In 2016 a prominent computer scientist said we should stop training radiologists because deep learning would replace them within five years. Ten years on there is a radiologist shortage. The same forecast was made about lawyers, in the same tone, with the same confidence, and it was wrong in the same way.
The forecast was not wrong because the technology underdelivered. Image classification got extremely good. It was wrong because it modeled a profession as a task, and professions are not tasks. They are bundles of tasks wrapped in licensure, liability, institutional trust, and the requirement that someone identifiable be answerable when it goes badly. The technology dissolved the task and left the wrapper intact.
Any prediction about law that ignores the wrapper will be wrong the same way. So here are mine, each with the mechanism I think drives it and — the part legal futurism always omits — the specific thing that would prove me wrong. Check back and score me.
The frame
One distinction does most of the work.
The cost of producing legal work is falling fast. The cost of everything gated by rules, courts, insurers, and trust is not.
Drafting, research, review, summarization, extraction — the production layer — is dropping in cost by an order of magnitude over a decade. Meanwhile the requirement that a licensed human sign, that a court accept a filing, that an insurer underwrite the risk, that a client trust the advisor, moves at institutional speed, which is to say roughly not at all.
Almost everything below is a consequence of those two clocks running at different rates.
Prediction 1 — The billable hour survives, and its share falls sharply
By 2035, hourly billing remains the dominant model for genuinely unpredictable work — litigation of uncertain scope, crisis work, complex transactions — while defined-scope work has largely moved to fixed fees.
Mechanism. The hour is a proxy for value that works only when nobody can estimate the work. AI attacks estimability first: the more predictable a task becomes, the more absurd it looks to price it by duration. Add the ethics constraint — the ABA’s Formal Opinion 512 is explicit that if a task takes less time, you bill less time — and hourly billing on compressible work becomes a shrinking revenue line that a firm is professionally obligated to keep shrinking. Fixed fees resolve that contradiction by letting the firm keep the efficiency it created.
What would falsify it. Hourly billing holding steady or growing as a share of small-firm and consumer-facing revenue through 2030. If firms simply keep billing the same hours for less work, I am wrong about the constraint binding — and the access argument in this whole section collapses with it.
Prediction 2 — Document review as a staffed activity mostly disappears
By 2032, linear human review of large productions is rare. The reviewing lawyer’s job becomes designing, validating, and sampling a system’s output.
Mechanism. This is the most automatable high-volume task in litigation and it is already halfway gone. Technology-assisted review has been judicially accepted for over a decade; language models are dramatically better at it than the classifiers TAR was built on. What remains is not the reading — it is the defensibility: sampling protocols, recall estimates, privilege QC, and a lawyer who will certify the process.
What would falsify it. Courts rejecting model-assisted review protocols, or a high-profile privilege waiver traceable to automated review that makes the practice uninsurable. Either would freeze this for a decade.
Prediction 3 — Malpractice insurers regulate AI use before the bars do
By 2029, professional liability applications routinely ask detailed questions about AI verification workflows, and premiums differentiate on the answers.
This is my least obvious prediction and the one I would bet hardest on.
Mechanism. Insurers move at underwriting speed; bars move at rulemaking speed, which is years slower and consensus-bound. Insurers already ask about cyber practices, calendaring systems, and conflict-checking procedures, and they price them. They have every incentive to ask “do you have a citation verification process” long before any bar requires one — and unlike a bar, they can act unilaterally, immediately, and with a financial instrument rather than a disciplinary one.
The result is that the effective regulation of AI in practice arrives as a question on a renewal form. Firms that can document a verification gate get a better rate. That is a faster and, frankly, better-targeted mechanism than discipline, because it is prospective rather than punitive.
What would falsify it. Carriers treating AI as a general practice risk without workflow-specific underwriting through 2030 — pricing it like weather rather than like a controllable exposure.
Prediction 4 — Courts shift from “did you use AI” to “did you verify”
By 2030, most standing orders requiring AI disclosure have been replaced by rules requiring certification of verification.
Mechanism. The first wave of judicial responses after the fabricated-citation cases demanded disclosure of AI use. That framing does not survive contact with reality, for the obvious reason that it will shortly describe every document filed in every court — the same way “did you use a computer” stopped being a meaningful question. What courts actually care about is already covered by existing rules: Rule 11 requires a certification “formed after an inquiry reasonable under the circumstances,” and that language does not need amending to reach a lawyer who filed a brief they did not check.
Expect consolidation toward a standard that asks what was verified and how, not what tool was used.
What would falsify it. A durable, widely adopted disclosure regime — a model rule requiring affirmative AI disclosure that most jurisdictions adopt and keep past 2030.
Prediction 5 — Regulatory divergence widens before it resolves
By 2032, a handful of additional states permit some form of entity regulation or non-lawyer ownership; the large majority do not; and the resulting interstate friction becomes the central regulatory fight.
Mechanism. Arizona eliminated Rule 5.4’s ownership restrictions and licenses alternative business structures; Utah is running a sandbox on a fixed clock. Stanford’s five-year data shows the entities serving individual consumers and small businesses rather than corporate clients — which is exactly the population the access gap describes. That evidence will pull some states in. Most will not move, because the profession in most states does not want it.
The interesting pressure is not within any state, it is across them. An Arizona ABS delivering services to a Minnesota resident over the internet raises a question no one has a satisfying answer to, and that question — not the merits of ownership rules — is what forces the issue.
What would falsify it. Rapid convergence in either direction: a wave of adoptions after a favorable multi-state study, or Arizona and Utah retrenching and the experiment being written off.
Prediction 6 — The paraprofessional tier grows quietly and works
By 2033, licensed-paraprofessional programs exist in a substantial minority of states and handle a meaningful share of routine family, housing, and debt matters.
Mechanism. This is the least glamorous and most likely of my predictions, because it is already happening and because AI changes its economics decisively. The historical objection to limited-license providers was competence at the margins — a non-lawyer facing an unusual issue. A paraprofessional with good AI support and a defined escalation path is a materially different proposition than one working from a checklist.
Minnesota’s program is the template: it began as a pilot and was made permanent effective January 1, 2025, with supervision by a licensed attorney and a defined set of case types. Narrow, evidence-driven, unexciting, and durable.
What would falsify it. A repeat of Washington’s LLLT outcome at scale — programs launched and then sunset for low uptake. Note the trap I have written about before: these get judged against the ideal of full representation rather than against the actual alternative, which is nobody.
Prediction 7 — The bottom of the market gets served by something that is not a law firm
By 2032, a significant share of routine consumer legal need is met by products, not practitioners — and the fight is about unauthorized practice, not quality.
Mechanism. Ninety-two percent of the civil legal problems of low-income Americans currently get no adequate help. That is not a market being served badly; it is a market not being served. When the cost of producing competent guidance on a standard eviction defense approaches zero, something will fill it, and the binding constraint is not capability — it is UPL statutes written when the alternative to a lawyer was a stranger with a form book.
Note the shape of the coming argument. It will be framed as consumer protection. The honest version of the question is whether an imperfect product is worse than nothing, and the answer for the 92% is usually no.
What would falsify it. Vigorous and successful UPL enforcement against AI legal products, or consumer-harm evidence strong enough to justify it.
Prediction 8 — Leverage compresses, and small firms gain
By 2033, the associate-leverage model that funds large firms is materially weaker, and the cost advantage of small firms over large ones narrows or reverses.
Mechanism. Large-firm economics rest on billing junior time at a multiple of its cost, on work that is now the most compressible in the building — document review, first drafts, research memos, due diligence. Meanwhile a three-lawyer firm can now access research and drafting capacity that used to require an institution. The gap in capability per lawyer shrinks; the gap in overhead per lawyer does not.
This does not kill large firms. Bet-the-company litigation, cross-border transactions, and regulatory work with genuine institutional depth are not going anywhere. It compresses the middle, which is where most of the associates are.
What would falsify it. Large-firm headcount and leverage ratios holding steady through 2032, or firms successfully repricing junior work as supervision rather than production.
Prediction 9 — Verification becomes a named, billable function
By 2030, “verification” is a recognized role in litigation practice the way document review was in 2005.
Mechanism. If production is cheap and certification is not, value concentrates in certification. Someone has to establish that the authority exists, says what it is cited for, and remains good law; that the extracted date is the right date; that the summary did not invent an admission. That work is not compressible in the way drafting is, because it is precisely the check on the compressed output.
The interesting consequence is status. Verification is currently invisible, unglamorous, and mostly done by whoever is most junior. When it becomes the scarce input, it stops being junior work.
What would falsify it. Verification being absorbed entirely into tooling — automated checking good enough that no human function forms around it. Possible for existence and quotation; I do not believe it for “does this case stand for this proposition.”
Prediction 10 — Privilege and work product get litigated hard, and the answers are messier than anyone expects
By 2029, there is meaningful appellate authority on the discoverability of AI prompts, retrieved context, and model outputs in litigation.
Mechanism. Every one of these systems generates artifacts: the prompt, the retrieved documents, the intermediate draft, the transcript. Are they attorney work product? Are they privileged communications? Does a litigation hold reach them? Does sharing a client’s facts with a vendor waive anything?
These questions have plausible answers under existing doctrine and no authority, which is exactly the condition that produces expensive discovery fights. My weak prediction is that the answers land roughly where work-product doctrine already sits — protected as reflecting counsel’s mental impressions, with the usual fights at the edges. My strong prediction is that firms will be badly surprised by how much of this they retained without deciding to.
What would falsify it. The question being resolved early and cleanly by rule amendment rather than litigation.
Prediction 11 — The signature does not move
Through 2035 and beyond, every document filed with a court is signed by a licensed human who is personally answerable for it.
Mechanism. This is the wrapper. The signature is not a technical requirement that a sufficiently good system could satisfy — it is the mechanism by which the legal system locates responsibility. Courts need somebody to sanction. Clients need somebody to sue. Insurers need somebody to underwrite. Bars need somebody to discipline. A system with no answerable person breaks all four institutions simultaneously, and none of them will accept that trade for any level of accuracy.
I would bet against full autonomy in anything with a signature block more confidently than I would bet on any prediction above.
What would falsify it. Any jurisdiction accepting a filing not attributable to a licensed human. I do not expect to see it.
Prediction 12 — And the one that is not a prediction
Here is the thing none of the above determines, and it is the only one that matters.
Falling production cost does not automatically mean falling prices. It means falling cost. Whether that reaches clients depends entirely on whether competition, regulation, or professional conscience forces it through. If it does not, the efficiency becomes margin — the same work, the same bill, a better year.
That is not cynicism, it is the default. Efficiency gains flow to whoever has the pricing power, and in most legal markets that is not the client. A firm that halves its production cost and holds its rates has done nothing for access to justice. It has just gotten richer, which is fine, and is not what any of this was for.
So the actual open question about the next decade is not technical and not really predictive. It is a choice the profession makes one firm at a time: whether the gain shows up as a smaller bill and a case you could previously not afford to take, or as a larger distribution.
Everything in this section — the boards, the pipelines, the verification gates, the connectors — is only worth writing down if the answer is the first one. That is the whole reason it is published free, and the only thing this firm asks in return.
Scoring this
I have tried to write these so they can be checked rather than admired. If you are reading this in 2035, the ones I expect to have gotten wrong are Prediction 2 (I suspect I am early — defensibility is stickier than capability) and Prediction 5 (regulatory prediction is a mug’s game; divergence may simply persist without resolution).
The one I expect to be most right about is Prediction 3, and it is the one nobody is watching.
Sources
- ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512 (July 29, 2024)
- Legal Services Corporation, The Justice Gap (2022)
- Stanford Law School, Regulatory Innovation at the Crossroads: Five Years of Data on Entity-Regulation Reform in Arizona and Utah (June 2025)
- Minnesota Judicial Branch, Legal Paraprofessional Program
- Minn. R. Civ. P. 11.02
Speculative commentary on the future of the profession. Not legal advice, not ethics advice, and emphatically not investment or career advice. Predictions are the author’s opinion and are offered to be tested, not relied upon.