The LexText Team
Updated:
On August 6, BARBRI published research built on interviews with ten leaders across nine firms, from global Am Law 100 partnerships to a firm operating outside the traditional model. Asked to grade their own technology rollouts, those leaders gave themselves a C. The finding underneath the grade was blunter: firms "know who has activated AI tools, but almost none of them know who has actually changed the way they work." [1][2]
That is a real finding. It is also being read wrong by nearly everyone who cited it this week.
The obvious take
The consensus reading landed within a day. Firms bought licenses, skipped the training, and now have expensive software sitting unopened on associate desktops. The fix, in that reading, is curriculum: competency frameworks, structured programs, an L&D function that actually talks to the innovation team. BARBRI's own framing supports part of it. No firm in the study had built the AI competency framework its associate pipeline needs, and the report names the billable hour as the most stubborn barrier to adoption, on the theory that nobody volunteers to write off the hours a better tool just saved. [1][3]
Every legal technology outlet ran some version of that story this week. It is not wrong. It is one layer too shallow.
The missing metric is verification
Firms are not failing to measure behavior change because they lack a framework. They are failing because license dashboards do not measure verification.
Seat activation is trivially available. Your vendor mails it to you monthly. It tells you a lawyer opened a tab.
It does not tell you whether the citation in the brief that lawyer filed on Thursday was checked against a source before it went out the door. Call that second number the verification rate: the share of AI-produced citations and factual propositions confirmed against a primary source before leaving the building. Almost no firm in the United States can produce it for last month.
Here is why that gap matters more than the training gap. Lawyers are already verifying. In a survey of 207 US plaintiff-side legal professionals conducted by Thirdside for Supio and released July 23, 99 percent said they will not use AI-generated content they cannot verify. Ninety-six percent described themselves as very or extremely concerned about untraced AI output. Seventy-six percent said direct integration with verified and authoritative legal research would raise their confidence in AI output. [4][5]
Read together, the BARBRI and Supio findings point to a more precise question. Lawyers say verifiability is essential, but firms often cannot see whether verification actually happens. License data shows who used the tool. It does not show whether AI-assisted citations and factual assertions were checked before the work left the firm. Under a compressed deadline, the firm may have no record of what was checked, by whom, or what was missed.
Uninstrumented work is unmanaged work. Unmanaged verification is precisely the failure surface courts have spent the last eighteen months writing opinions about.
The gap is not a failure of knowledge management. Most vendor dashboards are designed to measure adoption through active users, query volume, and feature use. Those metrics show whether a tool is being used. They do not show whether AI assisted work was verified before it reached a client or court. None is tied to a matter, a filing, or a verification decision. Firms need controls that record what was checked, how it was checked, and who approved the final work.
What this actually means for litigators
Picture the specific moment this fails. It is 4:40 p.m., the opposition is due at midnight, and a fourth-year has a draft with eleven cites in the argument section, nine of which she pulled herself and two of which came back from the model. She checks one. She means to check the other. The partner's redline lands at 5:15 and the section gets rewritten anyway. Nobody in that building will ever know which of those two cites went out unchecked, including her, and the firm's dashboard will still count her as an active user that month.
Three consequences, in ascending order of how much they should bother you.
Rule 11 is assessed at the filing, not at the license. Damien Charlotin's AI Hallucination Cases database, updated August 6, catalogs 1,847 decisions worldwide in which a court found that a party relied on hallucinated content. 1,278 come from the United States. [6] No court has ever asked a firm how many seats it provisioned. Courts ask who checked the cite.
Your AI process may become discoverable. On May 18, Magistrate Judge Thomas O. Farrish ordered the plaintiff in Conservation Law Foundation, Inc. v. Shell Oil Co. to produce any prompts or queries used by an expert and her team in preparing an expert report. The court reasoned that an expert’s methodology is discoverable and that, under the facts of the case, using AI to reduce the opposing party’s document production to a working subset was part of that methodology.
In United States v. Heppner, Judge Jed S. Rakoff held that a defendant’s independent exchanges with a publicly available AI platform were protected by neither attorney client privilege nor the work product doctrine. The defendant had used the platform without counsel’s direction, and the court relied in part on the provider’s privacy policy in concluding that the exchanges were not confidential.
One month later, Morgan v. V2X, Inc. distinguished Heppner. The court held that a pro se litigant’s AI assisted litigation preparation could receive work product protection, while still requiring him to identify the platform he had used with confidential discovery material.
These decisions do not establish a general rule that AI prompts are discoverable or unprotected. They show that courts are beginning to examine who used the system, for what purpose, under whose direction, with what information, and subject to which platform terms. Firms should be prepared to describe that process before a dispute forces them to reconstruct it.
You cannot defend a process you cannot describe. Which brings us to the only part of this post worth acting on.
Five questions. Run them Monday.
For last month, what percentage of AI-assisted citations in filed documents were checked against a primary source before filing?
Who checked them: the drafting associate, a second reviewer, or a tool?
Can you produce a record of that check, by matter, without asking anyone to reconstruct it from memory?
When a deadline compresses, which step gets cut first, and do you know how often that happened last quarter?
If a judge issued an order to show cause tomorrow asking how your firm verifies AI-assisted work, would your answer be a policy document or a log?
If the answer to question 3 is no, what you have is an AI policy, not an AI control.
Where LexText fits
LexText was built by litigators rather than by a general model vendor adding a legal veneer. Hallucination Guard checks every citation the system returns against the source before it reaches your draft, and says so plainly when a case does not exist. That check is logged. The verification trail is a byproduct of doing the work instead of a separate compliance chore somebody has to remember on a Friday afternoon.
That is the part that speaks to the measurement problem directly. A firm running LexText answers question 3 above by pulling a report, not by circulating a survey.
We are SOC 2 Type II. Stanford Law licensed us in February 2026. Our backers include former general counsel of OpenAI, Oracle, Uber, Twitter, Dropbox, Mozilla, and SoFi, a group with an unusually low tolerance for vendors who cannot show their work.
The billable hour rewards output, not verification
The billable hour is often described as a barrier to AI adoption. But adoption creates another operational question: how is verification performed, recorded, and reviewed?
Verification is real legal work. ABA Formal Opinion 512 recognizes that lawyers may charge for time reasonably spent reviewing AI generated work for accuracy and completeness. The problem is not that verification is inherently unbillable. The problem is that ordinary usage dashboards do not show whether it occurred.
That distinction matters when deadlines compress. A policy may require lawyers to review every AI assisted citation and factual assertion, but the firm may still have no reliable way to determine what was checked, how it was checked, or whether the review was completed.
Seat activation measures adoption. Documented verification shows whether a control was performed. Firms need visibility into both.
Close
BARBRI co-CEO Lucie Allen put it plainly: "Rolling out the technology is only the first step." [1] The second step is not a curriculum. It is a number. Pick the number a judge would ask you about, instrument it, and put it in front of your executive committee next to the license count.
If you want to see what verification looks like when it is logged rather than assumed, we will walk you through it on a live matter. Book a demo
Sources
Bob Ambrogi, "Law Firms Are Rolling Out AI Faster Than They Can Measure Changes in Lawyer Behavior, New BARBRI Research Finds," LawSites, August 6, 2026. https://www.lawnext.com/2026/08/law-firm-are-rolling-out-ai-faster-than-they-can-measure-changes-in-lawyer-behavior-new-barbri-research-finds.html (accessed August 9, 2026)
"Law Firms Are Rolling Out AI Faster Than They Can Measure Changes in Lawyer Behavior, New BARBRI Research Finds," LawFuel, August 6, 2026. https://www.lawfuel.com/law-firms-are-rolling-out-ai-faster-than-they-can-measure-changes-in-lawyer-behavior-new-barbri-research-finds/ (accessed August 9, 2026)
"Legaltech governance latest: New Akerman head, EU AI Act warning, & BARBRI AI adoption findings," Legal IT Insider, August 7, 2026. https://legaltechnology.com/legaltech-governance-latest-new-akerman-head-eu-ai-act-warning-barbri-ai-adoption-findings/ (accessed August 9, 2026)
"Trust, Not Cost, Is Blocking AI Adoption in Plaintiff Law, New Supio Report Finds," PR Newswire, July 23, 2026. https://www.prnewswire.com/news-releases/trust-not-cost-is-blocking-ai-adoption-in-plaintiff-law-new-supio-report-finds-302832991.html (accessed August 9, 2026)
"AI Realism, ILTACon, Legal Innovators New York + UK," Artificial Lawyer, August 7, 2026. https://www.artificiallawyer.com/2026/08/07/ai-realism-iltacon-legal-innovators-new-york-uk/ (accessed August 9, 2026)
Damien Charlotin, AI Hallucination Cases database, last updated August 6, 2026. https://www.damiencharlotin.com/hallucinations/ (accessed August 9, 2026)
"Produce the Prompts: A Court Says Expert AI Inputs Are Fair Game in Discovery," Alston & Bird Privacy, Cyber & Data Strategy Blog. https://www.alstonprivacy.com/produce-the-prompts-a-court-says-expert-ai-inputs-are-fair-game-in-discovery/ (accessed August 9, 2026)
"AI, Privilege, and Discovery in View of 'Heppner' and 'Morgan'," Sterne Kessler. https://www.sternekessler.com/news-insights/insights/ai-privilege-and-discovery-in-view-of-heppner-and-morgan/ (accessed August 9, 2026)