Most lawyers describe their work the same way most professionals describe their work from the inside.
Every matter is different. Every client is different. Every fact pattern has its own wrinkles. You cannot automate what we do, because what we do is judgment applied to specifics, and the specifics are always new.
That description is partly true.
It is also incomplete in a costly way.
The bespoke parts of legal work are real. They are not, however, the whole work. A large share of legal production is scaffolding around judgment: gathering, organizing, classifying, drafting, checking, tracking, formatting, routing, summarizing, comparing, and preparing.
AI is going to compress that scaffolding.
The firms that understand the distinction will have an advantage. The firms that insist everything is bespoke will still see the scaffolding compress, but they will have less control over how.
The remembered work is not the whole work
The parts of legal work that feel most novel are the parts lawyers consciously remember.
The statute of limitations question with a wrinkle. The clause that interacts with a jurisdictional issue. The strategic call in active litigation where the right answer depends on facts, forum, opposing counsel, and client risk tolerance. The case where the normal pattern does not quite fit.
Those moments are real. They are also not the only thing that fills a week.
Much of the week is research, document assembly, discovery organization, intake, deadline tracking, routine correspondence, standard-form drafting, status updates, conflict checks, filing logistics, and matter hygiene.
That work is real. It is not trivial. It is also not equally bespoke.
Every profession overestimates its bespoke fraction because the bespoke parts are where attention goes. They become the stories. The scaffolding fades because, once mastered, it feels like background work.
Then the profession generalizes from attention to volume.
Engineering went first
I have spent fifteen years building software and have worked on AI infrastructure since 2018. Engineering made this mistake first. The broader argument is in AI Didn't Replace Software Engineers. Legal Is Entering the Same Phase, and the judgment problem shows up again in AI Makes Everyone Look Competent. That Is the Problem.
Senior engineers spent years saying real software work was too bespoke to compress. Architecture was bespoke. Security was bespoke. Edge cases were bespoke. Scaling behavior was bespoke. Domain modeling was bespoke.
All of that was true.
AI did not need to replace the bespoke parts to transform the profession. It compressed the scaffolding around them: boilerplate, glue code, first drafts, standard patterns, test scaffolds, small internal tools, routine debugging, documentation, and translation between familiar shapes.
The judgment stayed where it was, with the engineer who could specify the right thing, inspect the output, reject the plausible-but-wrong answer, and understand the consequences of a decision the model could not see.
The senior-to-junior gap did not disappear. In many teams, it widened.
Senior engineers with AI tools can move faster because they know what correct looks like. Junior engineers can produce more code, but they are also more likely to accept output that passes a shallow test and fails under conditions they have not learned to anticipate.
The tool amplified judgment. It did not replace it.
Legal has the same shape
Legal work has its own professional obligations, its own rules, and its own stakes. The engineering analogy should not be stretched past usefulness.
But the shape is similar.
The first legal work to compress is the scaffolding:
- document assembly
- discovery organization
- intake classification
- routine correspondence
- standard-form drafting
- deadline tracking
- matter summaries
- document comparison
- initial research synthesis
- status and next-action organization
The work that compresses last is the work where legal judgment is actually load-bearing:
- fact-sensitive strategy
- professional responsibility calls
- client counseling
- settlement judgment
- litigation posture
- credibility assessment
- statutory interpretation where the answer is not settled
- questions where forum, judge, client objective, and consequence all matter
This distinction is the practical one.
Legal work is not easy. The scaffolding around legal judgment is larger than many lawyers admit, and AI systems are now good enough to compress a meaningful portion of it.
Legal may compress faster than lawyers expect
There is one uncomfortable reason legal scaffolding may compress faster than some engineering scaffolding did.
Legal is often more convergent.
Engineering problems can have many valid architectures. Two qualified engineers may design very different systems for the same product problem, and both designs may be defensible.
Legal work is constrained differently. Statutes, rules, precedent, procedure, forum, deadlines, and burdens of proof narrow the solution space. Two qualified attorneys may disagree about strategy, risk, and emphasis, but on many questions they are working inside a more bounded framework.
That boundedness is exactly the kind of thing AI systems operate inside well.
Not perfectly. Not without review. Not as a replacement for the lawyer. But well enough to compress repeatable legal scaffolding faster than a profession that thinks almost everything is bespoke may expect.
What this is not saying
I am not arguing that lawyers are not skilled.
It is not an argument that anyone can practice law.
It is not an argument that AI should replace lawyers.
It is not even an argument that legal reasoning is simpler than engineering reasoning. Convergent does not mean easy. A constrained answer can still require deep training to identify correctly.
The argument is narrower:
Legal work contains more scaffolding than the profession usually admits, and that scaffolding is going to compress.
The lawyers who benefit will be the ones who can distinguish scaffolding from judgment and build workflows around that distinction.
The better classification
Do not ask whether your practice is bespoke. Every practice has bespoke elements.
Ask:
- which parts require my judgment?
- which parts require a record?
- which parts require review?
- which parts require a source?
- which parts require a client conversation?
- which parts are repeatable preparation for those moments?
That classification is where the work begins.
A demand letter may contain judgment about tone, leverage, facts, damages, audience, and timing. It also contains repeatable scaffolding: party identification, factual chronology, demand structure, deadline language, supporting documents, and next-step framing.
A patent office action response may contain judgment about claim strategy and prior art. It also contains repeatable scaffolding: issue identification, reference summaries, claim mapping, deadline tracking, draft argument structure, and review state.
A legal-aid intake workflow may contain judgment about urgency, eligibility, and referral path. It also contains repeatable scaffolding: issue classification, jurisdiction, income screen, document checklist, deadline awareness, and handoff routing.
AI belongs there first.
Not replacing the judgment. Compressing the preparation around it.
Where this leaves the profession
The firms that adapt will not be the firms that declare themselves "AI-first" the loudest.
They will be the firms that can name the difference between scaffolding and judgment, then build systems that compress the first without hiding or weakening the second.
That means matter records, source grounding, review states, approval boundaries, provenance, and workflows that make it clear when output is preparation and when professional judgment has actually been applied.
The bespoke work is real.
It is just smaller than the profession wants it to be.
That should not scare lawyers. It should clarify the work. Let the scaffolding compress. Protect the judgment. Build systems that know the difference.