Evaluating AI Legal Research Tools: A Vendor and Ethics Compliance Guide for Solo and Small Firms

If you're a solo practitioner or running a small firm, you've probably been pitched at least three AI legal research tools in the past six months. Lexis+ AI, Westlaw Precision with CoCounsel, Harvey, Spellbook—each promises to compress hours of research and drafting into minutes. Some of those promises are real. But the gap between a compelling demo and an ethically defensible deployment is wide, and most attorneys don't have a framework for crossing it.

The American Bar Association took a major step toward filling that gap in July 2024, when it issued Formal Opinion 512, its first formal ethics guidance on generative AI tools in legal practice. The opinion doesn't ban AI—far from it—but it makes clear that ignorance of how these tools work is no longer an acceptable posture. You need to understand the technology, verify its outputs, protect client confidentiality, and supervise its use just as you would any nonlawyer assistant.

This guide gives you a practical, four-part framework for evaluating AI legal research and drafting tools: (1) understanding the ethics floor established by ABA Model Rules and state bar opinions, (2) comparing vendor data-handling practices, (3) negotiating the contract terms that actually matter, and (4) designing audit-ready workflows. We've written it for attorneys who don't have a full-time IT director or a $50,000 Am Law budget for vendor due diligence.

The Ethics Floor: What ABA Model Rules and State Bar Opinions Require

Before you evaluate any tool, you need to understand the ethical obligations that apply regardless of which vendor you choose. Five ABA Model Rules form the core framework, and state bar opinions have begun translating them into AI-specific guidance.

Model Rule 1.1: Competence and Technological Literacy

Model Rule 1.1 requires lawyers to provide competent representation, which Comment 8 explicitly extends to keeping abreast of "the benefits and risks associated with relevant technology." ABA Formal Opinion 512 reinforces this: attorneys must understand how generative AI works at a reasonable level—not as engineers, but well enough to evaluate risks of hallucination, data exposure, and output limitations. As the UNC Law Library noted in its analysis of Opinion 512, "the days of blatant AI misuse are over" and maintaining technological competence with evolving technologies like AI is now a requirement.

The Texas Professional Ethics Committee made this explicit in Opinion 705 (February 2025), drawing on its earlier cloud-computing guidance in Opinion 680 to conclude that a lawyer using generative AI must "acquire a general understanding of how the technology works," review the terms of service, learn about data-security protections, and train staff on appropriate use. The opinion warns that lawyers "should not unnecessarily retreat from the use of new technology that may save significant time and money for clients"—but if they do use it, they must understand it.

Model Rule 1.6: Confidentiality of Client Information

Rule 1.6 is where most AI deployment risks concentrate. The duty of confidentiality extends to all information related to the representation—not just privileged communications. When you type a fact pattern into a generative AI tool, you may be disclosing client confidential information to a third party. Formal Opinion 512 warns that lawyers must understand whether a tool is "self-learning"—meaning it stores and incorporates user inputs into its training data—which could result in confidential information being revealed in responses to future queries by other users.

The Florida Bar's Ethics Opinion 24-1 (January 2024) was one of the first state-level AI ethics opinions and sets out concrete steps that remain among the most practical guidance available. Florida requires lawyers to research the program's policies on data retention, data sharing, and self-learning before using it with confidential information. The opinion specifically recommends that lawyers ensure the provider has an enforceable confidentiality obligation, investigate the provider's security measures and policies, and determine whether the provider retains information after service termination.

Model Rules 1.4 and 5.3: Communication and Supervision

Rule 1.4 requires lawyers to keep clients reasonably informed about the status of their matters. Formal Opinion 512 suggests that, depending on the use case and the sensitivity of the information involved, attorneys may need to inform clients that AI tools are being used in their representation and obtain informed consent before inputting confidential data. The California State Bar's Generative AI Practical Guidance similarly recommends disclosing AI use to clients, particularly when it involves confidential information.

Rule 5.3 extends supervisory responsibility to nonlawyer assistants—and by extension, to AI tools. You remain responsible for the work product generated by an AI tool just as you would for output from a paralegal or contract attorney. The Texas committee's Opinion 705 draws a direct parallel: the duty of supervision "extends not only to the lawyer's own employees but over entities outside the lawyer's firm with whom the lawyer contracts." If an AI tool produces a hallucinated citation, the supervising lawyer bears the ethical and professional responsibility.

The cautionary tale here is Mata v. Avianca, where two New York attorneys were sanctioned in June 2023 for filing a brief containing non-existent judicial opinions fabricated by ChatGPT. As the court noted, "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance"—but blindly trusting outputs without verification is professional malpractice.

Vendor Data-Handling: What Each Tool Does With Your Queries and Documents

The single most important vendor evaluation question is: what happens to the data you put in? Not all AI legal tools handle data the same way, and the differences have direct ethical implications under Rule 1.6.

Lexis+ AI

LexisNexis has positioned Lexis+ AI as a closed-system alternative to consumer AI tools. According to LexisNexis's published privacy and security commitments, the platform does not use customer prompts or uploaded documents to train its underlying AI models. Responses are grounded in LexisNexis's proprietary legal content database, which reduces hallucination risk compared to general-purpose LLMs. For solo and small-firm attorneys, the key advantage is that the data handling is designed for the legal market from the ground up—you're not adapting a consumer tool to a legal context.

Westlaw Precision with CoCounsel

Thomson Reuters acquired Casetext (the original developer of CoCounsel) in 2023 and has integrated it into Westlaw Precision. Thomson Reuters has stated that CoCounsel uses a "human in the loop" approach, with attorney editors who work with technologists to train the system and validate results. The platform is built on Thomson Reuters's proprietary legal content, similar to Lexis+ AI's approach. According to Thomson Reuters's product documentation, CoCounsel includes administrative controls for data access and retention, and the company has published security and privacy documentation for enterprise customers.

Harvey

Harvey is a newer entrant that has gained significant traction among larger firms. Its platform agreement explicitly states that the service is "a research tool, and its Output is not legal advice" and that "Output is AI-generated, and it may contain errors and misstatements or may be incomplete." Harvey's agreement includes a Data Processing Addendum and Security Addendum, and notably provides that Harvey cannot update its terms in ways that "detract from its obligations... with respect to Confidential Information, Customer Data, Customer Content, or security, without express written authorization." This is a stronger data-protection commitment than most consumer AI tools offer. However, Harvey is primarily enterprise-focused, and solo practitioners may find pricing and minimum seat requirements prohibitive.

Spellbook

Spellbook focuses on contract drafting and review rather than legal research. It integrates with Microsoft Word and uses OpenAI's underlying models. According to Spellbook's published privacy guidance, the platform offers enterprise customers a data processing agreement and does not use customer data to train its models when customers are on appropriate plans. Solo attorneys should verify which plan level triggers the no-training commitment—Spellbook's lower tiers may have different data handling terms than its enterprise offering.

The Key Distinction: Closed vs. Open Systems

The critical vendor evaluation question is whether the tool is a "closed" system—meaning your inputs are not used to train the model and are not accessible to other users—or an "open" system that incorporates user inputs into its training data. Consumer tools like the free version of ChatGPT historically used user conversations for training unless users explicitly opted out. Legal-specific tools from LexisNexis and Thomson Reuters are designed as closed systems. Harvey's platform agreement reflects a closed-system posture. But the burden is on you to verify this in the current version of the terms of service, because vendors update these terms regularly.

The Texas ethics committee's Opinion 705 puts the obligation squarely on the lawyer: "with any generative AI tool, the lawyer should be reasonably satisfied that the program will not reveal confidential information to others or permit the use of such information to the disadvantage of the client. If the lawyer is not so satisfied, the lawyer should—at a minimum—not input any confidential information to the program without client consultation and consent."

Contract Terms to Negotiate: Data Retention, Training Opt-Out, and Security Certifications

Once you've identified a tool that meets the ethics floor, the contract negotiation phase is where you lock in protections. Most solo and small-firm attorneys won't have leverage to negotiate a from-scratch MSA with LexisNexis or Thomson Reuters, but even standard enterprise agreements have negotiable terms—and smaller, newer vendors like Harvey and Spellbook are often more flexible. Here are the terms that matter most:

1. Training Data Opt-Out (Written, Not Toggle-Based)

Ensure the contract explicitly states—in the body of the agreement, not just a privacy policy—that your firm's inputs (queries, uploaded documents, and generated outputs) will not be used to train the vendor's models. A settings toggle that defaults to "on" is insufficient; if the vendor changes its UI or resets preferences, you could lose the opt-out without knowing. Florida Ethics Opinion 24-1 specifically advises lawyers to "determine whether the provider retains information submitted by the lawyer before and after the discontinuation of services or asserts proprietary rights to the information."

2. Data Retention and Deletion

Negotiate specific retention periods and deletion obligations. How long does the vendor retain your inputs after your session ends? After your subscription terminates? Is deletion cryptographic (irreversible) or logical (flagged for overwrite)? The vendor should provide written confirmation of deletion within a defined period after contract termination.

3. Security Certifications and Audit Rights

At minimum, require evidence of SOC 2 Type II certification. For tools that will handle particularly sensitive matters, ask for ISO 27001 certification as well. If you're a solo practitioner, you may not have leverage to demand a full audit right, but you can request the vendor's most recent SOC 2 report, which details their security controls and any exceptions noted by the auditor. Harvey's platform agreement references a Security Addendum, and LexisNexis publishes its security certifications publicly—both are good signs.

4. Breach Notification

The contract must require the vendor to notify you of any data breach affecting your firm's data within a defined window (72 hours is the standard under GDPR and many state laws). Florida Ethics Opinion 24-1 recommends ensuring that the provider "will notify the lawyer in the event of a breach or service of process requiring the production of client information." Without this clause, you may not learn of a confidentiality breach until long after your duty to notify clients has been triggered.

5. Subprocessor Disclosure and Limitation

Many AI legal tools rely on subprocessors—typically cloud infrastructure providers like AWS or Azure, and sometimes model providers like OpenAI. The contract should disclose all subprocessors, require notice before adding new ones, and flow down your confidentiality and security requirements. If the vendor uses OpenAI as a subprocessor, verify that OpenAI's enterprise API terms (which prohibit training on customer data) apply, rather than the consumer terms.

6. Indemnification for Data Exposure

Push for indemnification covering losses arising from the vendor's breach of its data protection obligations. Most vendors will resist broad indemnification, but a carve-out for confidentiality breaches caused by the vendor's negligence or willful misconduct is a reasonable ask—and one that smaller vendors are more likely to grant.

Output Verification and Hallucination Mitigation Protocols

Contracting with the right vendor is necessary but not sufficient. Every AI tool—regardless of vendor—can produce hallucinated outputs. Your firm needs a verification protocol that would withstand scrutiny if a bar grievance or malpractice claim ever questions your AI-assisted work product.

Citation Verification

Every case citation generated by an AI tool must be independently verified against a primary source. This means pulling the case from Westlaw, Lexis, or a court's website and confirming that the proposition for which you cite it actually appears in the opinion. Do not rely on the AI tool's built-in citation linking—verify independently. The Mata v. Avianca sanctions arose precisely because the attorneys failed to take this step.

Statutory and Regulatory Verification

AI tools can cite outdated or repealed statutes. Verify that any cited statute is current by checking the most recent codification. For regulations, verify against the current Code of Federal Regulations or state administrative code. This is particularly important in fast-moving regulatory areas where AI training data may lag behind amendments.

Document-Specific Verification for Drafting Tools

For tools like Spellbook that generate contract language, verify that proposed clauses are legally enforceable in your jurisdiction and consistent with the deal terms. AI-generated boilerplate can include provisions that are unenforceable under state law (non-competes in California, for example) or that conflict with the negotiated business terms.

The "Trust but Verify" Standard

Formal Opinion 512 establishes that lawyers must review AI-generated work product with the same level of care they would apply to work product from a junior associate or paralegal. As we've discussed in our AI Ethics for Lawyers compliance checklist, the supervising attorney remains fully responsible for the accuracy and completeness of all AI-assisted work. There is no "the AI made me do it" defense.

Designing Audit-Ready Workflows

The final piece of the framework is workflow design. An audit-ready workflow is one that produces contemporaneous documentation showing you exercised professional judgment at each stage of AI-assisted work. If your state bar, a court, or a malpractice carrier ever asks how you used AI, you should be able to produce a clear record.

Step 1: Tier Your Use Cases by Confidentiality Risk

Not all AI use involves client confidential information. Using an AI tool to summarize a public statute involves no confidentiality risk. Using it to analyze a client's confidential settlement agreement involves significant risk. Tier your use cases: green (no confidential information), yellow (general case theory or public legal questions with anonymized facts), red (client confidential information requiring informed consent and a closed-system tool). This approach aligns with our broader guidance on audit-ready AI governance for law firms.

Step 2: Document Your Vendor Due Diligence

Keep a file for each AI tool your firm uses that includes: the current terms of service, the data handling policy, security certifications (SOC 2 report), your analysis of whether the tool is a closed or open system, and any client communications about AI use. Update this file whenever the vendor updates its terms—Harvey's platform agreement notes that it provides 15 days' notice of material updates, which gives you a window to review changes.

Step 3: Log AI Use in the Matter File

For each matter where AI tools were used, note in the file: which tool was used, what task it performed, what verification steps the attorney applied, and the outcome. This doesn't require elaborate software—a simple notation in your practice management system or a log entry in the matter file suffices.

Step 4: Establish a Firm AI Use Policy

Even solo practitioners should have a written AI use policy. At minimum, it should specify: which tools are approved for client work, what types of information may and may not be entered into AI tools, the verification protocol for AI-generated outputs, and the procedure for obtaining client consent when confidential information will be used. This is also good practice for firms implementing broader lawyer-in-the-loop AI workflows.

Step 5: Train Everyone Who Touches the Technology

If you have staff or contract attorneys, they need training on the AI tools your firm uses and the ethical obligations that attach. Texas Opinion 705 identifies training as one of the four reasonable precautions a lawyer must take before using generative AI with confidential information. Document the training—date, topics covered, attendees—as part of your audit-ready record.

Actionable Next Steps

Here's what we recommend you do in the next 30 days:

1. Audit your current AI usage. If anyone in your firm is already using AI tools—including consumer tools like ChatGPT—identify what's in use, what data is being entered, and whether the current terms of service protect confidentiality.
2. Pick one legal-specific tool for a structured pilot. Lexis+ AI and Westlaw Precision with CoCounsel are the most established closed-system options. Request a demo, ask specifically about data handling and training practices, and pilot it on low-risk matters first.
3. Draft a one-page AI use policy. Specify approved tools, prohibited inputs, verification protocols, and client consent procedures. You can expand it later, but having something in writing now creates a baseline of firm governance.
4. Review your engagement letters. Consider adding a brief disclosure that the firm may use AI-assisted tools in the representation, with an offer to discuss any concerns. Some states may eventually require this; getting ahead of it now costs nothing.
5. Check your malpractice coverage. Confirm with your carrier that AI-assisted work is covered under your current policy, and ask whether the carrier has any specific requirements for AI use documentation.

The tools are powerful. The ethics obligations are clear. The gap between them is bridgeable—but only if you approach vendor selection and deployment with the same diligence you'd apply to any other aspect of your practice. ABA Formal Opinion 512 and the growing body of state bar opinions give you the framework. Your job is to implement it.