The Machine Can Draft. It Can't Decide.

Every few weeks a founder asks me some version of the same question: is AI going to replace lawyers? Usually there’s a link attached — a model that just passed the bar, a demo that drafts an NDA in nine seconds, a headline predicting the end of the profession. The tone is half worry, half hope. The worry is for me. The hope is for their legal bill.

I understand the question, and I think it’s the wrong one. It assumes the thing a good lawyer sells is the document. Automate the document and you automate the lawyer. But the document was never the product. It was the receipt.

The honest version of the question isn’t whether AI replaces lawyers. It’s which part of what lawyers do is legal work and which part is legal judgment — because those two things have been bundled and sold at the same price for a hundred years, and AI is about to pull them apart. The work is collapsing in cost. The judgment isn’t. What happens next depends entirely on whether the profession passes that collapse through to the people who need counsel, or pockets it.

The wrong question about lawyers and AI

Start with the framing, because the framing is where most of the panic lives.

“AI will replace lawyers” treats a lawyer as a function that converts a fact pattern into a document. Feed in the deal terms, get out a stock purchase agreement. If that’s the job, then yes, the job is in trouble, and it should be — nobody should pay $900 an hour for a task a machine does better and faster.

But sit in on the actual work and you see something else. A founder calls because a prospective acquirer sent a term sheet with a no-shop that runs ninety days and an exclusivity provision with teeth. She doesn’t need someone to explain what a no-shop is; she can read. She needs someone who has watched a dozen of these deals go sideways to tell her whether ninety days is normal or a trap, what it signals about the buyer’s seriousness, and whether pushing back costs her the deal or earns their respect. That’s not document generation. That’s judgment under uncertainty, with money and a company’s future on the line.

I made a version of this argument in my case for frontier optimism: the interesting question about a new technology is rarely “what does it destroy,” it’s “what does it finally make possible.” And in writing about AI agents and the law, I kept coming back to the same move — the actor is new, but the underlying question is old. The same move applies here. The tool is new. What clients actually buy from a lawyer is old, and the machine isn’t selling it.

Legal work was never the product

Here’s the uncomfortable thing the profession has avoided saying out loud: the traditional law firm is a pricing model wearing the costume of a profession.

The billable hour bundles two completely different things and charges for them as one. There’s leverage — the army of associates doing document review, first drafts, due diligence, and research, whose hours get marked up and sold. And there’s judgment — the partner who actually decides what risk is worth taking, which almost by definition can’t scale, because it lives in one experienced head. The genius of the model, from the firm’s perspective, is that it prices the scalable part like the unscalable part. A first-year’s contract markup and a senior partner’s read on a bet-the-company decision both bill by the hour, and both hours are expensive.

That worked because the two were welded together. You couldn’t get the judgment without paying for the labor, because the judgment was buried inside the labor — the partner formed her read by having the team turn over every rock. The bundle wasn’t a scam. It was a technology constraint. Producing good judgment required producing a lot of expensive work first.

That constraint is what AI dissolves. And once you can produce the work without the army, the bundle stops being a law of nature and starts being a choice.

The machine is good at exactly the expensive part

The thing that makes this moment different from every previous round of “legal tech” is what the machine is good at.

It’s good at the leverage. Draft this from our standard form. Read these forty vendor agreements and tell me which ones have change-of-control triggers. Summarize the diligence. Turn the term sheet into definitive docs. Find the market position on this clause across a thousand deals. This is precisely the work that used to require a room full of billable associates, and a well-built system now does the first eighty percent of it in minutes, at a cost that rounds to zero.

What it is not good at — not yet, and not in the way that matters — is the part you can’t outsource without also outsourcing the responsibility. It doesn’t have accountability; when the deal blows up, you can’t put a model on the phone with the board. It doesn’t have adversarial nerve, the willingness to be the one who says we walk. It doesn’t hold the relationship, the years of context that let a client trust a hard piece of advice they don’t want to hear. And it doesn’t own the judgment call, because judgment isn’t retrieval — it’s deciding what to do when the precedent runs out, which is exactly where I’ve argued the law itself gets interesting.

So the collapse is asymmetric. The cheap-to-produce, expensive-to-buy layer — the work — is going to zero. The expensive-to-produce, genuinely-valuable layer — the judgment — is not. A machine that’s brilliant at the leverage and useless at the decision doesn’t replace the lawyer. It strips the lawyer down to the thing that was always worth paying for.

Once the work is cheap, the bundle is a choice

I run Altum Legal as a fractional general counsel practice — four lawyers, embedded as senior counsel for frontier-technology startups and venture funds, going up against firms with hundreds of attorneys. We can do that because we did the thing this argument implies: we stopped organizing the practice around producing the work and started organizing it around the judgment.

The mechanics are their own subject — we built the firm a memory, so that four people can carry the recall and the positions of a firm many times our size, and so no precedent walks out the door when someone leaves. But the build isn’t the point of this essay. The point is simpler and, I think, more consequential than any of the tooling.

Once the cost of the work falls toward zero, the old bundle isn’t a constraint anymore. It’s a decision. A seed-stage company that generates no revenue could never buy a real general counsel — a competent GC plus even a small team runs well into seven figures a year, so young companies either go without and absorb the risk, or rent a big firm by the hour and ration their questions, which is its own kind of going without. The most consequential decisions a company ever makes get the least real counsel, purely because the judgment was priced like the labor it was buried in.

AI severs that link. The document assembly, the first-pass review, the research — the parts that used to require a billable army — now cost almost nothing to produce. The senior read is all that’s left to charge for. Which means access to good judgment is no longer gated by the cost of the work. That’s not a fact about technology. It’s an opening. What the profession does with it is the actual subject of this piece.

The failure mode is capture, not replacement

I’m an optimist about this, but optimism is not the same as assuming it goes well on its own. In the frontier optimism essay I argued the default outcome of a powerful new technology isn’t progress — it’s inertia, capture, and the quiet erosion of the capacity to build. The same risk lives here, and it’s worth naming plainly.

The comfortable path for the profession is to adopt AI internally, do the same work in a fraction of the time, and charge the client exactly what they charged before. The partner’s read still costs what it cost. The document still bills like it took a team a week. The only thing that changes is the margin, and the margin stays inside the firm. Under that scenario AI doesn’t democratize anything. It’s a productivity dividend for incumbents, and the founder who couldn’t afford real counsel last year still can’t afford it this year — she’s just being overcharged by a leaner operation.

That’s the failure mode. Not robots replacing lawyers. Lawyers using robots to preserve a pricing model the technology has already made obsolete. It’s the legal-services version of every incumbent that captured an efficiency gain and gave the customer none of it.

The alternative is a choice, and it has to be made deliberately: pass the collapse through. Price the judgment honestly and let the cost of the work fall to something close to what it now actually costs to produce. That’s the whole difference between building an AI-native practice and an AI-assisted one. Same tools, opposite theory of who the gains belong to.

The job gets more human, not less

There’s a quiet irony in all of this. The lawyers most afraid of AI tend to be the ones whose value was mostly the work — the volume, the turnaround, the sheer capacity to produce documents. The lawyers who should be thrilled are the ones whose value was always the judgment, because AI takes the least human part of the job and hands it to a machine, leaving them with more of the part that was human to begin with.

Strip out the document assembly and what remains is counsel in the old sense of the word: sitting with a founder the night before a hard board meeting, telling her what you’d do and why, being willing to be wrong with her and to own it. Reading a room. Knowing when the legally correct answer is the commercially disastrous one. The hard conversation about a cofounder breakup that no system will ever run. None of that compresses into a prompt, and I don’t want it to.

The evidence, for what it’s worth, points the same way: people who lean on AI past the edge of their own competence tend to do worse than people who don’t use it at all. The technology rewards the lawyer who knows, at a bone-deep level, that the model is not practicing law — it’s clearing the ground so the lawyer can. If you’ve spent fifteen years building judgment in a practice area, you’re sitting on exactly the asset this moment makes more valuable. Most lawyers with that asset don’t realize it yet.

Keeping the promise

The bar likes to describe itself as a helping profession, a service to clients, access to justice, counsel for those who need it. For a long time the economics quietly contradicted the story. Real judgment was a luxury good, rationed by the hour and reserved for those who could pay for the labor it was buried in.

AI dissolves the constraint that made that true. The work is getting cheap. The judgment is staying valuable. The only open question is whether the people who hold the judgment use the moment to widen access or to protect margin — and that question won’t be answered by the technology. It’ll be answered by the choices of everyone building on top of it.

I know which side of that I want to be on. The machine can draft. It can’t decide. Deciding well, for people who need it and couldn’t reach it before — that part is still ours, and now there’s no excuse not to share it.