Can AI Replace Lawyers? How AI Is Changing Commercial Legal Work

No: AI will not replace lawyers, though it is already replacing specific tasks inside legal work. “Will AI replace lawyers” is really two questions: whether the role disappears, which it does not, and whether the work changes, which it already has. Contract review, clause extraction, first-draft generation and research triage all run faster with AI than without it. Professional liability does not move: a lawyer still signs off on risk, runs the negotiation and advises the business. The job is being re-cut, not removed.

The split runs task by task, and it holds across the measured evidence on accuracy and adoption, the ethics duties that keep a human in the loop for in-house legal teams, and the parts of outside-counsel and transactional work that are moving first.

[cta-1]

Which legal tasks AI can do today, and which still need a lawyer

How to read the split

Read the table by column rather than by row. The middle column is production work on a document, and the right column is a decision about a transaction. First-pass contract review is the clearest case. A model working against a defined playbook can flag every clause in a hundred-page master services agreement that deviates from the standard position, and it can do that in minutes. What it cannot do is price the deviation. Whether a mutual indemnity capped at fees is acceptable depends on the counterparty, the revenue at stake, the insurance behind the cap and the precedent the team is willing to set on the next twenty deals.

Who maintains the playbook

The judgment in the right column lives in the playbook, and the playbook is a human artifact. The playbook decisions that stay with lawyers:

1. The standard position itself. Somebody decided that the opening position on limitation of liability is twelve months of fees rather than six or twenty-four.
2. The escalation threshold. Somebody set the point above which a deviation goes to the general counsel instead of being accepted by a contract manager, and somebody decided which deviations are not legal’s call at all, because the most negotiated terms in a commercial agreement are frequently owned by product, engineering, revenue or procurement.
3. The revision trigger. Somebody revises the playbook when the market moves, when a regulator changes the position, or when a renewal cycle exposes a term that no longer holds.

AI legal document review runs as the first pass on third-party paper and shortens the distance to those decisions, and the decisions themselves do not move. The practical consequence is that a team adopting AI on review work spends less time reading and more time maintaining the standard it reads against, which is a different allocation of a lawyer’s hours rather than fewer of them.

What the evidence says about AI replacing lawyers

Three measured sources carry most of what is currently known about AI replacing lawyers, and none of them describes a profession being automated away. Thomson Reuters surveyed in-house legal and tax professionals for the corporate edition of its Future of Professionals report and found 59% without access to professional-grade AI tools and 36% using unauthorized ones, which is an access problem rather than a replacement one. A Stanford RegLab and HAI study of leading AI legal research tools measured hallucination rates above 17% for two products and above 34% for a third, so the verification step a lawyer performs is load-bearing, and the court record of AI-fabricated citations is now large enough to count.

The access gap is the failure mode

The access figure describes the most common failure in legal AI adoption. A team without a sanctioned tool does not stop using AI. It uses a consumer chatbot on a personal account, which is where the confidentiality exposure sits. For a legal team choosing a tool, the useful question is not whether AI is capable of the task but whether the sanctioned path is faster than the unsanctioned one, because that gap determines which tool the work runs through.

Why verification is load-bearing

The Stanford result is more specific than a hallucination rate, and the specifics are what make the verification duty concrete. The failure mode the researchers isolate in their study of retrieval-based legal tools is the misgrounded response: an answer that cites a real case, correctly reported, that does not support the proposition it is cited for. The error is expensive because:

1. It survives the check most lawyers actually run, which is confirming that the citation exists and pulls up.
2. Catching it requires reading the authority against the proposition, which is the judgment the tool was supposed to save.
3. It fails silently, because a confident summary of a real case reads exactly like a correct one.

What the court record shows

The downstream record of skipping that step is now large enough to count. The AI Hallucination Cases database records 2,184 court decisions involving AI-fabricated material, with the United States accounting for 1,492 of them, and the quarterly counts are rising rather than falling. That pattern is consistent with adoption outpacing the verification habits that are supposed to accompany it, which is a training and supervision problem inside firms rather than evidence that the tools cannot be used safely.

Can AI replace outside counsel?

AI cannot replace outside counsel, though it is already moving a defined slice of outside-counsel work back in house. The work that moves is the volume layer: first-pass review of third-party paper, routine nondisclosure agreement and order-form turns, and the initial read of a diligence set. Those are the matters an in-house team used to send out because it lacked capacity rather than expertise.

Capacity against expertise

Whether a matter moves in house depends on whether it was sent out for capacity or for expertise. A team sends a matter out when it does not have the hours, the specialism, or the independence an outside opinion carries. AI changes the first reason and leaves the other two intact.

Work that stays with the firm:

1. The bet-the-company matter, where the cost of being wrong exceeds the entire legal budget
2. A novel structure with no precedent in the team’s own files
3. A jurisdiction the team does not practice in, where local counsel carries the risk
4. The second opinion a board or an audit committee wants on the record from someone who is not an employee
5. Anything likely to be litigated, where the work product is being built for a court rather than for a counterparty

Work that moves in house:

1. Volume turns on standard paper, where the team’s own playbook is the standard
2. Clause extraction and obligation tracking across an existing contract base
3. The first read of a diligence set, before the issues that matter are known
4. Low-value renewals and amendments that were outsourced on hours alone

What changes at the panel review

The mechanism is a change in matter mix rather than a change in panel. A team that gains first-pass review capacity does not usually fire a firm. It stops sending out the work it was reluctant to send out at all, which raises the average complexity of what the panel sees and changes what the relationship is for. The same split between AI and lawyers runs inside the team on the matters it keeps, and the consequence shows up at the annual review, where the question becomes which categories of work the firm should still see rather than what the hourly rate on all of it should be.

Will AI replace transactional lawyers?

AI will not replace transactional lawyers, and transactional practice is nonetheless where it has moved furthest into the work. Transactional work is document-dense, pattern-heavy and governed by precedent language, which is exactly the shape of task a model handles well. The judgment that remains is allocational: which risk the client accepts, which position to open on, and when a deviation is worth the fight.

The leverage question

The exposure is to leverage rather than to the role. Transactional practice has been staffed on a pyramid, and the base of that pyramid is document hours. The exposed tasks are the document-dense ones:

1. Reading the data room on a diligence exercise
2. Turning the schedules and exhibits across a signing set
3. Checking consistency of defined terms and cross-references between documents
4. Marking up third-party paper against a precedent

Corporate practice sits in the same position for the same reason, because a corporate lawyer’s day is made of the same document-dense tasks as a transactional one.

Reading a commercial position

The skill that appreciates is reading a commercial position: knowing which of the counterparty’s asks are real and which are opening positions, what the client’s actual tolerance is as opposed to its stated one, and where the deal breaks if a term is conceded in the wrong order. That skill used to be acquired as a by-product of the document hours, which is the training problem AI creates for a practice group. Teams that address it do so deliberately:

1. Juniors join the negotiation earlier, as a second chair rather than as a note-taker
2. Review shifts to the judgment in the output rather than the diligence on the input
3. The precedent choice is made explicit and discussed, instead of being inherited with the file

The same pattern holds one level down, where the question of whether AI can take over paralegal work turns on who is accountable for the output rather than on who produced the first version of it.

Why AI cannot replace legal judgment

AI cannot replace legal judgment because judgment is the act of choosing between defensible options under uncertainty, and a model optimizes for the likeliest next token rather than the best outcome for one client. A language model can tell a lawyer that an indemnity is unusually broad. It cannot decide whether this client should accept it to close this week.

Why the mechanism matters

The gap persists because of that mechanism rather than a shortcoming in a particular product. A model predicts the next piece of text from the patterns in its inputs. It does not hold a client’s risk appetite, it does not know what was promised in a board meeting, and it has no stake in the relationship the contract is supposed to support.

Consequences in practice

1. The unusual term is not always the wrong term. A model trained on market language flags the deviation from the market. A lawyer knows when the deviation is the point, because the client’s position in this relationship is not the market’s.
2. Bias travels with the inputs. A system trained on a contract base reproduces the positions in that base, including the ones a team has outgrown. The check is a human one: somebody has to ask whether the standard being enforced is still the standard the business wants.
3. Confidentiality is a workflow property, not a model property. What protects privilege is where the data goes, who can read it and how long it is kept, and those are procurement and configuration decisions a lawyer is responsible for.

Who carries the output

Litigation raises the same question with higher stakes. For a commercial team the practical test is whether a named person can be held to the output. If a term goes wrong eighteen months later, the client asks who approved it, and the answer has to be a lawyer who understood what was being approved. No part of that sentence is satisfied by a model, which is why the useful question about a legal AI tool is not how accurate it is in the aggregate but how cheaply it lets the responsible lawyer check the particular thing in front of them.

[cta-2]

The ethics duties that keep a lawyer in the loop

Lawyer ethics duties when using AI are already written down, and they put a named human on every output. ABA Formal Opinion 512, issued in July 2024, applies the duties of competence, confidentiality, communication and reasonable fees to generative AI, and it requires an independent review of AI output rather than reliance on it.

What the state guidance adds

State guidance says the same thing in operational terms. The State Bar of California’s practical guidance on generative AI in the practice of law, issued by its Standing Committee on Professional Responsibility and Conduct and revised at the request of the California Supreme Court to address agentic systems, is explicit that a lawyer’s professional judgment cannot be delegated to AI and remains the lawyer’s responsibility at all times. It draws three distinctions that matter for a commercial team deciding how to deploy a tool:

1. Competence is two duties, not one. A lawyer has to understand the system’s capabilities, data sources and material risks before using it, and then has to review, verify and correct what it produces.
2. Confidential inputs require informed consent. A lawyer must not input confidential client information into a generative AI solution that presents material risks to confidentiality or security without informed client consent as to those risks, and reasonable efforts to assess that risk mean reading the terms of use, the privacy policy and the data-retention position rather than relying on marketing assurances.
3. Autonomy raises the supervision bar. The greater the autonomy a system is given, the more oversight the lawyer owes it, and an agentic system may not be deployed in a way that lets it make substantive legal determinations, give advice or file documents without meaningful supervision and review.

An ethics opinion sets duties whose application depends on the client, the matter and the tool. The supervision duty is the part commercial teams most often miss, because it applies to subordinates using AI as much as to the lawyer using it personally, which makes a tool’s configuration a supervisory question rather than an IT one.

The diligence that follows

The work this creates is familiar diligence in an unfamiliar place:

1. Data minimization on what goes into the tool in the first place
2. Anonymization or redaction of party names where the matter allows it
3. Encrypted transfer and a written position on how long inputs are retained
4. Vendor due diligence proportionate to the sensitivity of the files involved
5. A record of who approved the deployment and who supervises its use

When will AI replace lawyers?

No credible timeline has AI replacing lawyers, and the timelines that do exist describe hours moved rather than roles removed. Thomson Reuters, writing up its own Future of Professionals research on its legal industry blog, puts the saving at nearly 240 hours a year per professional, and the same analysis has 41% of law firms and 47% of in-house legal departments already using generative AI, up from 28% and 23% a year earlier. Adoption is running well ahead of replacement.

What the hours data says

The same research projects the saving in hours rather than in headcount. Thomson Reuters projected a saving of twelve hours per week within five years, starting at about four hours a week in the near term. A team can spend recovered hours in these ways:

1. Bill fewer hours on the same matters, which converts the saving into a client-facing price change
2. Take on more matters with the same headcount, which converts it into capacity
3. Move the team onto work it was previously too busy to do, such as playbook maintenance or contract-base cleanup

Which of the three a team picks is a business model decision, and it is the decision that determines whether AI looks like a threat or like capacity.

Which tasks have a visible horizon

Asking when AI replaces lawyers is usually asking about a horizon for a specific task, and that question has an answer. The tasks with a visible horizon are the ones already moving:

1. First-pass review of third-party paper against a known standard
2. Extraction of obligations, dates and defined terms
3. First drafts generated from a precedent
4. Diligence triage, before the issues are known

The tasks without one are negotiation, advice a client relies on, and anything where a named person has to carry the risk. No current research describes a mechanism by which those move, so a team planning against this does better to ask which of its own tasks sit on the first list and resource accordingly than to plan for a date.

Where AI fits in a commercial contract workflow

Lawyers use AI at four points in a commercial contract workflow: intake, first-pass review, drafting, and benchmarking a term against the market. Each point has a named human step after it, which is where the workflow differs from automation. Spellbook’s Review feature reads third-party paper against a firm’s playbook and proposes redlines as native Word tracked changes, and a lawyer accepts or rejects each one.

Intake and first-pass review

Intake is the point where a request becomes a matter, and the AI contribution is triage: classifying the paper, identifying the counterparty’s standard form and routing it to the right standard. First-pass review then runs the document against the playbook, which is the task the table above puts squarely in the AI column. The human step is the one that cannot be skipped: a lawyer decides which flagged deviations are acceptable on this deal, and the escalation threshold in the playbook sets the point at which that decision belongs to someone more senior.

Drafting from precedent

Spellbook’s Draft feature works in the other direction, generating clause and agreement language from the team’s own precedent inside Microsoft Word. That addresses the first-draft problem the task table names: producing the draft is production work, and choosing the precedent and the opening position is not. The human step is the selection, which is why a drafting tool is only as good as the precedent set it has been pointed at.

Benchmarking a term against the market

Spellbook’s Compare to Market feature answers the question a playbook cannot, which is whether the standard position is still a market position, by comparing a term against a set of 20M+ real contracts. This is the capability that turns playbook maintenance from an annual guess into a check against evidence. The human step is the interpretation: a market position is a distribution rather than a rule, and whether this counterparty should get the market term is still a judgment about the relationship.

Multi-document review

Spellbook’s Associate feature handles the multi-document end of the same workflow, where the question spans a signing set or a diligence folder rather than a single agreement. The practitioner value is consistency across documents, which is the error class a single-document review cannot catch: a defined term amended in one place and not another, or a cap that does not match the schedule it references. The human step is deciding which of the inconsistencies found actually change the deal.

Where the platform stops

Spellbook is a commercial-contract specialist, which means it is not a legal-research platform and not a litigation tool, and that distinction matters doubly here, because the hallucination rates in the evidence above were measured on legal research products rather than on contract review. A team evaluating on the configuration axis should also weigh the security position, which is where the diligence duties above land in a procurement questionnaire:

1. SOC 2 Type II
2. Zero data retention
3. Single sign-on (SSO) and multi-factor authentication (MFA)

The wider market of legal AI tools divides along the same line, between tools built for research and tools built for contract work. For the evidence behind how commercial terms are actually moving, read the State of Contracts report.

[cta-3]

Frequently asked questions

Can AI replace human lawyers in contract review?

No. AI can replace the first pass of contract review, which is the mechanical comparison of a document against a standard, and the review is not finished when that pass ends. A lawyer decides which flagged deviations are acceptable on the specific deal, and that decision carries the professional liability. The practical division is that the tool produces a list of candidate issues and the lawyer produces the position.

Can AI adapt legal documents to different jurisdictions?

Only as a first draft, and only where the tool has been given jurisdiction-specific source material to work from. A model can apply the drafting conventions it has seen for a given jurisdiction and flag terms that commonly fail there, such as a non-compete in a jurisdiction that will not enforce one. It cannot confirm that the adapted document is enforceable under local law. That confirmation needs a lawyer qualified in the jurisdiction, which is why cross-border work tends to keep local counsel in the loop even where the drafting is automated.

Will lawyers be automated?

Parts of legal work are being automated, and the role is not. The pattern across the available evidence is that document-dense production tasks move to AI while the decisions attached to them stay with lawyers, because professional liability does not transfer to software. The measurable effect so far is on hours rather than on positions: the research points to a saving measured in hundreds of hours a year per professional, and to adoption in the 40s as a percentage of firms and legal departments, with no corresponding account of roles removed.

What legal AI tools work without requiring lawyers to change how they work?

The tools that fit an existing workflow are the ones that live inside the application the work already happens in, which for commercial contract work means a word processor. A tool that runs as an add-in to Word or Google Docs puts review and drafting where the document already is, and the lawyer keeps tracked changes, comments and the existing file as the working artifact. The alternative pattern is a separate workspace that documents are uploaded to, which adds a step at both ends and tends to be adopted for repository and reporting functions rather than for drafting. The axis worth evaluating on is integration rather than feature count, because a tool that requires a new workspace competes with the habit it is trying to replace.