Skip to main content

An online professional reference on the use of generative artificial intelligence in legal practice.

Module 05

Evaluating Legal AI Tools

Choose by task, not reputation.

This module covers the AI tools available for legal work. We cover what tools exist, the concepts and technology behind them, and how to use them.

The Landscape

01

Legal Tech AI

Specialized tools built for legal work. Connected to legal databases with verified sources. Enterprise-grade security for confidential matters.

02

General AI

Versatile assistants for broad tasks. Drafting, brainstorming, summarization. Require more verification for legal accuracy.

Most lawyers will use both. The question is which tool to use for which task.

Here is a non-exhaustive list of legal AI tools that lawyers are familiar with. If you understand how these work, you can get a good idea as to what’s most effective for your practice.

Legal-tech capabilities

Research-first platforms and legal workflow systems.

Grounded ResearchAuthority LinksDocument AnalysisDrafting HelpWorkflow Actions
WWestlaw
CoCounsel
Lexis+ AI
HHarvey
Clio
Built-in
Partial / plan-gated
Not available

Legal-tech feature diff

Expanded side-by-side view of where the legal platforms actually differ.

Feature Diff
WWestlaw
CoCounsel
Lexis+ AI
HHarvey
Clio

Authority-grounded research

How directly the tool works from a proprietary legal corpus rather than generic web or model memory.

Native

AI-assisted research runs inside Westlaw authority.

Native

Research uses Thomson Reuters legal content, including Westlaw and Practical Law.

Native

Conversational legal research sits directly on Lexis content.

Assisted

Strong knowledge workflows, but research depth depends on firm setup and connected sources.

Limited

Not a primary legal-research destination; AI is oriented around matter operations.

Uploaded-document review

Whether the platform is built to take in your files and analyze, summarize, or extract from them.

Limited

Westlaw is strongest once you are inside its authority set, not as a general document-review surface.

Native

Built for document review, summarization, extraction, and legal analysis.

Native

Document Analysis is a first-class Lexis+ AI workflow.

Native

Document and contract analysis are core Harvey product surfaces.

Assisted

Matter documents drive summaries and next-step suggestions inside Clio workflows.

Drafting and rewrite help

How naturally the tool moves from research or matter context into producing editable work product.

Assisted

Research outputs can seed drafting, but drafting is not the main surface.

Native

Drafting is part of the main CoCounsel legal workflow.

Native

Lexis+ AI supports drafting, summarization, and uploaded-document workups.

Native

Ask, analyze, and draft are all core Harvey motions.

Native

Clio drafts client updates, motions, letters, and matter communications in context.

Matter and workflow automation

Whether the platform can turn outputs into next actions inside an operational legal system.

Limited

Westlaw is a research stack, not an execution system for firm operations.

Assisted

CoCounsel connects work across Thomson Reuters tools, but it is still assistant-first.

Assisted

Lexis+ AI helps with research and drafting more than end-to-end workflow automation.

Assisted

Harvey helps generate work product, but workflow automation depends on deployment pattern.

Native

Clio's AI is explicitly built to create tasks, deadlines, invoices, updates, and matter actions.

Source trace and citator posture

How directly the tool exposes source links, citator workflows, or validated legal provenance.

Native

Authority links and citator workflows are part of the core Westlaw posture.

Native

Outputs are framed around validated Thomson Reuters inputs and linked authority.

Native

Lexis+ AI pairs answers with Lexis sources and Shepard's workflows.

Assisted

Useful outputs still need source checking against the underlying authorities you trust.

Limited

Clio is not the place you go for citation provenance or citator depth.

General AI

General-AI capabilities

General assistants and answer engines used alongside legal-specific systems.

Web ResearchFile AnalysisDrafting HelpSource LinksReusable Workspace
ChatGPT
Claude
Gemini
Perplexity
Built-in
Partial / plan-gated
Not available

General-AI feature diff

Expanded side-by-side view of where the general assistants differ in day-to-day legal use.

Feature Diff
ChatGPT
Claude
Gemini
Perplexity

Live web and current-info research

How directly the tool is built to pull in current web information during the answer.

Native

ChatGPT can use web search for current, source-backed answers.

Native

Claude web search is designed for live-web grounding with citations.

Native

Gemini responses can draw on web-linked sources and double-check flows.

Native

Perplexity is built around web search and sourced answer synthesis.

File-grounded analysis

Whether the tool can take in your materials and work directly from them instead of only from chat context.

Native

Files can be uploaded for analysis, summaries, and shared workspace work.

Native

Claude supports uploaded documents in chats and project knowledge.

Native

Gemini supports uploaded documents, spreadsheets, code, images, and more.

Native

Perplexity supports attached files and follow-up questions within the same thread.

Persistent workspace and reusable instructions

How much the tool supports a standing matter setup rather than a one-off prompt.

Native

Projects group chats, files, and shared instructions under one objective.

Native

Claude Projects provide knowledge, instructions, and project-level retrieval.

Native

Gems package repeatable instructions and can be shared or edited.

Native

Spaces combine instructions, sources, and thread history in one workspace.

Source visibility

How clearly the tool shows where the answer came from when you are doing research work.

Assisted

Strong when search is invoked, but not every response is source-first.

Native

Claude web replies cite sources, and Claude also supports document citations.

Assisted

Gemini can show sources and related links when available, but not every reply has them.

Native

Perplexity positions cited answers and source links as the main interaction pattern.

Long-form drafting surface

How well the tool supports turning research into an editable drafting or artifact workflow.

Native

Canvas gives ChatGPT an explicit co-writing and editing surface.

Native

Artifacts and file creation make Claude strong for iterative deliverables.

Native

Gemini Canvas supports creating and editing docs, apps, slides, and code.

Assisted

Perplexity is strongest for research synthesis; drafting is useful but not its signature mode.

Features and pricing change frequently. Check each platform’s current product page for the most up-to-date information.

The Concepts Behind the Tools

Every AI tool listed above—legal tech or general—is built on the same set of underlying technologies. Understanding these concepts helps you evaluate any tool, including those that do not yet exist.

The Four-Layer Architecture

Most legal AI tools operate on four layers. The differences between tools come from how each layer is configured.

The Four-Layer Architecture

1
Interface

The chat window, document upload, and controls you interact with.

Determines the user experience and what workflows the tool supports.

2
AI Engine

The large language model that generates responses.

Different tools use different models. Most use third-party models like GPT or Claude.

3
Knowledge Base

The content the tool can access—legal databases, firm documents, the web, or only training data.

One of the biggest differentiators between tools. A tool like Westlaw augments its output with verified legal sources. A tool like ChatGPT does not.

4
Security Layer

Encryption, access controls, retention policies, training opt-outs.

Determines whether you can use the tool for confidential work. Review the TAR Check from Module 2.

RAG: Retrieval-Augmented Generation

RAG is a technology that most legal AI tools use to access a vast library of legal sources. Instead of relying solely on what the model learned during training, a RAG-enabled tool retrieves relevant documents from a database and includes them in the model’s context before generating a response.

Query"Title VII burden-shifting in 7th Cir."
RetrievalPulls top-k opinions from legal DB
ModelGenerates grounded in retrieved docs
AnswerSynthesis with inline citations

Retrieval-Augmented Generation. Every handoff can still fail: a missed document, an outdated result, or a misread passage.

Think of it as the difference between a closed-book exam and an open-book exam. A pure language model answers from memory, while a RAG-enabled tool can look things up before answering. The tool searches its knowledge base for documents relevant to your query, pulls them into the model’s working memory, and then generates a response grounded in those documents.

While legal tech companies have said that their RAG tools eliminate hallucinations, this does not seem to be the case. The retrieval system can miss relevant documents, return outdated ones, or surface results that the model misinterprets. The model can also ignore what it retrieved and generate from the training data anyway.

This is why hallucination rates for legal tech tools, while lower than general AI, are still significant. RAG improves the tools by giving the model verifiable sources that can help mitigate hallucinations, but it does not make them reliable enough to skip verification.

How a Westlaw-style RAG query works in practice

Retrieval narrows the source set first. The model answers second.

1. Query
Natural-language research ask

Has any federal court held that using a consumer AI tool can waive attorney-client privilege? Prioritize federal authority and explain the reasoning.

federal privilege waiver consumer AI draft legal materials
2. Retrieval
Westlaw ranks the authority set

The system searches the legal database, narrows by jurisdiction and topic, and selects a small set of opinions, headnotes, and secondary sources.

· Top result: federal privilege-waiver opinion
· Related headnote: confidentiality and third-party disclosure
· Secondary source: practice guide on AI-assisted drafting
3. Context pack
The model reads grounded excerpts

Only the retrieved passages and metadata go into the model context window, so the answer is anchored to specific authorities instead of pure model memory.

"Disclosure to a consumer AI platform may destroy confidentiality when the user accepts data-retention terms."

"Work-product analysis may differ if the tool is treated as an internal drafting aid rather than a disclosure to an adversary."

4. Answer
Synthesis with citations

The model returns a short synthesis with case cites and quoted reasoning, but the lawyer still needs to open the cited sources and verify the proposition.

Short answer: some courts treat consumer-AI use as a confidentiality risk, but the result depends on doctrine and platform terms.

Cites appear inline so the lawyer can open the cases and confirm the proposition before relying on it.

Cites: Heppner; Warner; Practice guide notes

RAG lowers hallucination risk, but it can still miss a case, retrieve stale material, or misread the excerpt it found. Verification stays outside the model.

Data Confidentiality Mechanisms

How your data is handled also depends on the platform, the tier, and the current terms of service. Module 2 covered the TAR Check. Here are some common protection mechanisms the tools advertise and what they mean.

Mechanism
What It Means
Zero Data Retention
Your inputs are deleted immediately after the response is generated. Nothing is stored. Available on some enterprise and API tiers.
No-Training Clauses
The provider guarantees your inputs will not be used to train future models. Standard on most business and enterprise tiers. Typically not the default on free tiers.
Data Processing Agreements
A formal contract specifying how data is handled, stored, and protected. Required for serious enterprise use with client data.
SOC 2 Compliance
An independent audit certifying the provider meets specific security standards.
Audit Logs
A record of who used the tool, when, and on what matter. Important for demonstrating supervision and compliance.

How the Concepts Map to Each Tool

Now that you understand the architecture, here is how each tool configures those layers. This is where the practical differences become clear.

Harvey AIHarvey
CoCounselThomson Reuters
Lexis+ AILexisNexis
ClioClio
ChatGPTOpenAI
ClaudeAnthropic
GeminiGoogle
PerplexityPerplexity AI