Can lawyers use AI on client files? What ABA Opinion 512 and Texas Opinion 705 require

What Opinion 512 asks of a firm

Formal Opinion 512, issued by the American Bar Association on July 29, 2024, answers the question managing partners ask first: can our lawyers put client files into an AI tool? The opinion is advisory, and your state's rules govern. Texas has its own opinion, covered below. Opinion 512 still works as a checklist anywhere, because it walks through the Model Rules one duty at a time. This guide is general information, not legal advice, and the responsible lawyer must check the rules and orders that govern each matter.

On competence, lawyers "need not become GAI experts", but they must have a reasonable understanding of the capabilities and limitations of the specific tool they use. Relying on output without an appropriate degree of independent review could violate Rule 1.1, and the right level of review depends on the tool and the task. The opinion's own example matters for any firm planning a build. A lawyer summarizing many long contracts need not review the whole set by hand if the tool was first tested for accuracy on a smaller subset. Whatever level of review the lawyer picks, the lawyer remains fully responsible for the work.

On confidentiality, the lawyer must evaluate the risk that client information is disclosed to or accessed by people outside the firm, and inside it too, such as lawyers behind an ethical wall. For tools that learn from what users type, the opinion says informed consent is required before client information goes in. It adds that general, boilerplate provisions in an engagement letter do not amount to informed consent. As a baseline, lawyers should read and understand the tool's terms of use and privacy policy.

Communication duties depend on the facts. A lawyer must tell the client about AI use if the client asks or the engagement terms require it, and should consult the client before client information goes into a tool or when AI use bears on the fee. The opinion adds a third trigger, output that will influence a significant decision in the matter. For court work, the opinion warns that even an unintentional misstatement to a court can be a misrepresentation, so analysis and citations are checked before anything is filed.

Two more parts of 512 shape how a firm sets AI up. Managerial lawyers must establish clear policies on permissible AI use and train people on them, and the usual outsourcing checks apply to any outside provider, covering confidentiality, security, breach notice, data retention and ownership. On fees, hourly billers must bill for their actual time, which includes the minutes spent prompting and the time spent reviewing the draft. A lawyer may not charge a client to learn a tool the firm uses across its practice.

Texas Opinion 705 and the proposed Rule 13 changes

The Professional Ethics Committee for the State Bar of Texas reached similar conclusions in Opinion 705, issued in February 2025. Texas Rule 1.01 "almost certainly does not require the use of generative AI", but a lawyer who uses it must have a reasonable and current understanding of the technology. The committee recommends reviewing, and potentially renegotiating, a tool's terms of service. Without reasonable satisfaction that a program will not reveal confidential information, a lawyer should not input that information without client consultation and consent.

Two lines in 705 matter most for a practice plan. Lawyers "cannot blindly rely upon or use answers given by generative AI tools", and a lawyer may not charge hourly fees for the time that was saved by using the program. Both lines drive the worked example later in this guide.

Texas courts may also add a signing rule. On June 26, 2026 the Supreme Court of Texas gave preliminary approval to a rewritten Rule 13 of the Texas Rules of Civil Procedure. Under proposed Rule 13(c)(1), a lawyer who signs a court filing represents that the filing was reviewed and its accuracy verified, including that legal citations, quotations and holdings are not fabricated. The Court's comment says this proposed text mainly addresses improper use of artificial intelligence in filings. Comments closed on September 1, and the Court said it expected the amendments to take effect on October 1, 2026. When this guide was written, the Court's rules advisories page listed no final approval order for the proposed Rule 13 changes, so treat the text as a proposal until one appears.

Court rules in North Texas

Federal courts in Texas already have AI rules. The Eastern District of Texas, which covers Plano and the rest of Collin County, says in Local Rule AT-3(m) that a lawyer who uses generative AI "continues to be bound by the requirements of Federal Rule of Civil Procedure 11" and the court's other standards, and "must review and verify all content" to make sure it meets those standards. The rule also cautions that certain technologies may produce factually or legally inaccurate content.

The Northern District of Texas, which includes Dallas, adds a disclosure duty. Local Civil Rule 7.2(f) requires a brief prepared using generative AI to say so on the first page under the heading "Use of Generative Artificial Intelligence". A brief without that heading certifies that no part of it was prepared with generative AI. A firm that files in both districts needs one step in its filing checklist that asks the drafting lawyer whether AI touched the brief.

What the Heppner ruling did and did not decide

In February 2026, Judge Jed Rakoff of the Southern District of New York ruled in United States v. Heppner that documents a criminal defendant prepared with the consumer version of Anthropic's Claude were protected by neither attorney-client privilege nor the work product doctrine. The defendant used the tool on his own and later sent the documents to his lawyers, who conceded they had not directed him to do it.

The court's reasons were specific to those facts. The tool was not a lawyer, and the provider's privacy policy allowed collection and disclosure to third parties, so the defendant had no reasonable expectation of confidentiality. Documents that were not privileged when created did not become privileged when sent to counsel. The court also rejected work product protection because no lawyer had directed the work. That reasoning came from a criminal case, and in civil cases Federal Rule 26(b)(3)(A) also covers qualifying material prepared by or for a party or its representative, so it should not be read as a general rule. Commentary from Venable and from the Washington Legal Foundation stresses that the decision did not address enterprise tools and did not hold that AI use waives privilege as such.

Our reading for a firm build, which is analysis and not a holding: run the agent inside the firm's own tenant under enterprise terms with no training on inputs, have it act at a lawyer's direction on a named matter, respect the matter permissions and ethical walls in the document system, and keep a log of what it read and produced. No setup can guarantee privilege. These steps line up with the confidentiality analysis in Opinion 512 and avoid the facts that sank the documents in Heppner.

The controls that answer these rules

Put together, the rules point to a short list of controls for any agent that prepares legal work. Outside Texas, the State Bar of California's 2026 Practical Guidance adds two limits for agentic tools: attorneys must not let AI systems file documents on their own, and must not let them send client information outside the firm without appropriate safeguards and human review. The table maps each control to the rule it answers. Ask any vendor or builder to show you each control working on your own files before a live matter goes near the tool.

A one-page firm AI policy

Many firms still have no policy at all. In 8am's 2026 Legal Industry Report, a survey of more than 1,300 legal professionals, 43% said their firm had no formal AI policy and no plans for one, and only 9% had a policy they actively enforce. Opinion 512 expects managerial lawyers to set one. A single page that everyone in the firm reads is enough to start, and it should cover these points:

• Which tools are approved, and on which account type, with consumer accounts barred for client work.
• What may go into each approved tool, from public material up to client files.
• Who approves a new tool or a new use, and how that approval is recorded.
• How output gets reviewed, including sample testing before a tool is used on live matters.
• Marking AI-drafted material when it is saved to a client file, as footnote 52 of Opinion 512 suggests.
• Engagement letter language that names the tools and the data involved, written for informed consent.
• The billing rule: actual time only, with tool costs treated as overhead unless matter-specific and disclosed.
• Reporting an incident, such as client data pasted into the wrong tool, to a named partner the same day.

Worked example: the billing and review math

This is an illustrative example with assumed numbers, not a client result. Take a 30-lawyer firm whose litigation group wants an agent to prepare first drafts of written discovery responses. Before go-live, the firm tests the agent on 20 closed sets and compares each draft with what was actually served, the sample-testing approach Opinion 512 describes. At an assumed 2 associate hours per comparison, that is 40 hours, spent once.

Now take one set on an hourly matter. Today it takes 6 associate hours at an assumed $300 rate, so $1,800 billed. With the agent, the associate spends 0.3 hours setting it up and 2.2 hours reviewing and editing, 2.5 hours in all, so $750 billed. The 3.5 hours saved cannot be billed, under Opinion 512 and Texas Opinion 705. On a flat-fee matter priced at an assumed $1,500 a set, Opinion 512 says it may be unreasonable to charge the same flat fee with the tool as without it, so the firm reviews its flat-fee schedule instead of keeping the old price by default.

The gain for the firm is capacity. Assume the group handles 5 sets a week and 4 are routine enough for the agent. Today that is 5 sets at 360 minutes, or 30 associate hours. With the agent, the 4 routine sets take 150 minutes each and the fifth still takes 360, so 960 minutes, or 16 hours. That gives back 14 associate hours a week, about 60 a month. At an assumed $85 loaded cost per hour, that is roughly $5,100 a month of capacity, and it turns into revenue only if the time goes to other matters.

Starting small with AI on client files

Write the one-page policy first, then pick one job and one approved tool. Pull ten or twenty closed files where you know the right answer and run the agent on copies inside the firm's systems, since closed files carry the same confidentiality duties as live ones. Then compare its drafts with what your team produced. The comparison is the evidence Opinion 512 asks for, and it shows your real routine share before anyone counts on the hours.

Our five-day AI Workflow Sprint runs roughly that test on redacted or synthetic samples. It costs $2,500 and the fee comes off the build, with a full refund if we do not find 10 staff hours a week. A build is from $7,500 per workflow, half to start and half when the acceptance criteria pass, and running it is $1,200 a month. Whoever you work with, ask to see the controls in the table above working on your own files.

Common questions