Joseph Tiano opines: “Law Firms Must Re-Engineer AI Use to Prevent Unauthorized Actions” —

• “As I read the news this weekend, I’m coming around to the view that the next-generation AI problem for law firms won’t be a hallucinated answer or an unverified court filing. It will be a capable tool that took an action nobody approved. Firms need to start thinking about this now.”
• “Two incidents from this week’s news show how general AI risks in other scenarios may translate over to the law firm context.”
• “Incident one. During testing, Claude was running example tasks on randomly selected webpages. It landed on PhillyUnsolvedMurders.com and filled out a form suggesting it had information about an unsolved murder. Police say the submission was marked as spam and never reached investigators. Apparently Anthropic reported the incident”
• “Incident two. In a report Friday, Anthropic disclosed that another model submitted forms to an undisclosed government website when it was supposed to stop before submitting. Anthropic calls this agent persistence.”
• “The failures differ, but they share a gap: no human approved the specific action. In the first, nothing limited where the agent could act. In the second, a stop-point existed and the agent went past it.”
• “AI didn’t create this gap. It exposed it. Today, if the duties of competence, diligence and supervision are followed, a lawyer is usually the last step before anything is filed or sent, so most firms never had to define anything beyond the final human approver control. Agents remove that step by default.”
• “Law firms have three options:”
  • “1. Resist. Ban agents, and watch shadow use appear anyway.”
  • “2. Patch. Issue a policy memo. (just like outside counsel guidelines, such a memo is a hopeful control at best.)”
  • “3. Re-engineer. Scope what each tool can reach, require human approval at the final step, and log every action.”
• Only the third option really holds up. Bans get routed around, and memos can’t be enforced particularly well.”
• “To re-engineer, I think it starts with an analysis of existing data pools: billing data, LPM data and DMs data where you can break each service into workflows by task and phase. Then map where an agent can act on its own and where it must stop for approval.”

Tiano also offers some compliance and risk management advice in another post: “California Regulates Lawyers Use of AI: Self Regulation, Delegation, and Liability Concerns” —

• “California just regulated how lawyers use AI. It’s well intended, but it raises a question I can’t stop asking as a legal ethics guy: what happened to lawyers regulating themselves? Was ABA Formal Opinion 512, our own guidance on generative AI, structurally inadequate?”
• “SB 574 takes effect January 1, 2027. Hundreds of decisions have sanctioned lawyers for hallucinated citations, so I understand the impulse. But SB 574 asks lawyers to meet standards the statute never defines, and the profession was not asked to define them. “
• “Lawyers may use AI ‘to assist in the practice of law’ but may not ‘delegate’ it. I kinda get the difference but the line is undefined, and so is the term it depends on. ‘Practice of law’ has never had a clean definition. Unauthorized practice of law (UPL) rules vary by state, courts often decide case by case, and the definition has long been murky at the edges: document preparation, legal information versus legal advice, and now software that drafts and analyzes. If ‘practice of law’ means everything lawyers do, no AI drafting is allowed. If it means legal judgment, the statute should say so.”
• “None of the four duties in the statute has a stated remedy, and the statute creates no cause of action. The only express remedy, sanctions under § 128.7, covers the citation rule alone, runs to opposing parties, and deters rather than compensates. So if a lawyer violates the statute, what can a client assert? Malpractice? Breach of fiduciary duty? Is a violation negligence per se, evidence of the standard of care, or nothing? And what are the damages? A fake citation may impact a court more than the client. A confidentiality breach raises harder questions: what is the loss, and was privilege waived? The statute is also silent on the witnesses and victims whose addresses and phone numbers it protects. And this is just a knee jerk reaction I had. “
• “What I’d do: measure and monitor: Log which tool touched which task, who verified what, and who made the legal judgment call, then monitor those records over time. When duties come without defined liability, enforcement defaults to State Bar discipline, malpractice carriers, and clients.”
  “Right away, I’d pull billing records from ten recent matters and ask whether you could show, from your own records, where a lawyer exercised judgment and where AI assisted. If you can’t, that gap is your compliance risk.”

And for those intrigued by these future-forward and emerging issues, see this fascinating thought exercise: “Law Department 2.0: When Clients Take Control (Part II)” —

• “What happens when clients really have command of their data?”
• “Law firms are governed by Outside Counsel Guidelines (OCGs). In addition to billing guidance, OCGs increasingly focus on how work gets done and when automation is expected. What if clients directed more of the systems that law firms use?”
• “What if firms were required to perform certain work within a client-controlled research environment, perhaps by leveraging a client instance of a legal research service? The client could gain additional insights into how the work was done, including search strategies and the AI prompts used to approach the matter. That information could help with risk assessment and decision-making. It would also become part of the client’s institutional knowledge so future matters have greater context.”
• “The same could hold true for workflows defined by the client using its preferred AI platform, or legal operating system, as some vendors call their platforms. Think Claude for Enterprise, Harvey, Eudia, or other solutions.”
• “In the future, panel participants might be required to use a ‘legal operating system’ to complete work. “
• “In AI, a harness is the surrounding infrastructure that constrains and directs a model’s operation. Perhaps OCGs evolve into something resembling Outside Counsel Harnesses (OCHs), defining not just billing rules but the technology and processes that firms must use to perform the work.”
• “Matter information already travels with a lateral between firms to put the client first. Wouldn’t it make more sense for the matter information to just stay with the client in the first place?”
• “This may be jolting for law firms, and to be fair, there would be much to think through. There will be situations where a firm has proprietary know-how and shouldn’t expect to give it up. Complicated situations, such as multiparty litigation, would pose challenges too. To be clear, I’m not suggesting that all work becomes controlled by the client, but the seesaw may tilt more in that direction. “
• “The very expertise that clients rely upon may atrophy if firms can’t benefit from the knowledge gained on a matter, too. That’s not good for anyone. It’s also in clients’ interest for firms to retain their ability to specialize. “
• “In the future, clients will require firms to conform even more closely to their preferences regarding how work is completed.”

And when AI meets Time Entry: “Finally, A Way To Measure AI ROI: Be Careful What You Ask For” —

• “In case anyone wasn’t listening, Legora, the legal AI provider, and Laurel, a work intelligence platform provider, recently announced a ‘partnership.’ A partnership that reportedly has two main parts. It will offer what sounds like a fairly standard timekeeping tool. But it holds out the promise of an intriguing future tool that raises some possibilities.”
• “What Legora and Laurel are promising is to give law firm management the ability to determine the ROI on their purchase of Legora. The ROI for most AI platforms has indeed proven elusive, except for administrative and nonbillable work where AI replaces human work. I can only assume those trying to determine Legora’s ROI have faced the same difficulty. Hence the effort.”
• “The whole premise is that the AI will do work, perhaps much of which had previously been done by paralegals or associates. That’s billable work, and for firms that use the billable hour model, that’s hardly a return on investment. To the contrary, it’s a loss. Try explaining at the next partners’ meeting that you’ve just spent a ton of money on something that reduces their year-end shares. And be ready for the question: is it too late to back out of this deal? Time saved is not necessarily money earned.”
• “The Legora/Laurel partnership presumably will measure the time saved by Legora AI. Which is easily convertible to dollars. The question then becomes: What does a law firm do with those dollars? Does it pass them on to the client? And what if clients want to see those numbers?”
• “As in, ‘We want to see the savings your tools came up with so we can determine whether our bills reflect that reduction.’ What happens when the tool for justifying the technology investment becomes negotiating leverage for the client? Once you start measuring AI’s value, the next question is who gets to keep it.”