement itself, a fund formation engine for the firm’s private equity clients.¹²
But the size of the check will not decide the outcome.
A proprietary platform is worth exactly as much as the changed practice it is wired into. If Kirkland spends half a billion dollars and rebuilds how its lawyers actually work, the investment could become a moat no competitor can rent. If it spends half a billion dollars and leaves the workflows intact, it will have installed a very expensive motor into the old driveshaft.
The hard question is not whether Kirkland can build or buy powerful technology. It obviously can, but procurement is not the same thing as absorption. The hard question is whether a firm that profitable can force itself to change the work that made it profitable in the first place. That is the question every incumbent faces.
The Absorption Business
If absorption is the constraint, the most valuable asset in the market is whatever moves capability from the fast clock to the slow one without breaking the institution on the way. Someday that might be a product. Today it is usually just a person: someone who knows the work well enough to do it the old way and the tools well enough to rebuild it the new way, sitting inside the firm while the rebuild happens. Almost no one is doing this job, and almost everyone is about to need it.
The technology industry already has a job title for this person. Palantir invented it twenty years ago and called it the “forward-deployed engineer,” someone who moves into the customer’s operations and rebuilds the work around the software, because the software never deploys itself. For most of that time the role looked like a Palantir eccentricity. This spring it became the position everyone with money is copying. OpenAI stood up an entire deployment company around it in May with more than four billion dollars behind it. Anthropic launched an AI-native services firm with Blackstone, Goldman Sachs, and Hellman & Friedman to embed its engineers inside client companies. The sellers of intelligence have concluded that capability without absorption produces nothing, and that absorption is a person’s job.
But notice whom that person works for. A forward-deployed engineer works for the vendor. For most businesses that is a fine trade. A factory can run its logistics on the same vendor platform as every competitor, because logistics was never the edge; the widgets were. A law firm has no widgets. Its work runs on client confidences. The platform underneath the work is one every competitor can rent. And its procedures encode the firm’s own method. Let the lab’s engineers write that method on the lab’s rails and it tends to migrate into the lab’s product, where the firm next door can subscribe to it. For a law firm, that person should work for the institution rather than the vendor, and sooner rather than later.
None of this means the vendors have no role. Anthropic, Palantir, Snowflake, and their peers may well end up building the data architecture a firm’s rebuild runs on. But the work above the plumbing belongs to lawyers, because a software company has no more idea than anyone else outside the firm how to build the prompts and workflows that encode that accrued know-how of the practice itself. A week before this essay was published, Palantir’s own chief executive, Alex Karp, spent a CNBC interview telling enterprises to own “the means of production” behind their AI rather than rent them. He is selling something, of course. But he is also right.
That is why change management, the least glamorous phrase in business, is about to become one of the most valuable kinds of work there is. Not the old kind of change management that produces stakeholder maps and adoption dashboards, but a new kind that turns expert judgment into automated procedures a machine can run and an institution can trust. Every workflow rebuilt makes the next one cheaper to rebuild, every partner converted converts others, and the firm that started eighteen months early is, by the time anyone notices, a different kind of firm.
For a century, an institution that knew it had to change called the management consultants, stood up a steering committee, and commissioned the roadmap. Firms are running that play on AI right now, and it is the wrong play. It worked, when it worked, because the changes it managed were organizational: reporting lines, cost structures, which division to sell. A smart generalist from McKinsey can map all of that from across a conference table.
But the change everyone wants from AI is not “organizational.” It lives down in the practice itself, in the thousand small decisions that make up a single matter: does the model produce the first markup or only an issues list; what does it need to know about the client’s borrowing base before it touches the covenants; which of its case citations does a human re-pull and which get spot-checked; when does the partner read every word, and when does she read the exceptions memo and go home. A management consulting firm can’t answer those questions because the answers themselves constitute what is quickly becoming the 21st century’s version of the practice of law. Only the people who do the legal work can properly redesign that work.
It still takes a push from the top. A partner will not spend a hard week rebuilding how he practices unless the firm has made clear that this is a strategic imperative, not a hobby. But the push only sets the direction. The rebuild happens at the individual lawyer’s desk, one workflow at a time, and it looks nothing like an “innovation” program.
The Workshop
Here is what the absorption business looks like.
A partner brings in the kind of task that already fills his week: a contract to review against a client’s business preferences, a term sheet and cap table that need to become financing documents, a research question where the law is unsettled and confidence levels matter, a redline from opposing counsel he has to explain to a client in plain English without flattening what the changes do to the deal.
Most lawyers hand the model a document and a simple command. Summarize this. Review this. Research this. Fix this. Then they look at the predictably generic answer and conclude the tool itself is generic. But the model did what they asked. The lawyer gave it the task and withheld everything that would have allowed it to perform well: the context, the detail, the posture, the judgment.
A serious instruction carries what a good lawyer would tell a good associate: what matters, what does not, what the client is worried about, what the audience will notice, what the answer must not assume, what level of uncertainty is acceptable, and what to verify before the work product leaves the building.
In my workshops, I teach that as an anatomy: task, background, judgment, constraints, deliverable, verification. Once lawyers see the anatomy, they recognize the failure in their own prompts. They asked for a “summary” when they needed a client-ready explanation of what moved in the redline, what held, and what still needs a decision. They asked for “research” when they needed a bottom-line-up-front memo that separates settled law from open questions and independently verifies every citation. They asked for “contract review” when they needed a ranked list of the provisions that change negotiating posture, with proposed counter-language for each.
Notice what that anatomy does not contain: anything technical. No code, no syntax, no settings. Every word of a serious instruction is plain English, the same English a partner already uses across the desk from an associate. This is a mindset, not a software skill: treat the model like a brilliant new associate who has read everything and knows nothing about your client, and brief it accordingly. The lawyers who take to this fastest are rarely the youngest or the most technical. They are often the best delegators, the ones who spent a career learning to move judgment down the table.
That is the prompt layer. It matters because it proves the tool can do real work when the lawyer gives it real instruction. But for a firm, the prompt layer is only the beginning.
The question is how an institution built around human labor absorbs a technology that will make a growing share of that labor cheap, fast, and abundant.
A firm cannot meet that problem just by teaching isolated lawyers to write better prompts and hoping the change spreads. The work has to move from individual instruction to institutional capability.
That is where prompts become automated workflows.
A prompt tells the model what to do on this matter. A workflow is a procedure that tells the model how a particular lawyer, practice group, or firm does that category of work. In practice there is nothing exotic about it. A workflow is a plain-English playbook the model reads before it starts, the standing instructions a partner gives a new associate on day one, except the model follows them on every matter, every time, without being reminded. Some procedures are mechanical: how to mark up a Word document without corrupting it, how to run citation verification as a separate pass, how to validate formatting before anything goes out. The more valuable ones are substantive: how a particular lawyer reviews a contract, which provisions she always checks, when she asks for more facts, when she proposes counter-language, when she rejects the premise because the record does not support it, where she slows down because experience has taught her that this is where mistakes hide.
That is at the heart of what most legal-AI products still miss. They package a model in a legal interface and ask the lawyer to conform to the product’s workflow. The better approach runs in the other direction: make the model conform to the lawyer’s practice. The durable asset is not the wrapper. It is the firm’s own method, written down precisely enough that the model can follow it, lawyers can supervise it, and the institution can improve it over time. Which is exactly why it should not live inside someone else’s product.
A good contract-review workflow is not a template. It is the partner’s review sequence: termination, liability caps, indemnification, discretion standards, IP ownership, survival, issue-by-issue recommendations, actual counter-language, and a final pass asking what would embarrass the lawyer if the client saw it. A lessons-learned file is not knowledge-management theater. It is the compounding mechanism. The model misses something, the lawyer corrects it, and the correction becomes part of the next run.
The hard part is not technical. The files are mostly plain text. The hard part is getting senior lawyers to externalize the thing they do almost unconsciously: the sentence they distrust, the case they never cite without checking, the clause they read twice because it once cost a client money, the commercial point that matters even though it is not the doctrinally interesting one. None of that appears cleanly in the final work product. It has to be extracted while the work is happening.
That is why the firms most serious about AI adoption are not treating this as a novelty. The AI wave is not going to stop at better summaries. It is going to press on staffing, pricing, training, quality control, client expectations, and the internal distribution of power between people who can work this way and people who cannot. The firms moving seriously are trying to turn their best lawyers’ instincts into infrastructure before that pressure arrives all at once.
That is the market split I am seeing up close. Some firms are still approving software. Others are preparing to absorb a new production function into the practice of law.
The second group is going to be very hard to catch.
Either Way
As an aside, nothing in this essay requires the fast clock to keep ticking at its current rate. Even if, as naysayers suggest, we are in the middle of an enormous AI bubble and the progress stops tomorrow (which, for the record, is not what it looks like from here), models at the level of Opus 4.8 and Fable 5 are already more than enough to transform how law is practiced, and institutions will have to absorb that capability to keep up regardless. And if the clock does keep ticking, the question only gets sharper: when the machine can produce nearly everything, what is left for lawyers?
My answer is that the high end of the law will not commoditize, and its lawyers will not get replaced. As I have written elsewhere, the premium will move from execution to judgment, and the faster the clock runs, the larger that premium gets.¹⁶
The Invoice Fiction
At the end of all this transformation, the human part that will survive is the judgment call.
Not “judgment” as a professional compliment lawyers pay themselves. The actual call: which risk matters, which fight is worth having, which concession looks harmless but will hurt later, which argument a court might accept, which point the client thinks is legal but is really commercial. This is ultimately about decision-making under pressure and uncertainty.
That is what clients have been trying to “buy” from elite lawyers all along.
Yet the invoice made that hard to see.
For a century, law firms billed for the visible, quantifiable work: research, drafting, diligence, cite-checking, redlines, signature pages, closing sets. Some of that work required real skill. Much of it was grind. All of it took time, and time was easy to measure, so time became the unit the profession sold.
That unit had a logic. The work had to be done by people. The juniors who did it learned by doing it. The partners who supervised it converted the juniors’ time into margin. The client paid because there was no other way to get the deal closed, the brief filed, the diligence finished, or the record reviewed.
But the billable hour also blurred the distinction between production and decision-making. Production is the skilled work of gathering, drafting, comparing, summarizing, formatting, checking, and organizing. Decision-making is the moment when a lawyer takes all of that material and tells the client what to do.
Clients paid for both, but they cared most about the second.
A board does not hire the elite deal partner because it wants more diligence hours. It hires her because she has seen enough deals to know where this one can break. A defendant does not hire the great trial lawyer because he wants more time spent on rote discovery. He hires him because he wants someone who can decide which three points matter and how to make them land. A founder does not hire a lawyer to admire a markup. She hires a lawyer to say: concede this, fight that, and do not let them take this clause because it will matter later.
AI changes the economics because it attacks production first. It drafts the first pass. It compares the documents. It summarizes the record. It checks the citations. It conforms the signature blocks. It runs the tedious review that used to justify a large share of the bill. Imperfectly, and not without supervision, but fast enough and well enough that the old relationship between time and value can no longer hold.
When production was expensive, the client’s bill was full of production. When production gets cheaper, the scarce input is the person who knows how to direct the machine, test the answer, understand the client’s objective, and make the recommendation when the answer is hard. The value moves toward responsibility: the human being who has enough experience to know what matters and enough accountability to stand behind the advice.
This is where the predictions that AI will “replace” lawyers go wrong. Given what I am watching lawyers do with these tools, and given that judgment is the input that holds its value, the likelier arithmetic is that half of each legal job gets replaced, as opposed to half of the legal jobs being replaced altogether. The machine takes the production half. The part people actually went to law school for stays, and at the high end the job gets better, albeit more intense, because more of the week gets spent on the difficult cognitive work that was always the point. The low end of the legal services market is a different story: where the stakes are low and the work is rote, legal services may truly commoditize, and the margin will compress toward the price of the tokens that process client requests. But for the most elite law firms, like the ones I am working with now, the judgment will remain valuable and the work will not commoditize.
The transition will be hard for juniors because the grind was not only the thing firms sold. It was also how lawyers learned. First-pass research, first-pass drafting, first-pass review, diligence, checklists, and closing mechanics were not glamorous, but they created repeated exposure to the raw materials of judgment. If AI compresses that work, firms cannot pretend that the old apprenticeship will keep functioning on its own. They will have to design training around decision-making deliberately.
The best juniors will move faster than ever. They can ask the model to explain the whole deal instead of grinding through their piece in the dark. They can see the structure, test their instincts, compare alternatives, and get closer to partner-level reasoning earlier. The weaker ones will lose the camouflage that volume used to provide.
Recruiting will need to reflect this new reality. For decades, firms hired for law school grades and law review membership, proof that a candidate can follow instructions and survive brutal hours, because the pyramid ran on volume and the volume had to be survivable. But notice that this is not the same as (or even necessarily particularly correlated with) being a good lawyer. The associate worth hiring now looks different: early indicia of judgment, agency, commercial savviness, and people skills. A clerkship, where a young litigator spends a year watching a judge make difficult decisions. Time (in a non-legal capacity) inside a bank or a Fortune 500 company, where a future deal lawyer learns how clients talk and what they are actually trying to buy. Classes will get smaller. The bar will move. The associates who clear it may get something their predecessors rarely did: a more direct apprenticeship in decision-making, with more of the grind delegated to a machine that never wanted a weekend anyway.
Step back and the fiction comes into focus. What firms bill for today, the hours of junior and mid-level associate time doing grunt work, was never the thing clients actually valued from top firms. The hours were how the firm chose to invoice what the client actually wanted to buy, which is judgment and decision-making from the partner who signs the advice. So the lawyers who worry that ever more powerful AI will drain the value out of law have it backwards. The value always sat, perversely, in the one thing the bill never itemized, and that thing is not going anywhere. Concentrated judgment is an asset, the best firms own more of it than anyone, and the mature response to this technology is to protect that asset and finally price it, while the machine commodifies the part of the bill that was always just packaging.
Tearing Up the Floor
Everything in this essay points at the same unglamorous conclusion. Change management, done at the level of the practice, is now the most consequential investment available to any law firm (or large business enterprise, for that matter), bigger than any lateral hire, any practice launch, any office opening. The upside of getting it right is a compounding lead measured in years. The downside of getting it wrong is existential: a decade spent defending an hourly bill for work clients can buy more cheaply elsewhere, while AI-native competitors, funded in part by those same clients, pick off the work one practice area at a time.
The time to face it is now, while it is still a choice. Institutions change on one of two schedules, deliberately or in an emergency, and everything about an emergency makes a rebuild worse. The talent is leaving, the clients are renegotiating, and the executive committee is meeting about a competitor’s announcement instead of its own plan. The firms that start now get to rebuild while revenue is still setting records. The firms that wait will do the same work later, under pressure, with less of everything.
Investing in the rebuild means what it meant in the factories a century ago. The winners did not stop at swapping the steam engine for an electric one. They put a motor on every machine and let the line follow the task. The legal version is change at the level of the practice itself: judgment written down where a machine can run it and a lawyer can supervise it, workflow by workflow, group by group. The work is slow, personal, and invisible from the org chart, and it is the only kind of AI spending that changes what a firm actually does.
For the firms that can pull this off, the ultimate prize is the Coca-Cola fortune. Producing excellent legal work has always meant paying for floors of associates, and that cost is collapsing. A firm that moves off the hourly model and resets what clients expect to pay for keeps the thing clients will always want to buy (judgment, decision-making) and sheds most of the cost of making it. Margins at the high end will dramatically improve. Woodruff wanted a Coke within arm’s reach of desire. The firms that tear up the floor first will put elite legal judgment within arm’s reach of every hard decision in every business enterprise in the world.
Notes
- Anthropic Institute, “When AI Builds Itself” (Marina Favaro and Jack Clark, June 4, 2026), reporting that more than 80 percent of the code merged into Anthropic’s production codebase as of May 2026 was authored by Claude, up from low single digits before Claude Code launched in February 2025. The four-times figure comes from the report’s March 2026 internal poll of roughly 130 research staff, in which the median respondent put their output at about four times what it would be without AI; the report itself cautions that self-estimates of this kind tend to run high.
- Clive Thompson, “Coding After Coders: The End of Computer Programming as We Know It,” The New York Times Magazine, March 2026. Thompson interviewed more than seventy software developers at Google, Amazon, Microsoft, Apple, and elsewhere about how AI agents have changed the job.
- Zack Shapiro, “The Claude-Native Law Firm,” published on X, February 27, 2026: a firsthand account of running a two-lawyer practice rebuilt around frontier models.
- Letter from Andrew Dietderich of Sullivan & Cromwell to Chief Judge Martin Glenn, U.S. Bankruptcy Court for the Southern District of New York (April 18, 2026), in the Prince Group chapter 15 proceedings, apologizing for an emergency motion filed April 9, 2026 that contained dozens of inaccurate citations and other errors, including AI hallucinations. The errors were flagged by opposing counsel at Boies Schiller Flexner and widely reported, including by Bloomberg Law and Reuters.
- Dario Amodei made the prediction on the record in a May 28, 2025 interview with Axios’s Jim VandeHei and Mike Allen: AI could eliminate half of all entry-level white-collar jobs and push unemployment to 10 to 20 percent within one to five years. Amodei has since reached for gentler economics himself, invoking the Jevons paradox (automate most of a job and demand for the remaining human part can grow) onstage with JPMorgan’s Jamie Dimon at an Anthropic financial-services briefing (Fortune, May 5, 2026).
- The electrification story is told in Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review 80, no. 2 (1990), and Warren D. Devine, Jr., “From Shafts to Wires: Historical Perspective on Electrification,” Journal of Economic History 43, no. 2 (1983). Factories began electrifying around 1900; the measured productivity payoff arrived in the 1920s, once unit-drive motors let plants abandon the central-shaft layout.
- The phrase is Robert Woodruff’s, the longtime Coca-Cola president who set the company’s goal in the 1920s of putting a Coke “within arm’s reach of desire.” The line is quoted in the company’s own corporate histories and in Mark Pendergrast, For God, Country and Coca-Cola (1993).
- Kirkland & Ellis’s plan, first reported by the Financial Times and confirmed by Bloomberg Law in late May 2026, commits roughly $500 million over three to four years, beginning with about $100 million in 2026. Kirkland reported $10.56 billion in 2025 revenue, the highest of any law firm.
- Kirkland’s 2025 results, first reported by The American Lawyer in March 2026: gross revenue of $10.56 billion, up 20 percent, and average profits per equity partner of $11.1 million, also up 20 percent, across 595 equity partners. Kirkland was the first firm to clear $10 billion in revenue and the first to clear $11 million in average partner profits.
- Bloomberg Law and Law.com, February 27, 2026, reporting from Blackstone’s annual securities filing: Blackstone paid Kirkland $87.8 million in legal fees in 2025, down from a record $101.3 million in 2024, even as Kirkland’s overall revenue grew 20 percent. Blackstone discloses the payments because a Kirkland partner sits on its board.
- Norm Law launched in November 2025 alongside a $50 million Blackstone investment in its parent, Norm Ai, whose backers include Bain Capital, Blackstone, and Vanguard; in January 2026 it named Michael Schmidtberger, who had chaired Sidley Austin’s executive committee for seven years, as its chairman (Bloomberg Law, January 22, 2026). Venture funding for AI-native law firms has followed the same pattern: Crosby has raised more than $85 million from Sequoia, Index, and Lux, and Eudia raised a Series A of up to $105 million before launching an AI-augmented law firm in Arizona.
- Kirkland & Ellis and Palantir Technologies announced the platform on June 4, 2026, one week after the Financial Times first reported the firm’s $500 million commitment. The fund formation engine, exclusive to Kirkland, is built to carry fund documentation, side letters, obligation tracking, and closings across the private equity fundraising lifecycle for the more than 1,000 lawyers in the firm’s investment funds practice. Kirkland has said the platform’s architecture is model-agnostic, designed so the firm is not locked into any single AI provider.
- OpenAI announced the OpenAI Deployment Company on May 11, 2026, a standalone unit with more than $4 billion in committed capital led by TPG, launched alongside the acquisition of Tomoro, an applied-AI consultancy that brought roughly 150 forward-deployed engineers on day one. Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced their AI-native enterprise services firm on May 4, 2026, reportedly capitalized at about $1.5 billion (CNBC, May 4, 2026), with Anthropic Applied AI engineers embedded in its teams. The scramble follows a year of evidence that capability alone was not converting; MIT researchers reported in 2025 that, despite tens of billions of dollars in enterprise spending, 95 percent of organizations were seeing no measurable return on generative AI.
- Alex Karp, interview on CNBC, July 1, 2026, given alongside the announcement of Palantir’s sovereign-AI partnership with Nvidia. Karp said enterprise customers want to own the means of production behind their AI, their compute, models, data, and competitive edge, and dismissed vendor deployment ventures as arrangements that transfer that edge to a third party.
- Zack Shapiro, “The Input Layer,” published on X, March 25, 2026, on why the model’s output is only as good as the briefing it receives.
- Zack Shapiro, “The Judgment Premium,” published on X, March 2, 2026. The argument: as AI absorbs skilled production, the intelligence premium evaporates and the professional premium migrates to judgment, the layer where a person decides what to do when the answer is not clear and stakes a reputation on the call.
