The legal profession has a specific relationship with AI that differs from most knowledge work: the consequences of error are governed by a formal ethics framework, the confidentiality obligations are enforceable by a bar authority, and the value of a legal professional is partly defined by judgment that cannot be systematized. That context matters enormously for figuring out where AI actually belongs in a legal workflow versus where it creates risk.
Most coverage of AI in law either overstates what the tools can do or dismisses them entirely. Neither is useful. The accurate picture is more specific: AI is genuinely saving time on a defined set of tasks, it is unreliable on a different set of tasks, and the distinction between those two sets is something every lawyer who touches these tools needs to understand cold.
Contract Review: Where AI Helps and Where It Fails
Contract review is the use case legal AI vendors lead with, and there's a real reason for that. Reading a 60-page commercial agreement to flag non-standard terms, identify missing provisions, and spot potentially problematic language is time-consuming, cognitively draining work that follows a relatively consistent pattern. It is also exactly the kind of structured-analysis task that AI handles well.
Tools like Harvey, Ironclad AI, and Kira (now part of Litera) are purpose-built for this. General-purpose models like Claude and GPT-4 can also do meaningful contract analysis if you give them a clear framework. What works: asking AI to compare a contract against a specific standard (your firm's preferred positions on indemnification, limitation of liability, IP ownership, governing law) and flag deviations. What also works: asking it to identify what's missing from a contract of a specific type. A software services agreement that lacks an SLA provision, a data processing addendum, or a clear termination for cause mechanism is something AI can reliably catch.
What requires human scrutiny: jurisdiction-specific enforceability questions. AI does not reliably distinguish between what a contract says and whether a court in your jurisdiction would actually enforce it. Non-compete provisions are a sharp example. A clause that looks clean on its face may be unenforceable in California, partially enforceable in Texas, and subject to blue-penciling in New York. AI will not consistently surface this. The substance of the language is one question; the enforceability question is yours.
The same limitation applies to regulatory compliance in contracts. If a vendor agreement touches health data and you need to evaluate HIPAA compliance, or if a financial services contract needs to be read against Reg S-P requirements, AI can surface relevant provisions but cannot reliably evaluate whether they satisfy the regulatory standard. That analysis requires a human who knows the regulatory landscape.
Legal Research: The Westlaw and LexisNexis AI Layer
Both Westlaw and LexisNexis have added AI summary layers to their core research products. Westlaw's AI-Assisted Research and Lexis+ AI surface relevant cases and generate plain- language summaries of holdings and key reasoning. These tools are trained on verified legal databases, which addresses the most dangerous failure mode of general AI in legal research: hallucinated citations.
The Mata v. Avianca case in 2023 put the hallucination problem in the national conversation. Attorneys submitted a brief citing cases that didn't exist, generated by ChatGPT. The platform had invented plausible-sounding citations with real-sounding parties, dockets, and holdings. The court sanctioned the attorneys. This is not a theoretical risk. It is a documented failure mode that happens when lawyers use general-purpose AI for citation generation without verification.
The workflow that works: use platform-native AI tools (Westlaw AI, Lexis+ AI) for initial case identification because they are grounded in verified databases. Use the AI summaries to quickly assess relevance before reading full opinions. Never cite a case you haven't read. Treat AI-generated research summaries as a starting point for reading, not as a substitute for it. The efficiency gain is in faster triage, not in skipping the primary source.
For background research on areas outside your practice area, general AI (Claude or GPT-4) can give you a useful orientation to the legal landscape before you go into Westlaw. A prompt like "explain the current state of federal preemption doctrine in the context of state data privacy laws" can get you oriented in five minutes. You then verify everything with primary sources before relying on it.
Harvey AI: What It Actually Does
Harvey is the highest-profile legal AI product built specifically for law firms. It has raised substantial venture funding and signed partnerships with Am Law 100 firms including Allen and Overy and PwC Legal. The product is essentially a GPT-4 class model fine-tuned on legal documents and integrated into law firm workflows, with security and confidentiality architecture designed to meet firm requirements.
Harvey's strongest use cases are document drafting, contract analysis at scale, due diligence support (summarizing large volumes of documents in M&A), and matter-specific research where the documents are uploaded into the system rather than retrieved from an external database. The platform keeps data within the firm's environment, which addresses the confidentiality concerns that make sending client documents to consumer AI tools problematic.
The realistic assessment: Harvey is genuinely useful for high-volume document work at firms that can afford the enterprise pricing and have the IT infrastructure to deploy it properly. For a solo practitioner or small firm, the tooling cost doesn't match the use case. For a practice group running M&A diligence on 500-document data rooms, it changes the economics meaningfully.
Specific Tasks That Save 2-4 Hours Per Week
Even without Harvey or similar enterprise tooling, there are specific tasks where AI integration into a legal workflow produces consistent time savings.
Standard agreement first drafts. If you draft the same agreements repeatedly, NDA, MSA, consulting agreement, vendor contract, you can use AI to generate a first draft from a one-paragraph description of the deal and then edit against your preferred positions. The draft will require substantive editing. But editing a 90-percent draft is faster than starting from scratch or pulling from a prior matter and hunting through tracked changes. For a routine NDA, this workflow saves 30 to 45 minutes per matter.
Engagement letter templates. Engagement letters follow a consistent structure with variable fields (client name, matter description, fee arrangement, billing terms, conflict check language). AI can generate a complete engagement letter from a structured prompt, and you maintain a library of templates you've reviewed and edited. Drafting one from AI output takes 15 minutes. Starting fresh takes 45.
Deposition preparation. Given a deposition transcript, AI can generate a summary of witness testimony organized by topic, identify internal inconsistencies, and pull specific quotes relevant to a set of issues you define. For a 300-page transcript, this can reduce the time to prepare a usable summary from several hours to under one. You still need to read the sections that matter. But getting oriented quickly in a long transcript is genuinely faster with AI.
Client update emails. After a hearing, a settlement negotiation, or a significant procedural development, you need to communicate what happened and what comes next in terms a non-lawyer client will understand. AI can draft these from your bullet notes. The output typically needs tone adjustment and specificity added, but the structural work is done.
What AI Gets Wrong in Legal Work
Beyond the citation hallucination problem, AI has specific failure patterns in legal contexts that practitioners need to know. It tends to generate confident-sounding statements about legal standards that are either outdated, overgeneralized, or incorrect for the specific jurisdiction. It does not know about recent case law unless it has been trained on or given access to it. Its knowledge of state-level procedural rules, local court rules, and judge-specific preferences is unreliable.
AI is also unreliable on ethical rules. Legal ethics varies by state bar, and the model rules are implemented inconsistently. If you ask an AI to review a conflict-of-interest scenario or evaluate whether a particular fee arrangement complies with ethical rules, the output is not trustworthy without independent verification. This is a domain where the cost of error includes discipline proceedings, not just malpractice liability.
The Confidentiality Question
ABA Formal Opinion 512, issued in 2023, addressed AI and lawyer confidentiality obligations. The core guidance: lawyers must take competent steps to prevent the disclosure of client information, which means understanding how the AI tools they use handle data and obtaining client consent when appropriate. The opinion does not prohibit AI use, but it does require diligence.
What firms are doing in practice: larger firms are deploying enterprise tools with data processing agreements that specify confidentiality protections and prohibit training on firm data. Many prohibit the use of consumer tools (ChatGPT, Claude.ai on the consumer tier) for client matters by policy. Some are using on-premise or private-cloud deployments that keep data entirely within the firm's environment.
For solo practitioners and small firms: the safest approach is to use AI only with de-identified documents, to review the terms of service and data handling policies for any tool you use with client information, and to check whether your state bar has issued guidance. Several state bars have issued formal opinions or guidance documents since 2023. The ABA's 2023 opinion is the federal baseline; your state bar's position may be more restrictive.
The practical answer for most practitioners: general AI tools are appropriate for tasks that don't involve client-specific confidential information. Document templates, research orientation, legal writing improvement, and workflow support don't require feeding client data into external systems. Contract analysis and deposition summarization do. Know the difference and apply appropriate diligence to each.