See how Spellbook helps legal teams draft contracts faster than ever before.
Book Demo
Generative AI for legal practice now does real work in five places: contract drafting, contract review, due diligence, legal research and client correspondence. In each of them it compresses the mechanical layer of legal work while judgment, supervision and verification stay with the lawyer. Courts have now sanctioned filings built on generative AI output that nobody checked, which is why every workable use of generative AI for legal work pairs a drafting gain with a named verification step.
Teams that get value from generative AI have settled three things before the first matter: which workflow the technology owns, who verifies its output, and what happens when it is wrong. Teams that have left those undecided carry the real exposure, because individual lawyers then paste client language into consumer chatbots with no policy behind it. Commercial teams at law firms close that gap by scoping one workflow at a time rather than by buying a platform and hoping adoption follows.
[cta-1]
Generative AI in legal practice is software that produces new text, such as a clause, a summary or a memo, from statistical patterns learned across large bodies of documents, rather than sorting documents that already exist into categories. The distinction matters because the two technologies fail differently: a classifier mislabels a clause it has seen, and a generative model writes a confident sentence about a case that does not exist.
The models behind these tools have moved through several generations since 2022, and the current generation is materially better at long documents than the one most firms first tested. Where a vendor does not publish which model version sits behind a feature, evaluate the capability claim rather than the model name.
Traditional legal AI classifies and extracts. It reads a document set, assigns each clause to a category, and pulls named fields into a table. The output is a label or a value, and it is traceable to a span of text in the source document.
Generative AI writes. Given a prompt and a precedent, it produces language that did not exist before. An extraction error is visible against the document, and a generation error reads like competent drafting until someone checks the authority behind it.
Generative AI in the legal industry has changed how first drafts and first reviews get produced, it has not changed who is accountable for them, and it has added a verification step that did not exist in 2022. Adoption rose fastest in document-heavy transactional work, where the input is a contract and the output is a redline a lawyer reads line by line.
The starting point moved. A first-pass nondisclosure agreement (NDA) or a summary of a full data room now arrives before a lawyer has read the first page, and the lawyer’s time moves from production to checking and positioning.
Accountability for the output stays exactly where it was. The signed agreement still reflects a decision a named lawyer made.
The governance layer is where the industry is furthest behind its own usage. Written AI use policies, training, matter-level disclosure practices and any measurement of output quality all lag the number of lawyers already using these tools. The ABA Task Force on Law and Artificial Intelligence has treated that lag as a practice-management problem rather than a technology problem. As a result, the risk today usually sits in undocumented individual use, where lawyers already apply generative AI to their own day-to-day work without any written policy behind it, rather than in sanctioned deployments that a firm actually tracks.
| Use case | What generative AI produces | What the lawyer still does |
|---|---|---|
| Contract drafting | A first-pass clause or agreement from a prompt and precedent | Sets the position, checks the fallback, approves the language |
| Contract review | A list of flagged risks and proposed redlines | Decides which risks matter on this deal |
| Due diligence | Summaries and extracted terms across a document set | Confirms the extraction against the source document |
| Legal research | Candidate authorities and a summary of them | Reads and verifies every citation |
| Correspondence | A plain-language explanation of a term or issue | Confirms accuracy and sends it |
Drafting has structured inputs: a playbook position and a precedent, with an output that is checkable against a precedent the team already trusts. A model given those inputs produces a serviceable first draft of a standard commercial agreement, and the work that remains is the work that was always the lawyer’s: choosing the position and deciding how far the fallback goes.
A fluent draft can still be internally inconsistent, usually a defined term used in two senses or a notice period that contradicts the term in another clause, and catching that depends on how well generative AI for legal contracts is applied to the clause library and playbook behind the draft.
Review is where the volume problem bites hardest: a team facing more third-party paper than review hours triages instead of reading. Generative AI reads the counterparty draft against a position and proposes redlines in the document.
Whose paper it is determines how useful this gets. On your own template, deviations are the whole signal. On third-party paper, the tool flags every departure it finds and the lawyer decides which of them matter on this deal, which is judgment that does not transfer. Teams negotiating from the counterparty’s draft run into this split first.
Diligence changes coverage: a model can read every agreement in a data room rather than the sample a timeline allows, extracting the change-of-control, assignment and exclusivity provisions.
Confirm the extracted term against the source document, because an extraction attached to the wrong clause reads exactly like a correct one in a summary table. Allocating that verification across a deal team is the part of multi-document review that needs designing before the first data room opens.
Research use carries a verification burden on every output, because published evidence on these tools is not reassuring. Magesh and colleagues ran a preregistered evaluation of retrieval-augmented legal research tools and found they hallucinate between 17% and 33% of the time. At that published error rate, unverified research output cannot go into a filing.
Generative AI is useful for finding candidate authorities and unsafe for stating what they hold. The lawyer reads each case. Treating the output as a research lead rather than a research answer is the distinction that keeps AI-assisted research inside the duty of competence.
A plain-language explanation of an indemnity or a limitation of liability, drafted for a commercial counterpart who is not a lawyer, is work a model does well and work a partner is expensive to do.
This output leaves the firm. An explanation that softens a cap or omits a carve-out becomes a representation the client relied on, so the review standard for outbound correspondence is the same standard as for the clause it describes, not a lighter one.
Every one of the five pairs a production gain with a verification step, and the gain tracks how structured the input is.
The limitations of generative AI in legal work are structural rather than temporary: the technology predicts likely text, so it cannot know whether a case is still good law, cannot weigh a client’s risk appetite, and cannot take responsibility for advice. A model that has never been told which fallback the client accepted will propose the one that appears most often in its training data.
The human step is not a formality. It is the step where someone who knows the matter decides whether the output is right for this client, on this deal, under current law, and that decision is recorded against a name. The question about replacement therefore turns on accountability rather than on capability.
[cta-2]
The generative AI legal ethics risks lawyers actually face are documented rather than speculative: a public database of judicial decisions addressing AI-hallucinated material held 2,266 cases on November 5, 2026, 1,535 of them in the United States. ABA Formal Opinion 512, issued on July 29, 2024, maps these risks onto six existing duties: competence, confidentiality, communication, candor toward the tribunal, supervision, and fees.
Fabricated authority is the risk with the longest public record. The database of judicial decisions addressing hallucinated material tracks matters where a court found a party relied on content an AI tool invented, and the count has grown steadily rather than tailed off as the tools improved.
The canonical matter is Mata v. Avianca, where counsel filed a brief citing cases that did not exist and then submitted fabricated opinions when the court asked for them. The docket in that case is worth reading for what the sanctions turned on, which was not the use of the tool but the failure to check its output and the response once challenged.
The verification step is concrete rather than attitudinal, and a team that writes it down is a team that can show it happened. That record is what turns a hallucination caught internally into a working process rather than a near miss, and it is the record a court asks for when a filing is challenged.
ABA Model Rule 1.6 does not change because the recipient of client information is a model. Formal Opinion 512 analyzes the point directly: a lawyer who enters client information into a tool that retains it, trains on it, or exposes it to anyone outside the relationship may need informed client consent before doing so.
The practical distinction is between a consumer chatbot and a tool bound by its contract. A general-purpose consumer product may retain conversations and use them in training by default, while a legal tool can commit contractually to zero data retention, which means the inputs are not stored after the request completes. That difference is a diligence question for procurement, and the privilege analysis behind it is worth settling before the first matter rather than after.
Disclosure runs on two tracks. A growing number of judges have issued standing orders requiring a certification about generative AI use in filings, and those orders differ by courtroom, so the obligation is local and has to be checked per matter rather than per firm.
Client disclosure is the second track. Formal Opinion 512 treats the duty of communication as engaged where the use of generative AI is material to the representation, and the duty around fees as engaged where a tool changes how long the work takes. Neither obligation resolves to a single sentence in an engagement letter, and some state-level guidance remains in draft, which is why the disclosure position is something to document and revisit rather than settle once.
Copyright protection in the United States requires human authorship. The Copyright Office’s Part 2 report on copyrightability concludes that material generated entirely by a machine is not protectable, while a work containing AI-generated material can be protected to the extent of the human contribution to it.
For commercial legal work the consequence is narrow. Contract language is rarely where a client’s copyright value sits, so the exposure is usually in the client’s own AI-assisted product or marketing output.
To start using generative AI in legal practice, pick one bounded workflow, write the use policy before the first matter, choose a tool built for that workflow, name who verifies the output, and measure the change. The order matters: teams that choose the tool first end up writing a policy around whatever the tool happens to do.

The procurement question is fit rather than capability, and the shortlist is worth weighing against the way tools for commercial lawyers actually divide by workflow.
Spellbook is an AI suite for transactional legal work, which means its use of generative AI is deliberately narrow: commercial contract drafting and review, done in the document rather than in a separate chat window. It works in Word and Google Docs, and the Draft and Review features are grounded in the team’s own playbooks and precedents plus 20M+ real market contracts. Spellbook drafts and auto-completes using GPT-5.

That scope is a deliberate limit rather than a gap. Spellbook is a commercial-contract specialist, not a legal-research or litigation platform, which is why the research use case above is not a Spellbook use case. A team that needs authority checking and a team that needs faster redlines are buying two different things.
Within contracts, each feature covers a defined part of the workflow.
Verification applies to this tooling as much as to any other. A suggested redline is a proposal, the lawyer decides which departures matter on the deal, and the tool can help ensure the mechanical layer is consistent without taking a position on risk appetite. Pricing is custom per seat and scales by team size, and there is a free seven-day trial.
Try Spellbook free.
[cta-3]
Yes, within existing duties. ABA Formal Opinion 512 treats generative AI use as governed by the rules already in place on competence, confidentiality, communication, candor toward the tribunal, supervision, and fees, rather than by a new prohibition. The ethical question is therefore not whether to use it but whether the specific use satisfies those duties, which turns on what the tool does with client information and who verifies its output.
Sometimes, and it depends on materiality. The duty of communication is engaged where the use of generative AI is material to the representation, and the duty around fees is engaged where a tool changes how long the work takes. Courts add a separate obligation: a number of judges have issued standing orders requiring certification about AI use in filings, and those orders differ by courtroom, so the filing-side position has to be checked per matter.
Not the machine-generated part. Copyright protection in the United States requires human authorship, so material produced entirely by a model is not protectable, while a work containing AI-generated material can be protected to the extent of the human contribution. The practical test is whether the author can describe and show what the human contributed.
Generative AI is the underlying capability, and a legal AI tool is a product built around it for a defined workflow. The older generation of legal AI classified and extracted, producing labels and field values traceable to the source text. Generative tools write new language, which is more useful and harder to check, because a generation error reads like competent drafting until someone verifies the authority behind it.
No, not on current capability. The capability gap is in the parts of legal work that do not delegate. A model cannot confirm a decision is still good law without a live authority check, cannot weigh a client’s risk appetite, and cannot hold professional responsibility for advice. What it changes is the composition of the work, moving time from producing a first draft toward checking, positioning and deciding, which is a change in the task mix rather than in who is accountable.



Submission Received
Thank you for your interest!
Submission Received
Thank you for your interest!
We're connecting you with the best rep