Generative AI for Lawyers: What Actually Works in 2026

Generative AI for lawyers has moved well past the proof-of-concept stage. The 2026 General Counsel Report from FTI Consulting and Relativity puts in-house generative AI usage at 87%, up from 44% a year earlier. That is not a trend; it is a shift in how legal work gets done. But one number cuts against the optimism: only 22% of those users report high trust in what the tools produce.

That gap is the subject of this guide. Not whether generative AI belongs in legal practice, but how to work with it accurately and responsibly, and which tools hold up when the document that comes back matters.

Why adoption outpaced trust in legal AI

The jump from 44% to 87% happened because the tools became easier to access and the perceived cost of being left behind went up. Associates started using ChatGPT for initial research summaries. Partners asked for contract clause comparisons in a fraction of the usual time. The outputs looked good, which encouraged more use.

What took longer to surface was the error rate. A peer-reviewed study published in the Journal of Empirical Legal Studies in April 2025 measured hallucination rates on legal research tasks. GPT-4, without legal-specific tuning, hallucinated on 43% of queries. Purpose-built platforms did better: Lexis+ AI at 17%, Westlaw AI-Assisted Research at 33%. Better, but still not the near-zero standard that legal work expects.

The takeaway is not that AI is unreliable by nature. It is that reliability varies significantly by task type and by how the tool was trained. Understanding that distinction is the starting point for any useful AI policy inside a legal team.

What generative AI actually does in legal practice

There are two distinct jobs generative AI takes on in legal work, and they behave differently enough that treating them as the same task is a recurring source of problems.

The first is document analysis: reading a contract, an NDA, or a term sheet and extracting information. Flagging a missing limitation of liability clause. Identifying that the governing law is set to Delaware when both parties are UK-based. Comparing a vendor's standard terms against a master service agreement template. This is the area where AI performs most consistently, because the task involves pattern recognition within a defined document, not generation of new legal conclusions.

The second is legal research: finding how courts have interpreted a clause, surfacing relevant case law, understanding how a regulation applies to a specific fact pattern. This is where hallucination risk rises sharply. The model generates answers that look authoritative but may reference cases that do not exist, or misread the holding of a real decision. The 43% figure above was measured here, not on document extraction tasks.

In practice, the safest workflows treat these as separate activities with different verification requirements, not as one continuous AI task.

Contract drafting and review: where the numbers hold up

For transactional lawyers, AI-assisted contract review has delivered measurable results. Platforms built specifically for this task report 70 to 85% reductions in time spent per contract. That range deserves some scrutiny: it typically applies to routine agreements, standard NDAs, and high-volume procurement contracts where clause patterns are predictable and the model has been extensively trained on similar documents.

The efficiency gain is real. Where it can mislead is in unusual or bespoke agreements: joint venture structures with complex earn-out provisions, cross-border arrangements where the applicable law changes the risk profile of standard language, or heavily negotiated commercial contracts where every clause has a history. Here, the AI may return fewer flagged issues than expected, not because there are none, but because the patterns fall outside its high-confidence zone.

Spellbook is one of the more practical implementations for transactional work. It operates directly inside Microsoft Word, learns from a firm's existing contract templates, and handles clause suggestions, redlining, and risk detection within the drafting environment rather than requiring a context switch to a separate platform. The limitation is the same as for any tool built on a language model: it performs best when the document type is one the model has seen in volume.

For in-house teams reviewing vendor agreements at volume, purpose-built platforms offer deeper workflow integration, including obligation tracking, automated reminders for renewal dates, and records of which clauses were negotiated away from standard positions. The audit trail that comes with these platforms also matters for corporate governance, where showing what was reviewed and when has its own value.

Legal research with AI: the verification layer you cannot skip

Using AI for legal research carries a higher burden of verification than most practitioners expect when they first try it. The 43% hallucination rate cited above was not measured on obvious questions. It was tested on tasks resembling real research queries, where the output had to include accurate citations and correct legal reasoning on a specific point.

This does not make AI research assistance useless. It means the workflow must include a confirmation step that is treated as mandatory, not optional. Every case cited must be checked against the original source. Every statutory reference must be traced to the current version of the text. Every summary of a court's holding must be verified against the actual decision.

Lexis+ AI and Westlaw's AI layer show lower hallucination rates than general models partly because they operate within vetted databases and retrieve actual documents rather than generating citations from memory. The retrieval-based approach reduces the risk. It does not eliminate it, and practitioners who use these platforms have reported finding errors even there.

For smaller firms or practitioners without institutional access to Lexis or Westlaw, general AI tools can still accelerate the research process: identifying the right search terms, drafting initial research memos for attorney review, or summarising long decisions in readable form. None of that replaces checking the source, but it does reduce the time from starting a research task to having something reviewable.

How to choose between a general AI tool and a specialised legal platform

The practical question most legal teams face is not whether to use AI but which tool to reach for, and when.

General-purpose AI tools are accessible, flexible, and low-cost. They handle drafting, summarising, and generating first versions of documents reasonably well. They are less reliable for jurisdiction-specific legal analysis or precise citation research without human verification.

Purpose-built legal platforms are trained on legal documents and workflows. They tend to integrate with the tools lawyers already use, offer audit trails, and in many cases have made specific commitments on data confidentiality that matter for client information. The tradeoff is cost and setup time: a specialised platform requires more investment to deploy and configure correctly.

The practical split that many teams land on: general AI for internal drafting, summarisation, and initial analysis; specialised tools for anything that will be shared with counterparties or relied upon for a business decision. That boundary is worth making explicit in any AI policy rather than leaving it to individual discretion.

Jurisdiction also matters. A tool trained primarily on US case law and contract templates will give inconsistent results on English law MSAs or Swiss governing law clauses. For EU, UK, and CH-based practices, this is a point worth checking before committing to a platform.

Building a verification workflow that holds up under pressure

The 89.5% positive ROI figure in the FTI/Relativity report applies to teams with high trust in AI output. The 27.8% ROI figure applies to teams without it. The difference is not that some teams have better tools. It is that they have clearer protocols.

A workable verification framework for generative AI in legal practice has three steps.

First, define the task type before starting. Contract clause extraction and obligation identification are lower-risk tasks. Legal research that requires accurate case citations is a higher-risk task. The human review effort required scales accordingly, and treating both as equally reliable is a policy failure waiting to surface.

Second, check outputs against primary sources. For contract analysis, that means reading the flagged clause in its full document context, not just the extract the tool returns. Missing the sentence before and after a flagged clause is how the context changes the risk assessment entirely.

Third, track where errors occur. If a specific tool consistently misses a clause type in SLA agreements, that is information worth capturing and acting on. Over time, that log shapes which tasks the tool handles with lighter supervision and which require closer review by a qualified practitioner.

This is not glamorous work. It is the difference between a firm where AI adds reliable capacity and one where it adds a new category of risk to manage.

Three questions to ask before committing to any legal AI tool

These questions come up consistently when a firm is evaluating a new platform, and the answers separate tools that work in production from those that look good in a demo.

Where does the data go? AI tools that send client contract content to train a shared model create a confidentiality risk. Most enterprise-grade legal platforms now offer data isolation and contractual commitments on how content is stored and used, but the terms vary considerably. Reading the data processing addendum before signing is not optional, and verifying it against your bar association's guidance on client data handling is worth the hour it takes.

What jurisdiction does the model know well? This is the question that gets skipped most often in evaluations, and it matters most for cross-border practices and teams working under EU, UK, or Swiss law. Ask the vendor for specific examples of how the tool handles common clauses under your governing law. A tool that works well on California choice-of-law clauses may not apply the same precision to English law limitation periods.

What happens when it is wrong? Not if, but when. The relevant question is whether the workflow catches the error before it leaves the firm. A tool with high recall (it flags everything, including false positives) is preferable to one with low recall (it misses real issues) when the stakes are high. Understanding a tool's failure mode before deployment is more useful than understanding its success rate.

Generative AI is now a standard part of legal practice. The question is not whether to use it, but whether the protocol around it is solid enough to catch what the model gets wrong. In a practice where the cost of an error is measured in client outcomes rather than content quality, that is worth checking before the next contract lands in your inbox.