From Training to Execution: Embedded Safeguards for Responsible AI Use in Legal Practice

[EDRM Editor’s Note: EDRM is proud to publish the advocacy and analysis of the Hon. Ralph Artigliere (ret.). The opinions and positions expressed are their own. The author gratefully acknowledges the insights and suggestions of Professor William F. Hamilton, Rose Hunter Jones, Suzanne Clark, and Varun Perumal, which helped shape this work.]

Professional competence is not proven by what we remember to do in calm conditions, but by what well-designed systems help us execute under pressure.

There is no longer serious debate about whether artificial intelligence will be a central feature of legal practice. It already is. Lawyers, judges,¹ and legal professionals use AI to draft, analyze, summarize, and organize information at a pace and scale that would have been unthinkable just a few years ago. Recent survey data confirms that adoption is no longer theoretical. In 2025, Thomson Reuters reported that about half of professionals across legal, tax, accounting and audit, risk and fraud, and government were using generative AI in some fashion, and that use within legal organizations had nearly doubled year over year.² More recent law-firm data points in the same direction and expose a serious governance gap: the 8am 2026 Legal Industry Report reported that 69% of legal professionals use general-purpose generative AI tools for work, while 54% of law firms offer no AI training and 43% have no AI governance policy.³ As AI tools adapt to legal challenges, improve, and proliferate, they will become the norm for the profession, from top to bottom, in every type of legal practice and organization. The benefits of that adoption are real. So is the risk that the profession’s competence infrastructure—training, policies, workflow governance, and supervision—has not kept pace with the tools it is deploying.

The issue is not whether or when the profession will adopt AI. It is whether lawyers, judges, and legal professionals can develop and maintain the competence and judgment required to use it effectively and responsibly—and execute those duties consistently under the pressures of practice. Ethics authorities have correctly identified and emphasized competence, confidentiality, supervision, and verification.⁴ But the speed of AI development and the relative novelty of AI-assisted workflows in legal practice create a challenge that traditional professional education was not designed to meet. The path forward will require more than instruction alone. It will require systems and embedded safeguards that help professionals use the technology effectively and safely at the moment they engage it.

The question, then, is not whether legal professionals understand their obligations when using AI. Most do. The question is whether that understanding reliably survives contact with the pressures of actual practice. That challenge begins with the limits of the education model the profession has long relied upon.

I. The Limits of Traditional Education

Legal education has long relied on structured, periodic instruction. Continuing legal education, judicial education, law school curricula, internal firm training, and written policies all operate on a familiar model: gather professionals, present information, and expect retention and application over time. Since the emergence of generative AI, lawyers, judges, and law students have encountered that instruction in many forms, ranging from vendor blogs and short CLE or CJE presentations to university courses and more sustained hands-on training. Some of that education is quite effective.⁵ The difficulty is that even good instruction does not eliminate the gap between knowing what the rules require and executing them consistently under the pressures of practice, especially as the technology itself continues to evolve.

Assuming that CLE and CJE courses are made available in an effective way, that model has worked reasonably well for many core legal topics. Technology training and eDiscovery, however, have long presented special challenges in gaining attention, maintaining currency, and translating instruction into practice as the technology, data sources, and processes evolved. The technology competence gap among practitioners and judges in eDiscovery has been a persistent issue, as the profession’s experience with electronically stored information, technology-assisted review, and digital evidence has shown.⁶ But the challenge presented by generative AI is broader still because it reaches virtually every legal role and workflow.

Fourth, and most importantly, the problem is not simply one of access to education. It is the inability of educational efforts, standing alone, to ensure correct execution in practice. AI competence is not a one-time educational achievement. It is a continuously degrading condition unless it is reinforced at the point of use, especially when the underlying technology is itself changing rapidly.

II. The Real Risk: Knowledge Without Execution

The legal profession does not lack intelligence, training, or professional responsibility. Lawyers and judges are highly skilled professionals who routinely manage complexity and make consequential decisions. The more serious problem is that knowledge alone does not reliably translate into disciplined performance under real-world conditions.

Atul Gawande made an important distinction between failures of ignorance and failures of ineptitude. Some mistakes occur because professionals do not know enough. Others occur because they fail to make proper use of what they already know.⁹ That distinction helps explain much of what we are now seeing in legal AI use.

Many AI-related failures in law are not failures of ignorance. They are failures of execution. The lawyers involved generally know, at least in principle, that AI output must be verified and that confidential information and privilege must be protected. What fails is the disciplined application of those requirements in the press of actual work: under deadlines, under workload pressure, and amid the strong temptation to value speed and convenience over verification.

AI intensifies a problem already familiar from other high-stakes domains. Its output often arrives in a form that appears polished, authoritative, and complete, making it unusually easy to accept without sufficient scrutiny.¹⁰ The cognitive science literature describes “automation bias” as the tendency to over-rely on automated systems and under-exercise independent checking, especially when the system’s output is fluent, salient, and easy to accept.¹¹ That risk takes on special force in legal practice, where AI can mimic the conventional markers of reliable legal work: polished prose, plausible citations, confident framing, and organized structure.

That is why the recent hallucination sanctions cases are significant beyond their facts. Hundreds of cases involving hallucinated citations and misquoted authorities have been reported and are accelerating in number.¹² Courts and commentators have taken notice for good reason. But the deeper significance of these failures lies not in the erroneous citations themselves. Failures that have occurred among experienced lawyers and judges alike are examples of execution failure: lawyers and firms knowing the risks and requirements associated with AI use but not reliably accounting for them in practice. In Johnson v. Dunn, for example, the court imposed significant sanctions after false authorities generated by AI found their way into briefing, including public reprimands, disqualification from further participation in the case, and referral to disciplinary authorities. The case is especially instructive because the failure occurred in a sophisticated legal setting, reinforcing that policies and training do not by themselves ensure disciplined execution.¹³ Hallucinated citations are best understood as one visible symptom of that broader execution problem, not the central risk itself.

III. Lessons from Another High-Stakes Domain

There is a familiar, high-stakes domain that confronted this problem long ago: aviation. As a former aviator, I relied on checklists for every phase of flight: inspection, start-up, takeoff, landing, and emergency procedures. These were not optional crutches for the untrained. They were integral parts of safe operation. Checklists do not eliminate pilot judgment. They discipline and support it under pressure. Pilots do not use checklists because they are untrained. They use them because they are trained and understand the limits of human performance. Throughout an aviator’s career, checklists are not treated as bureaucratic ritual. They are welcomed as part of disciplined execution.

Aviation operates across a spectrum that maps more closely to the legal profession than many might assume. Commercial airline operations involve mandatory protocols, institutional oversight, and highly standardized procedures. Corporate and charter aviation often operate with significant but less uniform governance. Private pilots function under a regulatory floor but exercise substantial independence in day-to-day execution. Student pilots are still developing habits under supervision.

Across every tier, checklist culture is taught, expected, and practiced—not because an enforcement officer is present at every preflight, but because the profession internalized a fundamental truth: memory alone is structurally insufficient where complexity and consequence exceed reliable unaided recall.

The legal profession operates across a remarkably similar spectrum. Large law firms, corporate legal departments, and federal courts may adopt formal protocols and institutional oversight. Smaller firms, solo practitioners, chambers staff, and individual judges often operate with far more independence. Yet the cognitive problem is the same across the spectrum. A senior partner, a solo practitioner, a judge, and a new lawyer using AI all face the same underlying risk: complexity plus pressure can exceed reliable memory. That is the lesson legal AI practice should borrow from aviation. The value of checklists and structured verification does not depend on universal mandate. It rests on a deeper professional insight that systems are more reliable than memory under pressure.

IV. From Training to Systems

The legal profession is now confronting a comparable level of complexity in the use of AI. Lawyers are expected to understand how tools generate output, recognize hallucinations and incompleteness, identify bias, protect confidentiality and privilege, comply with ethical and court-imposed rules, supervise subordinate use, and verify all output before it is filed, sent, or relied upon.

Each of those obligations is manageable in the abstract. Taken together, across the full range of everyday legal work, they create a system of expectations that is difficult to execute flawlessly from memory alone.

Training is necessary. It is not sufficient. The profession’s divide is no longer between those who have AI and those who do not. It is between those who exercise judgment while using AI to its fullest advantage and those who mistake fluency for reliability.¹⁴

The next question is no longer how to teach these requirements in the abstract. It is how to help professionals execute them consistently in practice.

V. Embedded Safeguards: A Practical Framework

The profession should not think only in terms of AI training. It should also think in terms of embedded safeguards. These are structured interventions within the workflow that prompt, guide, or require verification at the point of use. They move education and compliance from the classroom, policy manual, or CLE outline into the actual performance of legal work. In that way, they reinforce education and help protect against predictable human error.

Importantly, embedded safeguard tools are not limited to enterprise-licensed platforms. Browser-based governance architectures can monitor and guide interactions across any AI interface—from institutional deployments to publicly available consumer tools—bringing consistent verification and compliance prompts to bear regardless of which tool the lawyer or judge is using.

Varun Perumal of OrcaWorcs AI described the value and necessity of embedded safeguards in legal work in a private communication shared with his permission:

“The gap between knowing the rules and executing them consistently under real-world pressure is precisely what periodic training cannot fix. Fatigue, context, and the novelty of the task at hand can erode even well-trained professionals at the moment of use. The durable answer is to move the guardrails into the workflow itself, so that verification, confidentiality checks, and bias flags surface at the point of action rather than relying on memory alone.”

For those reasons, some emerging tools are beginning to incorporate elements of this approach through just-in-time guidance within the interface itself, surfacing relevant policy, prompting verification, and logging interactions against governance frameworks in real time.¹⁵ That point matters because much unsupervised AI use in legal practice occurs not inside enterprise platforms, but in consumer tools accessed through ordinary browser sessions outside the organization’s formal compliance perimeter. The importance of that development is conceptual, not product specific. It suggests that the profession may be able to use AI not only as a productivity tool, but also as a system for reinforcing responsible AI use.

Viewed more broadly, a practical framework for embedded safeguards may be understood in three tiers.

This framework is the legal analogue to checklist culture in aviation and protocol culture in medicine. The point is not to reduce professional work to rote steps. It is to ensure that professional judgment is exercised consistently within a disciplined process, with human responsibility preserved at the points that matter most.¹⁶ This approach is designed not to replace judgment, but to make its disciplined exercise more reliable.

VI. Embedded Safeguards as Institutional Infrastructure

The safeguard framework described above was conceived in the context of today’s prompt-and-response paradigm. But the AI industry is already shifting toward more autonomous, workflow-oriented systems, which makes that framework more—not less—urgent.

AI’s value to the profession in all its forms lies chiefly in its strengths, not as a substitute for legal judgment. It can assist powerfully with large-scale review of information, identification of patterns and alternatives, and detection of communication, tone, or logical weaknesses. Properly integrated into legal workflows, it can also support responsible practice by prompting verification, surfacing safeguards, and reinforcing oversight at the point of use. But legal reasoning, professional judgment, and final decision making must remain human responsibilities. The distinction is more than rhetorical. AI performs what theorists of artificial intelligence describe as reckoning—the rapid computational structuring of information—but judgment, in the sense of responsible decision-making accountable to professional obligation, remains irreducibly human.¹⁷

When a system prompts source verification, warns against disclosure of sensitive information, flags missing support, or forces a pause before risky output is adopted, it is doing more than compliance work. It is reinforcing professional habits in real time. Education becomes iterative and embedded in the workflow itself. The legal professional becomes better at exercising judgment while using the tool.

Some safeguards are clerical in nature: they verify citations, preserve audit trails, or warn against disclosure of sensitive information. Others must be judgment-oriented: they require human review of reasoning, framing, strategic choices, and the sufficiency of support before AI-assisted work becomes professional work product. When safeguards are designed or adopted, their structure should reflect the risk profile of the tasks in which AI is being used. Prompts, warnings, pause points, and required review should be calibrated to protect against meaningful risk without imposing unnecessary friction that undermines efficient and thoughtful work.

The role of embedded safeguards becomes even more important as legal AI moves from stand-alone prompt-and-response systems toward more autonomous, workflow-oriented tools. Many current systems emphasize end-to-end workflows, automated drafting, review, analysis, and citation checking rather than isolated prompts and answers.¹⁸ As AI systems take on more complex sequences of work, the need for embedded verification points becomes more urgent, not less. Errors introduced early in a workflow can propagate quickly through later stages if no structured pause point exists. The more powerful AI becomes, the less sensible it is to rely on memory or after-the-fact review as the principal safety mechanisms.

The responsibility for building this culture, however, does not rest on individual lawyers alone. Vendors, firms, courts, bar organizations, judicial educators, and legal institutions all have roles to play. As AI becomes more embedded in legal workflows, responsibility for safe and disciplined use must likewise become more embedded in the systems and supervisory structures within which legal work is performed.

For lawyers, this framework fits naturally within existing supervisory doctrine. Embedding verification protocols at the workflow level is not merely a best practice. It is a structural discharge of the supervisory obligations that Rules 5.1 and 5.3 already impose, executed more systematically and with more documentation than ad hoc after-the-fact review.¹⁹ When AI is used in practice for document review, drafting, summarizing, research, and organization, the lawyer’s duties of supervision and review become more demanding, and the consequences of inadequate supervision become more serious. Embedded safeguards can be understood as a structural response to that supervisory burden, helping ensure that review and verification occur before defective output becomes professional work product. Without workable safeguard structures, the growing burden of supervision and perceived risk may chill beneficial AI use, leading organizations, supervising lawyers, and individual professionals to avoid not only higher-risk applications, but also lower-risk uses that could be responsibly adopted.

In the past, supervising lawyers could often rely more heavily on trust, training, and conventional review when delegating work to associates and staff. AI-assisted work changes that equation. For supervising lawyers, generative AI requires more than trust in one’s team, general training, and ordinary delegation. Those remain necessary, but they are no longer sufficient. When AI is used in legal workflows, supervision must be more structured and more clearly documented. Workflow safeguards and audit trails are not ancillary. They are part of how supervisory responsibility is now discharged in practice.

A concrete example from litigation practice shows how this can work without turning professional judgment into rote compliance. Rose Hunter Jones of Hilgers, PLLC, illustrates this point well. In her eDiscovery and litigation practice, she encourages responsible generative AI use by associates for defined tasks, but within a repeatable and documented playbook. When AI is used to synthesize key documents, testimony, or timelines in witness preparation or brief development, the associate must memorialize the process in a short workflow memo identifying the materials reviewed, ideally by folder or index, the prompts used, and the verification steps performed, including certification that each citation was checked against the source for accuracy and context. For briefs, the workflow memo, drafts, and source materials are kept together in a single working folder tied to the filing so that the process is transparent and auditable. In any organization, the form and detail of such a workflow memo should vary with the task, the risk profile, and the organization’s needs. Properly used, the memo communicates expectations, reinforces required steps, and creates a concise audit trail of how AI was used and verified.

As a further guardrail, Jones discourages string cites, both because they increase hallucination risk and because they multiply the supervisory burden on the lawyer who must ultimately verify the authorities and answer for the filing.

Beyond consistency, embedded systems offer something that ad hoc review cannot: a contemporaneous record. When a tool logs verification steps, flags acted upon, and decisions made at each stage of an AI-assisted workflow, it creates documentable evidence that the required process was followed— evidence that matters when competence is later questioned, when supervisory responsibility is reviewed, when AI-assisted workflows must be explained in negotiating an ESI protocol,²⁰ or when a court asks whether counsel exercised independent judgment before filing.

The same is true for judges and judicial staff. Judges and their staff are increasingly exposed to AI tools for research, drafting, case management, and administrative work. The governing framework differs from the Model Rules, but the execution problem does not. Courts, too, need a culture of structured verification if they are to take advantage of AI without compromising the integrity of the judicial process. AI may assist judges and staff in processing large volumes of information, summarizing materials, and supporting case management. But judges must retain full responsibility and accountability for the reasoning, accuracy, and basis of their decisions.

There is also a broader access-to-justice concern. If robust safeguard systems become available only through expensive enterprise platforms, the profession risks deepening a two-tier competence divide between well-resourced organizations and smaller practices. That is one reason courts, bar regulators, and professional organizations should be attentive not only to whether AI is used, but also to the quality and accessibility of the systems that govern its use. That risk also underscores the value of lightweight, tool-agnostic safeguard practices that smaller firms and courts can adopt without waiting for expensive enterprise solutions. Those practices should be taught and encouraged broadly across the profession so that responsible AI use does not become a luxury of scale.

CONCLUSION: The Path Forward

The legal profession has always adapted to emerging technologies through a combination of education, experience, and evolving rules. AI will require all those elements, but it will also require a change in emphasis.

Competence in the AI era will not be measured solely by what lawyers and judges know. It will increasingly be measured by whether their tools and workflows help them apply that knowledge reliably in practice. Highly trained professionals do not rely on memory alone. They rely on systems.

The path forward is not to train legal professionals once and expect compliance from memory. It is to build and adopt systems that reinforce responsible use, support human judgment, and make disciplined execution more reliable. If done correctly, AI will not only challenge the profession to improve its competence. Used for the tasks it performs well and constrained where judgment must remain human, it will help the profession reinforce and sustain that competence.

Notes

1. The use of AI by judges is under significant scrutiny and remains in transition as AI tools emerge and improve. See The Sedona Conference, Navigating AI in the Judiciary: New Guidelines for Judges and Their Chambers, 26 SEDONA CONF. J. 1 (2025). ↩︎
2. Thomson Reuters, 2025 Generative AI in Professional Services Report 5, 8 (2025), https://www.thomsonreuters.com/content/dam/ewp-m/documents/thomsonreuters/en/pdf/reports/2025-generative-ai-in-professional-services-report-tr5433489-rgb.pdf; see also Thomson Reuters, Generative AI for Legal Professionals: Top Use Cases (May 13, 2025), https://legal.thomsonreuters.com/blog/generative-ai-for-legal-professionals-top-use-cases-tri/ (reporting that legal organization use had increased from 14% in 2024 to 26% in 2025). ↩︎
3. Nicole Black, Generative AI Data from the 8am 2026 Legal Industry Report, 8am Blog (Mar. 20, 2026), available at https://www.8am.com/blog/ai-adoption-law-firms-2026-legal-industry-report/ (reporting that 69% of legal professionals use general-purpose generative AI tools for work, while 54% of law firms offer no AI training and 43% have no AI governance policy, “creating a widening gap between AI experimentation and institutional readiness”). ↩︎
4. Model Rules of Prof’l Conduct r. 1.1 cmt. 8 (Am. Bar Ass’n 2012) (requiring lawyers to keep abreast of changes in the law and its practice, including the benefits and risks of relevant technology); ABA Comm. on Ethics & Prof’l Responsibility, Formal Op. 512, Generative Artificial Intelligence Tools (July 29, 2024) (Opinion 512 addresses competence, confidentiality, communication, supervision, and reasonable fees); Fla. Bar Pro. Ethics Comm., Op. 24-1 (Jan. 19, 2024). ↩︎
5. Some courses are quite effective. For example, Professor William F. Hamilton’s eDiscovery course includes hands-on training in practical skills, including use of AI tools. ↩︎
6. See The Sedona Conference, Navigating AI in the Judiciary: New Guidelines for Judges and Their Chambers, supra note 1; R. Artigliere and W.F. Hamilton, From Competence to Judgment: How AI Compresses Litigation Work and Why That Makes Judgment More Important, EDRM (JD Supra, Apr. 2026), available at https://www.jdsupra.com/legalnews/from-competence-to-judgment-how-ai-4386570/. ↩︎
7. The term “risk profile” is used here in a practical governance sense, not as a rigid numerical score, but as a structured assessment of the risks associated with a given AI use case in context. That profile may change over time as risks are better understood, mitigated, or avoided through changes in the technology, the workflow, or the conditions of use. ↩︎
8. See ABA Formal Op. 512, supra note 4; Fla. Bar Pro. Ethics Comm., Op. 24-1, supra note 4. The ABA has emphasized that the degree of verification required is fact-specific and depends on both the tool and the task. ↩︎
9. Atul Gawande, The Checklist Manifesto: How to Get Things Right 9–13 (2009). ↩︎
10. In litigation, for example, AI’s most important contribution is not cognition. It is compression. AI collapses the time between raw data and apparent structure, making execution failures more likely. See Artigliere and Hamilton, From Competence to Judgment, supra note 6. ↩︎
11. See K. Goddard et al., Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators, 19 J. Am. Med. Inform. Ass’n 121, 121–27 (2012) (describing automation bias as the tendency to over‑rely on automated decision support and under‑exercise independent judgment). ↩︎
12. Damien Charlotin’s AI Hallucination Cases Database tracks legal decisions involving generative AI hallucinations, including false citations, fabricated descriptions of case holdings, and misattributed judicial language. As of April 29, 2026, the database reported 919 such cases in the United States and 1,356 worldwide. See Damien Charlotin, AI Hallucination Cases Database, available at https://www.damiencharlotin.com/hallucinations/. ↩︎
13. Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. 2025) (publicly reprimanding three attorneys, disqualifying them from further participation in the case, referring the matter to the Alabama State Bar and other applicable licensing authorities, and emphasizing counsel’s responsibility for false AI-generated authorities despite counsel’s experience, firm resources, and internal AI policies). See also R. Artigliere, When AI Policies Fail: The AI Sanctions in Johnson v. Dunn and What They Mean for the Profession, EDRM (JD Supra, Aug. 2025), available at https://www.jdsupra.com/legalnews/when-ai-policies-fail-the-ai-sanctions-9043268/; ABA Formal Op. 512, supra note 4; Fla. Bar Pro. Ethics Comm., Op. 24-1, supra note 4. ↩︎
14. “The emerging divide in the legal profession is not between those who use AI and those who do not. It is between those who exercise judgment in its use and those who mistake fluency for reliability.” Artigliere and Hamilton, From Competence to Judgment, supra note 6. ↩︎
15. See OrcaWorcs AI, Accord: AI Governance & Compliance Layer 1–2 (internal product overview, on file with author) (describing a browser‑native architecture providing just‑in‑time compliance prompts, real‑time policy propagation, and interaction logging). This reference is offered solely as an example of an emerging governance architecture and is not presented as an endorsement or independent validation of efficacy. ↩︎
16. Brian Cantwell Smith, The Promise of Artificial Intelligence: Reckoning and Judgment 108 (MIT Press 2019) (quoting John Haugeland). AI performs “reckoning” (computational structuring), but only humans perform judgment (responsible decision-making). See also Artigliere and Hamilton, From Competence to Judgment, supra note 6. ↩︎
17. Id. ↩︎
18. See LexisNexis, General Availability of Lexis+ with Protégé Sets New Standard for Automating Legal Work with Easy-to-Use Authoritative AI Workflows (Feb. 24, 2026), https://www.lexisnexis.com/community/pressroom/b/news/posts/general-availability-of-lexis-with-protege-sets-new-standard-for-automating-legal-work-with-easy-to-use-authoritative-ai-workflows (describing purpose-built, end-to-end legal AI workflows supported by agentic AI capabilities, including conversational research, personalized legal drafting, document upload, summarization, analysis, and Shepard’s citations). This reference is offered solely as an example of an emerging agentic platform architecture and is not presented as an endorsement or independent validation of efficacy. ↩︎
19. See ABA Formal Op. 512, supra note 4 (addressing supervisory responsibility in the use of generative AI tools under the Model Rules). ↩︎
20. A solid and documented validation workflow can be invaluable when justifying AI use in negotiating an ESI protocol or in explaining the process to a judge, if needed. See Sam Bock, 3 Adversaries You Might Meet Negotiating an AI-Friendly ESI Protocol, Relativity Blog (Apr. 16, 2026), available at https://www.relativity.com/blog/3-adversaries-you-might-meet-negotiating-an-ai-friendly-esi-protocol/. ↩︎

April 30, 2026 © 2026 Ralph Artigliere. ALL RIGHTS RESERVED (Published with permission.)
Assisted by GAI and LLM Technologies per EDRM’s GAI and LLM Policy.