The Texas Bar AI Opinion and the Confidentiality Standard for Legal AI

The Texas Bar AI Opinion and the Confidentiality Standard for Legal AI

The State Bar of Texas Committee on Professional Ethics issued Opinion No. 690 in March 2024, and it deserves a closer read than most firms have given it. On the surface, it looks like another "lawyers can use AI but be careful" advisory. Dig in, though, and you find something more specific and more useful: a confidentiality standard that effectively draws a bright line for how law firms can and cannot deploy generative AI tools.

If you practice in Texas or serve Texas clients, this opinion is now the governing framework for your AI adoption decisions. And even if you are outside Texas, the reasoning here is likely to ripple outward as other state bars finalize their own guidance.

What Opinion 690 Actually Says

The opinion addresses whether a lawyer may use generative AI tools in the course of representing clients, and it anchors its analysis in Texas Disciplinary Rules of Professional Conduct 1.05 and 1.01. Rule 1.05 is the confidentiality rule. Rule 1.01 is the competence rule. The committee treats both as independently binding constraints on AI use, but the confidentiality analysis is where the real teeth are.

The core holding: a lawyer may not input confidential client information into a generative AI tool unless the lawyer has a reasonable basis to believe that the information will be protected from unauthorized disclosure. The committee explicitly flags that free or consumer-grade AI tools, where inputs may be used for model training or are accessible to the tool provider's employees, do not meet this standard.

This is not a suggestion. Under Rule 1.05(b)(1), a lawyer shall not knowingly reveal confidential information of a client or former client. The committee reads "reveal" broadly enough to encompass submitting text to a third-party AI system that retains, processes, or trains on that data. If the AI vendor's terms of service allow them to use your inputs for any purpose beyond providing the service back to you, you have a confidentiality problem.

The competence prong is also worth noting. Under Rule 1.01, a lawyer must provide competent representation, which the committee interprets to include understanding the technology you are using. You need to know how the AI tool handles data, where it stores inputs, and whether outputs might be hallucinated. Blind reliance on AI-generated legal analysis, without independent verification, is a competence violation.

The Confidentiality Bright Line

The most operationally significant part of Opinion 690 is the confidentiality framework, because it creates a functional test that firms can apply to every AI tool in their stack. The test boils down to three questions:

• Does the tool's provider retain client inputs? If yes, you need contractual assurances that retention is limited, encrypted, and not used for training or product improvement.
• Can the provider's personnel access client data? If yes, you need the same kind of protections you would require from any third-party service provider handling privileged information, including confidentiality agreements and access controls.
• Is there a risk that client information could appear in outputs served to other users? This is the training data question. If the model learns from your inputs and surfaces that learning to others, you have disclosed confidential information. Full stop.

The committee does not prescribe specific technical architectures. But the practical effect of these three questions is to rule out most consumer AI products for any use involving client data. OpenAI's ChatGPT free tier, Google's Gemini consumer product, and similar tools all have terms of service that, at minimum, allow input data to be used for model improvement. The enterprise tiers of these products often have different terms, but you need to actually read them and confirm they meet the standard.

This also applies to AI features embedded in other software. If your document management system or email client has added generative AI features, you need to evaluate those features under the same framework. The fact that you already have a vendor agreement for the underlying software does not automatically cover the AI component.

The Implementation Playbook for Texas Firms

So what does compliance with Opinion 690 look like in practice? Here is a reasonable implementation framework based on the opinion's requirements and the disciplinary rules it interprets.

1. Audit Your Current AI Usage

Start with the assumption that attorneys in your firm are already using AI tools, because they almost certainly are. A 2024 Thomson Reuters survey found that 51% of lawyers reported using generative AI in their work, and a meaningful percentage of those were using consumer tools without firm approval. Your first step is an honest inventory. What tools are in use? What data is going into them?

2. Evaluate Vendor Agreements Against the Three-Question Test

For every AI tool that touches client data, pull the terms of service, the data processing agreement, and any applicable security documentation. Map each vendor against the three questions above. If a vendor cannot provide clear, contractual commitments on data retention, personnel access, and training exclusion, that tool cannot be used with client information under Opinion 690.

3. Establish a Firm AI Policy

The policy should specify which AI tools are approved for use with client data, which are approved only for non-client work (like internal research or marketing drafts), and which are prohibited entirely. Be specific. Naming approved tools is more effective than listing general principles and hoping attorneys apply them correctly.

4. Implement Technical Controls

Policy alone is insufficient. If you have prohibited the use of ChatGPT's free tier for client work, but any attorney can open a browser tab and use it, your policy is aspirational rather than operational. Consider network-level controls, endpoint monitoring, or approved AI platforms that route all usage through a controlled environment.

5. Train on Verification Obligations

The competence requirement under Rule 1.01 means attorneys must verify AI outputs. This is not optional, and it is not a general best practice; it is a disciplinary obligation. Your training should cover how to spot hallucinated citations (the Mata v. Avianca lesson from the Southern District of New York in June 2023 remains instructive), how to evaluate AI-drafted analysis for accuracy, and when AI-assisted work product requires the same level of review as work from a junior associate.

6. Document Everything

If a grievance is filed and the question is whether you took reasonable steps to protect client confidentiality when using AI, you want a paper trail. Documented vendor evaluations, a written AI policy, training records, and technical control logs all support a defense of reasonableness.

A Note on Informed Consent

Opinion 690 does not explicitly require client consent for AI use in every circumstance, but it strongly implies that disclosure is advisable when AI plays a significant role in the representation. Several other state bars, including Florida (Proposed Advisory Opinion 24-1) and California (Practical Guidance for the Use of Generative AI, November 2023), have been more direct about consent obligations. Texas firms would be wise to get ahead of this and consider adding AI disclosure language to engagement letters now, rather than retrofitting later.

How FirmAdapt Addresses This

FirmAdapt was built for exactly this kind of regulatory constraint. The platform's architecture ensures that client data submitted to AI models is not used for training, is not retained beyond the session, and is not accessible to vendor personnel outside of narrowly scoped support functions governed by contractual confidentiality obligations. These are not optional settings; they are defaults. That design maps directly to the three-question test that Opinion 690 effectively establishes.

For firms implementing the playbook above, FirmAdapt also provides the documentation and audit trail infrastructure that supports a reasonableness defense. Usage logs, data handling certifications, and policy enforcement tools are built into the platform, so compliance is a byproduct of normal operations rather than a separate administrative burden.