AI Use Cases in Law: 8 Workflows and the Rules That Apply

The AI use cases in law that hold up in daily practice are contract drafting, contract review, benchmarking terms against the market, legal research, document review and e-discovery, summarizing long documents, obligation and deadline tracking, and risk and compliance review. Each one compresses a task that law firms and in-house legal teams used to do by hand, and none of them takes the lawyer off the file. ABA Formal Opinion 512 leaves duties such as competence, confidentiality, client communication and reasonable fees with the lawyer who uses the tool, and courts have sanctioned filings built on cases a model invented. The practical question is therefore which legal workflows AI is accurate enough for today, what a tool has to prove before it touches privileged material, and where a lawyer still reads every line.

Key Takeaways

1. The AI use cases in law with the longest track record are contract drafting, contract review, market benchmarking, legal research, document review and e-discovery, summarization, obligation tracking and compliance review, in both law firms and in-house legal teams.
2. Legal research is the AI use case in law where the gap between speed and reliability is widest: a general-purpose model fabricates citations, and courts have struck filings built on them.
3. ABA Formal Opinion 512 creates no new rule for AI. It applies existing Model Rules to it, among them 1.1, 1.4, 1.5 and 1.6, so competence, client communication, fee reasonableness and confidentiality all remain duties of the lawyer.
4. Contract work is the AI use case in law where a tool can be grounded in a team’s own precedent, which is why accuracy there is higher than in open-web research.

[cta-1]

AI Use Cases in the Legal Industry at a Glance

Applications of AI in the legal sector run wider than these eight, and these are the ones with enough track record to plan a practice around.

AI Contract Drafting

AI contract drafting is the use case with the most settled workflow, because the first draft of a commercial agreement is largely assembly: known clauses, known positions, known fallbacks. A drafting tool grounded in a firm’s own precedent can produce a complete first pass with the standard provisions in place, leaving the lawyer to argue the two or three terms that actually carry the deal.

What It Does Well

1. Assembling a known structure. The standard provisions, defined terms and signature mechanics land in the first pass rather than being pulled from the last deal that looked similar.
2. Clause alternatives on demand. Fallback positions for a liability cap or an indemnity can be produced at the three or four levels a negotiator actually uses, which removes the hunt through closed files for the version that got signed.
3. Consistency across a deal family. The same defined terms and cross-references carry across a master services agreement, its order forms and its statements of work, which is where manual drafting usually drifts.

Where It Needs a Lawyer

1. Risk allocation. Which party carries which liability is a commercial decision, and a generated cap reflects the precedent set rather than this deal’s leverage.
2. Scope and deliverables. The description of what is being bought comes out of the negotiation, and a model has no access to what was agreed on the call.
3. Non-standard terms. The provisions a counterparty will fight over are, by definition, the ones with no precedent to assemble from.

Fallback positions are easier to standardize when the team’s preferred language is written down, which is what a clause index does: a named set of standard provisions a drafting tool can be pointed at instead of producing one unprompted.

AI Contract Review and Risk Flagging

AI contract review reads an inbound agreement against a defined standard and flags every clause that departs from it. The useful output is not a risk score but a list of specific deviations: an indemnity with no cap, a notice period shorter than the playbook allows, an assignment clause that survives a change of control.

The volume of defects in circulating paper is the reason the workflow exists. An analysis of 3,019 material-contract exhibits filed on the Securities and Exchange Commission’s (SEC) EDGAR system by more than 500 public companies, run by Spellbook Labs, found that 60% of SEC-filed contracts contain drafting errors, so an inbound review is screening a population in which error is normal rather than exceptional.

What It Does Well

1. Playbook deviation. A defined standard turns review into a comparison, and a comparison is the kind of task a model is reliable at: this cap is lower than the floor, this notice period is shorter than the minimum.
2. Missing provisions. The absence of a clause is harder to spot by eye than a bad one, and an extraction pass against a required list catches the omission.
3. Volume triage. A stack of renewals or supplier paper can be sorted by how far each document sits from the standard, so the lawyer reads the worst first.

Where It Needs a Lawyer

1. Materiality. A flag is a finding, not a decision, and which of twenty deviations is worth raising depends on the counterparty, the spend and the relationship.
2. Negotiating posture. What gets conceded is a judgment about leverage that no flag carries.
3. Interpretation. A clause that reads cleanly in isolation can conflict with a defined term thirty pages away, and resolving that conflict is legal analysis.

Where the automated pass sits inside the wider contract review process decides how much of it lands, because a flag list produced before the team has agreed its standard moves the disagreement rather than resolving it.

AI Contract Benchmarking Against Market Terms

AI contract benchmarking reports whether the term on the table is normal, which a playbook cannot do: a playbook states the team’s own standard, not the market’s. A tool with a large set of executed agreements behind it can report how a liability cap, a termination-for-convenience right or a payment term compares with similar deals, which turns a negotiating instinct into a position a lawyer can defend to a client.

What It Does Well

1. Positional evidence. A statement that a mutual cap at one year of fees sits inside the common range for this deal size is checkable, and a counterparty argues with it differently than with a preference.
2. Playbook calibration. Benchmarking surfaces a playbook standard that no comparable agreement meets, which is a standard worth revisiting.
3. Onboarding newer lawyers. A market range gives an associate a defensible starting position before the pattern recognition arrives.

Where It Needs a Lawyer

1. Comparability. A benchmark is only as good as the set it is drawn from, so whether those agreements resemble this one is the first question.
2. Deviation on purpose. Plenty of terms sit outside the market because the client wants them there.
3. The client conversation. Market position informs advice; it does not replace the recommendation.

AI Legal Research

AI legal research is the AI use case in law where the gap between speed and reliability is widest in practice. Research platforms built on retrieval over a licensed case database surface candidate authorities and summarize an argument quickly, but a general-purpose chatbot asked the same question can invent citations that look correct, and the lawyer who files them owns the consequence.

That allocation of risk is written into the rules of practice. A filing certifies that the legal contentions in it are warranted, a duty Federal Rule of Civil Procedure 11 places on the attorney who signs rather than on any tool, which is why a fabricated citation becomes the lawyer’s problem the moment it reaches the docket.

The research category is currently held by platforms that sit on licensed content rather than on the open web: Westlaw, Lexis+ and CoCounsel. Retrieval over a licensed database narrows the failure mode without closing it, because a real case can still be summarized wrongly or cited for a proposition it does not support. How AI legal research retrieves and ranks authority is what decides where the remaining error sits.

What It Does Well

1. Entry points. A plain-language question returns candidate authorities faster than a Boolean string, which shortens the gap between a new issue and a reading list.
2. Argument summaries. Condensing a long opinion into the holding and the reasoning is a checkable task when the opinion is open alongside it.
3. Secondary-source orientation. Treatise and practice-note retrieval gets a lawyer oriented in an unfamiliar area before the primary research starts.

Where It Needs a Lawyer

1. Every citation. The case must exist, say what it is cited for, and still be good law.
2. Weight of authority. A model has no sense of which circuit, which year and which posture makes a case useful here.
3. Jurisdictional fit. A proposition that is correct somewhere is not an answer for the forum the matter sits in.

AI Document Review and E-Discovery

Document review and e-discovery were automated long before generative AI arrived. Technology-assisted review (TAR) ranks large document populations by likely relevance and privilege so that a review team reads the top of the pile first, and the newer models add the ability to summarize what a flagged document says rather than only that it matched.

What makes a prioritized review defensible is the standard it answers to. The scope of discovery is what is relevant to a claim or defense and proportional to the needs of the case, so a ranked review is measured against proportionality rather than against a promise that every document was read with equal attention. Generative models have been layered on top of that older ranking machinery rather than replacing it, which is the practical shape of e-discovery AI today.

What It Does Well

1. Prioritization at volume. A population of several hundred thousand documents can be ordered so that the responsive material surfaces in the first passes rather than the last.
2. Near-duplicate and thread handling. Email threading and near-duplicate detection remove most of the repeat reading that manual review pays for twice.
3. Issue coding at scale. A model trained on a reviewed sample can extend an issue tag across the population, which a review manager then samples for accuracy.

Where It Needs a Lawyer

1. The privilege call. A privilege prediction is a prediction, and the production decision stays with the lawyer who signs off on it.
2. Sampling and validation. Someone has to set the recall target, draw the control set and defend the methodology if the protocol is challenged.
3. Narrative judgment. The documents that matter to the case are often not the ones that rank highest on relevance.

The Rules the Protocol Answers To

Production obligations are what give the ranking its stakes: the form in which electronically stored information must be produced is set by Rule 34, and a review protocol that cannot explain how it met the request is the one that gets challenged. A production in native format with the metadata intact is a different deliverable from a set of images, and a ranking method has to survive that difference without a second review pass.

Privilege is the part of the workflow that carries its own procedural safety net. Federal Rule of Evidence 502 governs the consequences when privileged material is produced inadvertently, and clawback agreements under it are standard practice for exactly the reason a predictive privilege ranking is not a privilege determination.

Where automated review ends and attorney review begins is the line that AI legal document review has to draw explicitly, because a protocol that leaves it implicit cannot be defended when the methodology is challenged.

The confidentiality exposure sits on top of the production question. A review platform that ranks across a privileged document population has to keep the privilege calls separate from the relevance calls, or a batch cleared for relevance goes out with privileged material still inside it. A general-purpose assistant run over a privileged document population is a direct route to an inadvertent waiver of privilege, which is why the privilege log and the clawback agreement are settled before the ranking starts rather than after a production has gone out.

AI Summarization of Documents and Transcripts

AI legal document summarization carries the least exposure of these eight AI use cases in law, because the source document is in front of the lawyer and every claim in the summary is checkable against it in seconds. Summarization is also the AI use case in law that a lawyer can audit fastest, because checking a summary means rereading a passage rather than verifying an outside authority. The practical uses are deposition transcripts, long-form filings, diligence data rooms and the fiftieth version of a master services agreement.

What It Does Well

1. Transcript compression. A day of testimony reduces to the admissions, the exhibits referenced and the lines worth pulling, with citations back to the page.
2. Data room triage. Hundreds of agreements in a diligence data room can be summarized on the handful of terms the deal turns on, such as assignment, change of control and exclusivity.
3. Version deltas. A summary of what changed between two drafts is faster to read than a redline when the question is whether anything material moved.

Where It Needs a Lawyer

1. The passages the summary turns on. A summary is a reading aid, and the sentence that decides the issue gets read in the original.
2. Omission. A summary reports what it found, and what it left out is invisible unless someone checks.
3. Characterization. Whether a provision is “standard” is a legal conclusion that does not belong in a summary.

Diligence is where the volume makes the case, because AI due diligence is this same summarization step run across a whole data room rather than one document, and the review memo is assembled out of those summaries.

[cta-2]

AI Obligation and Deadline Tracking

AI contract obligation tracking turns an executed contract back into a list of things somebody has to do. AI extracts renewal dates, notice windows, reporting duties and price-escalation triggers from a signed agreement and writes them into a calendar or a register, which is the use case that prevents an avoidable loss rather than saving time.

What It Does Well

1. Date extraction. Effective dates, renewal windows, notice periods and termination rights come out of the document and into a register without a manual read of every executed file.
2. Recurring duties. Reporting obligations, insurance certificates and audit rights are easy to miss, because nothing in the business process prompts them.
3. Portfolio visibility. A register built this way answers questions such as which agreements auto-renew inside ninety days, which no stack of signed documents can answer on its own.

Where It Needs a Lawyer

1. Escalation. An approaching renewal is information, and whether to renegotiate, renew or exit is a business decision taken with counsel.
2. Ambiguous triggers. A notice period that runs from an event rather than a date needs a person to decide when the event happened.
3. Register accuracy. An extracted date that nobody validated is a false sense of coverage, so the first pass is spot-checked against the paper.

AI Risk Assessment and Compliance Review

AI legal risk assessment and compliance review asks whether an agreement conflicts with a named framework: the General Data Protection Regulation (GDPR), the California Consumer Privacy Act (CCPA), the Health Insurance Portability and Accountability Act (HIPAA), a sanctions regime, an internal policy. AI can compare clause language against the requirement and report the gaps, and it is also the layer that catches altered or missing provisions in a document that came back from the other side looking unchanged.

The framework has to be named for the comparison to mean anything. A data processing addendum reviewed against the GDPR is a different exercise from the same document reviewed against the CCPA, and a tool asked for “privacy compliance” without a named regime returns the intersection of several, which is where a false pass comes from.

What It Does Well

1. Clause-to-requirement mapping. A required transfer mechanism, retention limit or subprocessor consent either appears in the document or does not, and that is a comparison AI can help identify.
2. Altered paper. Silent edits in a returned document, including a deleted sentence inside an otherwise untouched clause, surface in a comparison against the sent version.
3. Policy conformance at volume. An internal standard applied across a supplier base is a repeatable check rather than a sampling exercise.

Where It Needs a Lawyer

1. The legal conclusion. A gap report is not a compliance opinion, and whether the gap creates exposure is an attorney judgment.
2. Remediation sequencing. Which gaps get fixed, in which agreements, in what order, is a risk decision.
3. Regulatory change. The framework moves, and a check built against last year’s requirement reports a clean result against the wrong standard.

Compliance language also has to be precise about status. HIPAA has no certification scheme, so compliance is the only accurate claim available about a vendor, and a report describing any stronger status describes a category that does not exist.

Risk work of this kind is often where a team starts, which is why contract risk assessment tends to arrive before the drafting use cases do. It also has a clear internal owner, which makes it straightforward to fund.

How Many Lawyers Actually Use AI

The number of lawyers using AI is now high enough that unauthorized use is its own category. In the legal cut of the Thomson Reuters 2026 Future of Professionals Report, 34% of law firm professionals say they are currently using AI tools their firm has not authorized, which puts adoption ahead of firm policy.

Client expectations are moving at the same time. Thomson Reuters found in the main cut of its 2026 Future of Professionals Report that 78% of corporate clients called AI-enabled quality improvements very important or essential, while only 6% said most or all of their providers deliver them.

AI Adoption Figures in the Legal Profession

The same legal cut reports that 71% of in-house legal professionals expect outside firms to change their commercial models while 28% of firms say they are doing it, and that 24% of law firm professionals would decline an offer without professional-grade AI tools. That gap is commercial pressure that in-house legal teams, who buy the work those figures describe, apply directly.

The ABA Ethics Rules That Apply When Lawyers Use AI

The ABA ethics rules for lawyers using AI are the rules that were already there. ABA Formal Opinion 512 is the controlling guidance, and it creates no new duty. It reads existing Model Rules onto the tool, among them Rule 1.1 competence, including a reasonable understanding of how the tool behaves, Rule 1.6 confidentiality over everything relating to the representation, Rule 1.4 communication about material use, and Rule 1.5 fee reasonableness when the work takes minutes instead of hours.

The ABA Standing Committee on Ethics and Professional Responsibility issued it on July 29, 2024 as its first formal opinion on generative AI.

The Rules Opinion 512 Applies

1. Rule 1.1 (Competence): a reasonable understanding of the tool’s capabilities and limitations, which the opinion treats as part of existing technological competence rather than as a new obligation.
2. Rule 1.6 (Confidentiality): informed client consent before entering information relating to a representation into a tool that is not self-contained.
3. Rule 1.4 (Communication): telling the client where the use of AI is material to the representation, including where it bears on the scope or the cost of the work.
4. Rule 1.5 (Fees): a fee that stays reasonable when a task that took hours now takes minutes. The opinion’s own example is a lawyer who spends 15 minutes inputting information where the drafting previously took considerably longer.
5. Candor and supervision: the opinion also reaches the duty of candor to the tribunal and the supervisory duties that partners and managers owe over how the firm uses the tool, so an AI policy is a supervision question as well as a technology one.

Why It Is Guidance, Not a Bright-Line Rule

Formal Opinion 512 applies existing rules to a new tool rather than setting a bright-line test, and its published text frames the analysis as specific to the tool in front of the lawyer and the matter it is used on. Two tools with different data terms can land in different places under Rule 1.6 on identical facts.

Where AI in Law Still Falls Short

The limitations of AI in law cluster in three places: fabricated authority, unvetted tools inside privileged workflows, and outcome prediction. The first is documented in court. In Mercante v. Tarzia, decided in the Eastern District of New York on October 2, 2026, defense counsel admitted using an AI research tool and copying propositions without checking them, and the court found a fictitious case, false quotations and a materially misrepresented precedent.

That order is one of the decisions collected in the AI Hallucination Cases database, which tracks decisions in which a court identified AI-generated fabrications in a submission.

How Each Limit Shows Up

1. Fabricated authority. A model that cannot retrieve an answer may produce one that is formatted like an answer, and a citation is the easiest thing in a brief to format convincingly.
2. Unvetted tools on privileged material. A general-purpose assistant with no stated position on retention or training is a confidentiality question before it is a quality question, which is what Rule 1.6 reaches.
3. Outcome prediction. Predicting how a judge or a jury will react is the least supportable claim made for legal AI: the data is thin, the variables are not observable, and a settlement recommendation built on a prediction of that kind is advice a lawyer cannot support if asked how it was reached.

What Litigation Analytics Can Honestly Do

Litigation analytics have a narrower honest use: describing what has happened in a venue, before a given judge or across a motion type. That is a record of past rulings rather than a forecast of this one, and the distinction is worth holding in a client conversation.

AI hallucination in law reaches a docket the same way each time: a model produces a citation, nobody opens it, and the brief is filed.

What a Legal AI Tool Has to Prove Before It Touches Client Work

The client confidentiality requirement a legal AI tool has to meet starts with Rule 1.6, which binds the lawyer and reaches all information relating to a representation, so the vendor’s written position on what happens to an input is the first thing that has to be in hand before the tool sees a client document. The other questions that decide whether a legal AI tool is fit for client work are what its answers are grounded in, how it behaves inside the software the team already uses, and what it shows when it is unsure.

The Four Questions a Vendor Has to Answer

1. Retention and training. Does the vendor state in writing that inputs are not retained and not used for training, and does that statement cover subprocessors as well as the vendor? A tool with no written answer here is difficult to clear for client work whatever it offers on the other three.
2. Grounding. Does the output come from the team’s own precedent, a licensed database or the open web, and can a specific answer be traced back to its source document?
3. Workflow fit. Does the tool run in the document the lawyer is already in, or does the work have to leave the drafting environment to be processed?
4. Behavior under uncertainty. Does the tool report low confidence and decline, or does it answer anyway, which is the behavior that puts fabricated authority in a filing.

These are confidentiality questions rather than procurement questions, which is why they sit closer to attorney-client privilege than to a feature matrix. A tool that cannot answer them in writing, with its subprocessors named, is not ready for client work, and the four answers above are what a vendor conversation should open with rather than close on.

Only once those answers are in hand does a comparison of legal AI tools mean anything, because a tool that fails the confidentiality questions is not a candidate however well it compares on capability.

How Spellbook Fits These Use Cases

Spellbook is an AI platform for commercial legal work built for AI contract review and AI contract drafting, and it runs inside Microsoft Word and integrates with Google Docs. The Review feature reads an inbound agreement against the playbook a team has defined and proposes redlines as native tracked changes, the Draft feature generates clause language in the document, and the Compare to Market feature reports how a term sits against a set of 20M+ real contracts.

Which of the Eight Use Cases It Covers

The Review, Draft and Compare to Market features cover three of the eight workflows above: drafting, review and benchmarking. Legal research, e-discovery and litigation support are other categories, and the tool is a commercial-contract specialist rather than a research or litigation platform. More than 5,000 in-house teams and law firms across 80+ countries use it, and pricing is custom per-seat by team size with a free seven-day trial rather than an implementation project.

A partner at KMSC Law reviewed a 300-page franchise agreement by surfacing every section mentioning cost, charge or fee along with the sections that could constitute a franchisee breach, then had the tool prompt him for the clauses he would expect to find. Todd Strang, a partner at the firm, put the result this way: “It’s given us a degree of efficiency we didn’t have before, and we were already pretty efficient.”

Try Spellbook free, or see how the Review feature works in Word on your own contracts.

[cta-3]

Frequently Asked Questions

What Can AI Be Used for in Law?

AI is used across eight workflows that hold up in practice: drafting contracts, reviewing inbound contracts against a defined standard, benchmarking terms against comparable agreements, legal research, document review and e-discovery, summarizing long documents and transcripts, tracking obligations and deadlines in executed contracts, and checking documents against a named regulatory framework. Accuracy is highest where the source material is in front of the lawyer and lowest in open-web research, where a general-purpose model can invent an authority that does not exist.

Can AI Be Used in Legal Cases?

AI can support work in an active matter, and courts have sanctioned lawyers who filed its output without checking it. The verification duty sits with the attorney who signs the filing, so every citation an AI tool produces has to be confirmed to exist, to say what it is cited for and to remain good law. Several courts and individual judges now also require disclosure of AI use in filings, so the local rules and the standing orders in the forum are worth reading before the first submission.

Does AI Change the Billable Hour?

AI changes how long a task takes, which puts pressure on hourly billing for work that compresses the most, such as first-draft production and document summarization. ABA Formal Opinion 512 addresses the consequence directly through Rule 1.5: a fee has to stay reasonable when a task that previously took hours now takes minutes, and a lawyer cannot bill the former hours for the latter minutes. Firms are responding with fixed fees on compressible work and hourly rates on the judgment-heavy parts.

Can Small Firms Use AI Without a Technology Team?

Small firms can adopt the contract and summarization use cases without technical staff, because the tools that matter run inside the word processor and the browser the firm already uses. The work that does require attention is not installation. It is deciding which matters a tool may see, reading the vendor’s retention terms against Rule 1.6, and writing down who checks the output before it goes to a client or a court.