Key highlights:

• • • AI governance is hard to prove in practice — While our research shows that 44% of companies publish an AI strategy, 76% of those same companies show no evidence of having policies to evaluate the quality of data used to train AI systems.

    • Workers are being left under-prepared and under-protected — Only 14% of companies have policies to mitigate the negative impacts of AI on workers, and only 31% offer any reskilling or training programs around adapting to an AI-integrated workplace.

    • Human rights and ethics appear an afterthought in AI governance — Almost three-quarters (72%) of companies conduct no AI impact assessments, and less than 1 in 10 companies conduct ethical or human rights assessments.

There is a widening chasm at the heart of corporate AI governance, according to a new report, , published by the ¶¶ŇőłÉÄę Foundation and the United Nations Educational, Scientific and Cultural Organization (UNESCO).

The Foundation’s analyzed publicly available information from nearly 3,000 companies across 11 industry sectors, creating the most comprehensive picture yet of how organizations are managing AI.

Beneath the surface of corporate AI governance mechanisms, divergence between the speed of AI adoption and meaningful human oversight is growing. The report’s findings make clear that this is no longer a gap that organizations can afford to ignore, especially when backlash against is growing and are solidifying among consumers in the United States.

Data highlights the illusion of AI governance

Businesses of different sizes and across multiple sectors are adopting AI technology at a rapid pace. When governance exists only in the wording of a strategy or company vision, however, the people most affected by AI systems — workers, consumers, and communities — are left vulnerable. According to the report:

• • • 44% of companies publicly communicate having an AI strategy. However, a gap in AI governance is evident as more than three-quarters of those companies (76%) do not seem to have policies to evaluate the quality of data used to train AI systems.
    • 40% of companies report board- or committee-level oversight of AI. At the same time, strategic signals do not necessarily indicate operational capacity or day-to-day governance. In fact, less than one-third of all sampled companies claim to have an additional team or resource dedicated to AI governance. Moreover, limited information is publicly disclosed on the teams, processes, and accountability mechanisms that translate intent into action.

Workers are being left behind

Research by the International Monetary Fund finds almost , highlighting the acute nature of concerns about job displacement and declining opportunities for some groups. Without sufficient oversight, AI can threaten workers’ rights, amplify bias, and increase surveillance and work intensity, which can enable inhumane decision-making at scale.

The TR Foundation/UNESCO report notes that many companies are adopting AI without the safeguards needed to support workers and help them to adapt to the changes this technology brings. Less than one-third of companies were shown to offer training and reskilling programs for employees who may be adapting to an AI-integrated workplace. Even within the 31% of organizations in which these training programs exist, there is a vast variation in the scope and depth of the training offered.

In fact, many company training programs are not enterprise-wide or structured. Instead, they are ad-hoc or limited to leadership roles. This lack of investment in talent risks undermining the significant investment that companies are making in AI.

Despite growing pressure from regulators, policymakers and social justice campaigners, the ethical impact of AI appears poorly governed, with companies sharing limited information publicly.

The picture on worker protections is equally concerning. Only 14% of companies have public policies in place to mitigate the negative impacts of AI systems on workers, the report shows. This means the majority of companies either have no policies in place or do not publicly communicate them.

What is more troubling is that when workers experience harm, there is almost nowhere for them to turn. Only 2% of companies indicated they had a complaints mechanism — a critical early warning system for potential concerns. The findings suggest many organizations lack a mechanism for AI-related internal complaints beyond the broad generic complaint channel, and this is compounded by low awareness of the areas in which AI systems may infringe employees’ rights and protections.

Ethics and human dignity as an afterthought

Despite growing pressure from regulators, policymakers and social justice campaigners, the ethical impact of AI appears poorly governed, with companies sharing limited information publicly.

Human rights and ethical use of AI are treated as secondary considerations to compliance, according to our research. The majority of companies (72%) do not conduct any impact assessment with regard to AI. Only 7% publicly communicate conducting a fundamental or human rights impact assessment, and just 5% report conducting an ethical impact assessment.

Among those companies conducting some form of impact assessment, the focus skews sharply toward compliance rather than people. The most prevalent assessments are privacy or compliance-focused, with 18% of those companies that conduct some form of impact assessment reporting that they conducted a data protection impact assessment, and 14% reporting they conducted a privacy impact assessment.

How to center people in AI governance

Closing this governance gap is essential for companies in order to adopt AI responsibly and avoid costly legal, ethical operational, talent-related risks.

To support companies in navigating this challenge, offers a free survey to help companies map the areas in which AI is used across products, operations and services, and then benchmark those against peers their sector.

The report also contains case studies from companies that voluntarily shared their responsible practices with us. For example, German software company SAP intentionally designs and deploys its internal AI systems with a human-in-the-loop in which AI automates repetitive tasks and supports decision-making while final judgment and complex problem-solving remain firmly in the hands of employees.

As AI becomes part of core business infrastructure, companies must move beyond statements of intent and toward measurable AI governance.

In another example, BASF, a German chemical conglomerate, has jointly agreed with its workers’ councils on a general reskilling program that covers technical, hard, and soft skills. Finally, Canadian telecom company TELUS’ Indigenous Advisory Council provides guidance on AI ethics issues that directly affect indigenous communities.

Next steps for companies

The TR Foundation/UNESCO report highlights the most impactful concrete commitments that companies can take now to future proof against AI-related risk, including:

• • • investing in structured, enterprise-wide worker-reskilling programs that measure outcomes, not just participation;
    • establishing enforceable human rights impact assessments as a standard part of AI deployment, not as an optional addition; and
    • creating accessible, AI-specific internal grievance mechanisms so that workers and users have a genuine pathway to raise concerns and seek remedy.

As AI becomes part of core business infrastructure, companies must move beyond statements of intent and toward measurable AI governance. While this data demonstrates clear governance gaps, it also presents an opportunity for companies to take the lead on implementing responsible AI that operates openly in the public interest.

You can learn more about

Key insights:

• • • Cloud modernization created accumulation, not transformationĚý— Many enterprises scaled infrastructure faster than they integrated systems, leaving fragmented data, duplicated processes, rising costs, and weak links between IT investment and business value.

    • AI and regulation now expose weak data architectureĚý— Agentic AI, real-time decision making, and regulatory reporting depend on consistent, traceable, well-governed data — not fragmented systems or after-the-fact governance.

    • Enterprise architecture must be rebuilt around outcomes and economicsĚý— Instead of treating the cloud as the strategy, organizations should define business outcomes first, structure data as a reusable asset, and measure architecture by revenue, cost efficiency, regulatory accuracy, and decision speed.

In this two-part blog series about the current state of cloud architecture, we look into where this architecture has failed and, in the next part of the series, what the possible remedies might be.

For the better part of the last 15 years, IT enterprise architecture definition and management didn’t disappear, it was deprioritized and replaced by as-a-Service solutions. The rapid rise of cloud platforms such as Amazon Web Services, Microsoft Azure, Snowflake, and Google Cloud made it possible to stand up infrastructure, deploy applications, create advanced databases, and scale environments without the same level of architectural rigor that was once required. Speed replaced structure, and access replaced integration.

Today’s cloud realities

The problem is that what was built during this last technological explosion was not architecture — it was accumulation. Systems expanded, data proliferated, budgets exploded, and organizations convinced themselves that connectivity was the same as coherence, and that data replication was the same as the system-of-record.

These assumptions have now been exposed as false positives. Over the last three years, AI, real-time decision making, and regulatory transparency have fundamentally changed the requirements. These are not technologies that sit on top of fragmented environments, they are data-driven capabilities and outcomes that depend on precision, integration, and sequence. The arrival of agentic AI and its stringent objective-based principles cannot tolerate data ambiguity and fragmented architectural designs.

The problem is that what was built during this last technological explosion was not architecture — it was accumulation.

AI does not fail at rates exceeding 80% because models are weak, it fails because the underlying data is inconsistent, inaccessible, or economically misaligned. Regulatory frameworks do not struggle because rules are unclear, they struggle because data cannot be traced, reconciled, or produced in real time. What the cloud enabled — rapid deployment without disciplined integration — is exactly what now constrains performance. The issue is no longer whether systems can scale, but whether they can produce measurable, consistent, and adaptable outcomes.

This is where enterprise architecture returns, but just not in its previous form. The discipline cannot simply revert to academic frameworks and abstractions that were designed for a different software era. It must be rebuilt around a different sequence that sees business outcomes first, data second, and then systems engineered within those constraints. Today, enterprise architecture must be defined and managed by economic KPIs, value added, and its adaptability to rapidly changing business realities.

Where the model broke

The failures experienced today in cloud architecture are not singularly technological. Cloud platforms deliver exactly what they promise — scalable, resilient, highly available infrastructure. Rather, the failure is architectural, and more precisely, involves the economics of compartmentalized capabilities.

Enterprise value is not created at the infrastructure layer. It is created where data informs decisions, and decisions drive outcomes. By over-rotating toward infrastructure, organizations optimized the least differentiating component of the enterprise stack, while leaving the highest-value layers largely untouched.

The result is a structural imbalance in which data remains fragmented across domains, business logic continues to operate in silos, governance is applied inconsistently and often retroactively, and measurement frameworks fail to tie technology activity to financial performance.

In this model, the cloud amplifies existing conditions. If fragmentation exists, it scales fragmentation. If inefficiency exists, it scales inefficiency. Modern infrastructure, applied to legacy architecture, produces modernized dysfunction.

What makes the cloud’s illusion particularly persistent is that its failure is rarely framed in economic terms. Cloud investments are justified through technical metrics such as uptime, latency, migration progress, and consumption efficiency. And while these are necessary, they are not sufficient. They do not answer the only question that ultimately matters: “Did the investment improve the economics of the business?”

Enterprise value is not created at the infrastructure layer — it’s created where data informs decisions, and decisions drive outcomes.

In many cases, the answer is no — at least, not in a way that can be clearly articulated. Instead, organizations experience cost expansion without proportional productivity gains, increased data duplication that drive storage and processing inefficiencies, extended timelines for analytics and reporting despite real-time capabilities, and persistent manual intervention in regulatory and operational workflows.

The absence of a direct line between architecture and outcome creates a vacuum often filled with disconnected KPIs, measurement solutions, and most recently, AI-automation. And with this interoperable vacuum, activity and speed have been mistaken for progress.

Figure 1: Cloud accumulation meets enterprise architecture shifts

The data reality beneath the surface

The cloud did not fail to deliver transformation; rather it exposed why transformation had not occurred — and at the center of this exposure is data.

Most enterprises operate with data architectures that were never designed for interoperability, reuse, or regulatory-grade consistency. Definitions vary by function, pipelines are purpose-built and duplicative, and governance is layered on after the fact. Automation was designed using business rules, then software architectures, then what the data needed. Therein resides the structural disconnect for enterprise architecture in AI solutions: They are out of order.

When these legacy conditions are moved to the cloud, they do not improve, they accelerate. The organization gains speed without alignment, scale without standardization, and access without coherence. For regulated industries, this creates a compounding risk of inconsistent outputs across reporting channels, increased reconciliation overhead, reduced confidence in data lineage and auditability, and slower response to regulatory changes.

What appears to be a technology issue is, in fact, a failure of data design.

Reframing the problem

To move forward, the premise must change. The cloud is not the strategy; rather it’s the environment. Transformation does not occur when systems are moved, it occurs when the relationship between data, decisions, and outcomes is fundamentally redesigned.

This requires an organization-wide shift from infrastructure-led thinking to what is defined as value architecture. Simply put, value architecture includes data that is structured as a reusable, governed asset — not a byproduct of applications, and business outcomes that are defined upfront and used to drive architectural decisions. Its governance is embedded at the point of data creation and distribution, and it replaces redundancy by making reuse the primary scaling mechanism. Finally, measurement is tied directly to financial and operational impact.

This is not a rejection of the cloud; rather, it’s a repositioning of its role and value proposition.

The implication is both direct and unavoidable. If your current strategy cannot clearly articulate how technology investment improves the economics of your business, then your organization is operating within the cloud illusion. However, this is not a critique of past decisions. It is a recognition that the next phase of transformation requires a different operating model — one that explicitly connects architecture to economics. Moving forward, what was forgotten in the past is now a future core competency.

Most organizations using as-a-service software had assumed that the cloud provider, vendor, or combination of those dealt with the complex liabilities of making designs interoperable. The implication moving forward — as well explore more in the second installment of this series — is that service software architectures using the system ideation approaches within AI silos are failing miserably, and there are few who understand the designs and skills needed to guide enterprises in the future.

You can find more blog postsĚýby this author here

Key insights:

• • • Law firms have moved from restricting GenAI use (Don’t use tools that leak client data) to mandating it (Incorporate AI into your practice and market our firm’s GenAI capabilities)—ĚýNeither phase has given rank and file lawyers what they really need: Guidance on in which instances GenAI actually helps deliver better, cheaper, and faster legal services, where it introduces serious professional risk, and how to tell the difference.

    • GenAI’s capacity to transform legal work for the better is real, but so is its capacity to degrade itĚý—GenAI can significantly boost speed and quality on tasks involving breadth, synthesis, or straightforward analysis, but it can weaken performance on complex judgment and revision tasks — especially for stronger professionals — by encouraging overconfidence, missed issues, and superficial reasoning.

    • A use-mode framework can close theĚýgap— A proposed governance framework can give law firm leadership a practical tool for identifying in which situations GenAI enhances legal work, where it introduces serious risk, and where professional judgment is non-negotiable.

This article synthesizes findings from the author’s paper,

Your law firm undoubtedly has a policy around generative AI (GenAI), which probably tells lawyers to avoid tools that leak client data, admonishes them to look out for hallucinations, and encourages them to incorporate AI into their practice to satisfy client demands.

However, it likely does not tell them which cognitive functions they should delegate to GenAI, which they should not, and where the line between the two is absolute. In the space between restriction and mandate, lawyers are making consequential decisions about GenAI delegation every day. Meanwhile, most law firms have not addressed that space with meaningful governance.

GenAI can make legal work worse

GenAI’s capacity to transform legal work for the better is real, but so is its capacity to degrade it. Most law firm leaders know that AI can hallucinate; yet far fewer know that it can make expert legal judgment and work product actively worse.

The best evidence of this dynamic comes from a with consultants from the Boston Consulting Group, who were given similar tasks and allowed to use various levels of AI assistance, including no AI. For professional tasks requiring breadth and option generation, GenAI delivered, showing that output quality improved by 40% and consultants worked faster. For tasks requiring judgment and synthesis, however, something unexpected happened. Consultants using GenAI were 19% less likely to produce correct solutions than those working without it.

Governing GenAI’s uneven performance requires asking a question that most law firms are not asking: What cognitive function is being delegated to GenAI at each step in the workflow?

The same pattern appears in research evaluating GenAI use in legal analysis. An empirical in the Journal of Legal Education confirmed that AI dramatically improves performance on straightforward analysis while producing no measurable benefit for complex reasoning. And in the case of complex reasoning, GenAI use also introduced recurring failures, such as jumping to conclusions, missing less obvious issues, and generating confident prose that masks superficial analysis.

from the University of Minnesota focused on legal tasks showed that GenAI assistance on a synthesis task improved performance by nearly 60% and produced a surprising downstream benefit. Those participants who used AI for synthesis outperformed the control group on the subsequent independent reasoning task even after GenAI was removed. However, when GenAI was introduced at the revision stage, the picture changed. GenAI helped weaker performers, but it actively degraded the work of stronger ones. Indeed, the best lawyers in the study produced worse revised work product when they used GenAI than when they worked without it.

A use-mode governance framework

Given all these findings, governing GenAI’s uneven performance requires asking a question that most law firms are not asking. Instead of determining whether GenAI is appropriate for a particular deliverable — such as a brief, a contract, or a board presentation — the governance question instead should be: What cognitive function is being delegated to GenAI at each step in the workflow?

My proposed framework, outlined below, organizes common GenAI uses into seven recurring modes following the sequence in which lawyers actually use GenAI to produce legal work product. Then, governance controls are calibrated to the risk profile of each mode.

Modes 1 and 2: Retrieval and organization

At the mechanical end of the cognitive spectrum are two distinct functions. In retrieval mode (Mode 1), a lawyer reviewing a merger agreement asks GenAI to identify every representation and warranty in the document. In organization mode (Mode 2), a litigator reviewing 50 depositions asks GenAI to construct a timeline from the testimony. The first locates material that already exists. The second arranges it into a usable structure. No new content is created in either case, and both uses are low-risk and should be actively encouraged, subject to modest verification controls. Firms that unduly restrict these use modes are leaving value on the table.

Mode 3: Summarization

Summarization (Mode 3) introduces selection risk. In this mode, GenAI chooses what to emphasize, include, and omit. Consider a lawyer preparing a board presentation on the results of an internal investigation. GenAI can condense dozens of witness interviews into key points and themes in minutes; however, a summary may focus on procedural detail while missing credibility issues that a lawyer would immediately recognize as material. The appropriate control is to mandate meaningful review by a lawyer with first-hand knowledge of the source material. A lawyer encountering the summary cold has no reliable way to evaluate what GenAI missed.

Mode 4: Candidate generation

Mode 4 is exploratory. A lawyer drafting a brief might ask GenAI to generate a list of potential arguments, propose alternative framings, or identify supporting authority. This candidate material expands options and accelerates iteration. The work product is not filing-ready and must be treated as provisional. GenAI can suggest, but a lawyer must decide.

The authority verification obligation at this stage deserves special emphasis. GenAI will identify cases, summarize holdings, and weave them into an argument structure. Thus, the output will read fluently and cite real-looking cases. However, a lawyer cannot assume the model has accurately characterized the holdings or context, and any authority cited in an external filing must be independently read and verified. GenAI can help find the cases, but a lawyer must read and apply them.

Mode 5: Editing and rewriting

In Mode 5, a lawyer asks GenAI to tighten a dense contract provision or restructure a wordy paragraph, risking, of course, unintended meaning change. An edit may read cleanly while subtly narrowing a representation, softening a covenant, or eliminating a carve-out. The revision risk is not hypothetical. The University of Minnesota study referenced above found that stronger performers produced worse work product when GenAI revised their independently produced memos. In this mode, a lawyer must confirm that the edit produced no shift in meaning and introduced no new factual assertions.

Mode 6: Critique and stress-testing

Mode 6 may be the most underutilized GenAI capability. Before filing a brief or presenting to regulators, a lawyer can ask GenAI to identify weaknesses in their argument. In this way, GenAI finds vulnerabilities before adversaries do; and unlike every other mode, the risk here runs in one direction. Lawyers who skip this step are missing one of GenAI’s core value propositions. Law firms’ governance frameworks should not merely permit it but actually require it in appropriate cases.

Mode 7: Evaluation and decision

The boundary against AI delegation becomes absolute when GenAI is asked to evaluate or decide. A lawyer advising a board on whether an event requires disclosure cannot delegate that determination to GenAI. A litigator assessing settlement value cannot outsource probability judgments because these are core expressions of professional responsibility. In this mode, GenAI may inform background analysis, but it may not substitute for lawyer judgment in making the call. This is a categorical prohibition — professional judgment cannot be delegated.

Going forward with GenAI

Law firm leaders who have moved their GenAI policy from restriction to mandate without governing the space between have not finished the job. Their lawyers are making consequential decisions about GenAI use every day without the guidance they need and deserve.

The use-mode framework presented above gives firm leadership a practical tool for filling that gap. It identifies the instances in which GenAI enhances legal work, where it introduces serious risk, and where professional judgment is non-negotiable. Firms that govern at that level will capture GenAI’s value; and those firms that do not will have policies that look serious but govern nothing important.

The views expressed in this article are solely those of the author in his individual capacity and do not represent the views, positions, or opinions of Foley & Lardner LLP, its partners or clients, or the University of Wisconsin Law School.

Key points:

• • • AI capability is jagged — Ethan Mollick’s frontier metaphor describes a coastline of strengths and weaknesses, in which a model that excels at contract analysis can fabricate a citation in the same conversation.

    • Human intelligence is jagged too — A century of psychology, from multiple intelligences to the Big Five, shows that each lawyer has their own coastline of strengths and weaknesses.

    • Person-AI fit is the next discipline — Firms that take this seriously will move from one-tool deployments to portfolios that match each lawyer to an AI partner whose jagged edges meet theirs.

Welcome back to The AI Law Professor. Last month, I examined how AI first drafts can blind us to other lines of reasoning and hijack our legal judgment. This month, I want to take up what determines whether an AI works for any given lawyer at all: Not which model is best, but which model is best for this lawyer, on this kind of work, at this point in their career

Professor and author gave us the metaphor that started this conversation — the jagged frontier of AI capability. Picture a coastline, irregular and unpredictable. On one side, the model is capable; on the other, it fails, sometimes catastrophically. The line itself does not run where you expect. Tasks that look hard turn out to be easy, and tasks that look easy turn out to be hard.

In terms of legal work, this means that a model that has just produced a useful contract analysis will confidently invent a citation. A model that has summarized a 90-page deposition with insight will fail at basic arithmetic. The capabilities of AI form a coastline, with bays and inlets and the occasional cliff. Mollick’s contribution was to give us a way to see this clearly. AI is not uniformly competent or uniformly incompetent — rather, it is jagged.

Humans are jagged too. Psychology has been telling us this for a century, although the message is uncomfortable enough that we keep flattening it back into a single number. The single-number version is IQ; yet the deeper issue with IQ is that it pretends intelligence is one-dimensional.

Developmental psychologist Howard Gardner’s , whatever its empirical limits, points us toward a more honest picture, one in which linguistic, logical-mathematical, spatial, musical, interpersonal, intrapersonal, and kinesthetic intelligences, are each largely independent. People are not equally strong across all these dimensions. So, it follows that a great trial lawyer and a great patent lawyer are drawing on different intelligences, and each could be lost in the other’s territory.

Human intelligence, like AI capability, is jagged, and each of us has an edge. The jaggedness is not a flaw to be smoothed; rather, it’s a feature of being a unique individual.

When two jagged edges meet

Place the two coastline maps — the human and the AI model — side by side. Press them together at random and they grind, with gaps where neither side fills the space and ridges where both claim the same territory. The lawyer’s strength overlaps with the AI model’s strength, so neither is amplified. The lawyer’s weakness overlaps with the model’s weakness, so neither is covered. The pair produces less than either party would produce alone.

However, align the same two surfaces with attention to their contours and something different happens. The peaks of one fit the valleys of the other. The lawyer’s weakness is met by the model’s strength; and the model’s weakness is met by the lawyer’s strength. The pair becomes more capable than either party alone.

A law firm that takes this seriously will not deploy a single AI tool across all of its lawyers and call the rollout complete. It will offer a portfolio of models and configurations and help each lawyer find the AI partner that works with their actual mind.

Every foundational model now ships with a model card, a document describing the model’s intended uses, training data, performance characteristics, and known limitations. The cards exist because models are not interchangeable. Read three of these cards side by side and the matching question becomes clear. A cautious generalist that hedges and flags uncertainty fits a lawyer who already holds strong views and wants a partner that will test them. A citation-anchored specialist that refuses to invent cases and stays grounded in retrieval fits a lawyer in heavily regulated practice areas in which errors are catastrophic.

The matchmaking discipline

Organizational psychology has worked on a version of this problem for 50 years under the . When a person’s strengths, values, and working style align with the demands and culture of their role, performance and well-being both rise. When they misalign, performance drops and burnout follows.

The same logic applies to person-AI fit. On the human side, cognitive style, domain expertise, personality profile, and the actual tasks performed in a typical week are key. On the AI side, behavior under different prompt styles, default tone, willingness to push back, hallucination patterns, and the shape of strengths and weaknesses across the practice areas in question may matter most. Yet, law firms are still treating AI procurement as a software decision rather than a partnership decision.

A law firm that takes this seriously will not deploy a single AI tool across all of its lawyers and call the rollout complete. It will offer a portfolio of models and configurations and help each lawyer find the AI partner that works with their actual mind. The first generation of legal AI has been dominated by the question of which model is best; however, the second generation will be dominated by a different question: Not which model, but which pairing works best. Not capability, but fit.

Those lawyers that flourish with AI will not necessarily be the most technical or the most enthusiastic users. Instead, they will be the ones that found, by luck or by design, an AI partner whose jagged edges meet theirs.

When two jagged intelligences fit well together, they can accomplish more than what either — human or AI — could do alone. Today, fit is the frontier.

Tom Martin is CEO & Founder of LawDroid, Adjunct Professor at Suffolk University Law School, and author of the forthcoming

Ěý

Key insights:

• • • Accountability, not adoption, is the central governance challengeĚý— With many institutions using AI a variety of tasks, informal “shadow AI” use is expanding without consistent oversight.

    • Justice systems now face a parallel governance problem — They must find a way to regulate AI while using AI inside the institutions that enforce rights, while allowing responsible innovation that improves efficiency and access to justice.

    • AI needs to be integrated into broader justice reformĚý— Without strong data governance and clear boundaries between AI assistance and legal judgment, courts risk automating inefficiency, deepening inequities, and undermining public trust.

Even as AI governance frameworks remain mired in ongoing debate, justice systems are moving ahead with implementation. Courts and dispute resolution institutions are integrating AI into their operations to more efficiently digitize records and automate workflows.

This introduces the very real challenge of parallel governance. We must now determine not only how AI should be regulated, but how it operates within the very institutions responsible for enforcing rights.

And this intersection is no longer theoretical: Does AI governance strengthen fairness, preserve independence, and expand access — or does it undermine their very foundations?

From experimentation to embedded use

Across jurisdictions, AI is often framed as an administrative tool that can handle basic tasks such as transcription, translation, case triage, and more, as well as providing analytics to identify delays or inefficiencies.

These applications respond to real constraints, such as overburdened courts, limited resources, and persistent backlogs. Similarly, dispute resolution platforms are integrating AI to guide users through processes and structure negotiations.

However, this formal adoption tells only part of the story. AI is also entering justice systems informally. Judges, clerks, and lawyers are independently using general-purpose tools in their daily work, often without guidance, oversight, or a clear grasp of the tools’ implications for security, confidentiality, and discoverability. As one expert observed: “Shadow AI is already happening.”

The absence of governance does not prevent AI use; and, in fact, it may encourage misuse. This shadow AI simply pushes AI usage into unstructured and unmonitored areas — the risk then becomes not adoption itself, but uneven adoption that evolves beyond institutional control.

It’s no longer a question that justice systems need to engage with AI; however, that engagement has be done deliberately and in a way that allows governance frameworks to keep pace without constraining beneficial use.

While it’s no longer a question that justice systems need to engage with AI, that engagement has be done deliberately and in a way that allows governance frameworks to keep pace without constraining beneficial use.

Automating inefficiency?

Efficiency is often the entry point for AI in justice systems; but efficiency alone is not reform. And misapplied efficiency can often lead to its direct opposite: a scramble to repair broken systems or to plug technology and personnel gaps.

Many current AI initiatives remain isolated pilots — layered onto existing processes rather than integrated into broader institutional strategy. Without addressing underlying structural constraints like fragmented data, inconsistent procedures, and uneven infrastructure, AI risks automating inefficiency rather than resolving it. And without strong data governance, infrastructure, and institutional alignment, even well-designed AI tools will underperform or produce unreliable outcomes.

That means that efforts to tightly control AI deployment without addressing these foundational issues risk focusing on symptoms rather than the system itself. AI should not function as a parallel modernization effort; rather, it must align with broader justice system reform.

Clearly, the most consequential questions arise when AI tools begin to shape legal reasoning or outcomes. And while there is broad agreement that AI can support judicial work without replacing independent human judgment, in practice, however, the boundary between assistance and influence is not always clear.

Even administrative tools can shape decisions. Summaries may omit nuance, or suggested language can influence framing. Over time, reliance on system outputs can create subtle forms of dependency. In fact, this dynamic is compounded by what has been described as the myth of verification — the assumption that human oversight alone is sufficient. In reality, time constraints, cognitive bias, and limited technical fluency can make meaningful review difficult. And automation bias affects even experienced decision-makers.

Overall, these boundaries require deliberate definition. Left on their own, AI tools and their outputs will be shaped implicitly through practice rather than through principled governance.

Design determines outcome

Institutional capacity will determine how these dynamics play out because digital maturity varies widely across jurisdictions. Some courts operate advanced platforms, while others remain largely paper based. In lower-resource environments, infrastructure may not support even basic digitization. In more advanced systems, adoption may outpace governance.

Yet, one consistent challenge among all jurisdictions is reliance on external vendors. Without internal expertise, institutions risk adopting tools that meet technical requirements but fall short of rule-of-law standards, particularly in transparency, accountability, and data governance.

Justice systems are not neutral environments for technology adoption — they are the operational core of the rule of law.

Addressing this gap requires more than a procurement issue. It requires institutional literacy. Judges and administrators need a working understanding of how AI systems function, where risks arise, and how to evaluate them. Training efforts are underway, but scaling this capacity will take time. In the interim, governance gaps will persist and attempts to compensate for these gaps through overly rigid restrictions may limit adoption but do little to build the institutional capability required for effective oversight.

From adoption to accountability

Clearly, AI will not improve justice systems by default; rather its impact will be determined by institutional design, which includes clear boundaries on use, transparency around deployment, safeguards to protect independence, and mechanisms for oversight and accountability. It also requires alignment with broader justice system goals of efficiency, fairness, and accessibility.

Yet, justice systems are not neutral environments for technology adoption. They are the operational core of the rule of law. Their legitimacy depends on trust, which in turn requires accountability.

This makes the path forward not purely a technical one. It requires institutional self-assessment, alignment with human rights frameworks, and collaboration across policymakers, courts, technologists, and the public. The measure of success will not be the sophistication of the tools deployed, but whether they strengthen the system’s core functions of impartiality, accessibility, and trust.

AI tools can support those goals, of course, but only if they are designed into justice systems from the outset.

You can find other installments ofĚýour Scaling Justice blog series here

Key insights:

• • • Consensus among AI systems is not the same as correctness — Agreement between AI models often signals shared blind spots, not truth; and AI errors can be highly correlated across instances and even across model families.

    • Productive disagreement must be explicitly designed into AI advisors — Multi‑agent AI systems are most effective when they are intentionally built to preserve meaningful disagreement, not just to synthesize a unified response.

    • The future of AI advisory mirrors long‑standing human decision-making — Modern multi‑agent AI design has a long historical lineage; yet, across all examples, the same principle holds: The best decision systems are engineered for internal conflict.

In this new two‑part blog series, we explore why AI works best as an executive advisor not by delivering consensus answers, but by being intentionally designed to identify, preserve, and productively leverage disagreement. In the first part, we saw why a single AI advisor is structurally vulnerable; now, in this concluding part, we look at what happens when you design disagreement on purpose.

The academic evidence for multi-agent AI systems has been building rapidly, and the most important findings aren’t about the power of agreement. They’re about the danger of it.

In February, , a product that sends every query simultaneously to three frontier AI models (Claude, GPT, and Gemini) then uses a fourth chair model to synthesize a unified answer. The product’s value proposition isn’t that three models produce a better answer than one; rather, it’s that divergence between models is treated as a signal. When models converge, that indicates confidence, but when they diverge, that indicates the user should slow down.

Studies have borne this out. Multi-agent debate compared to single-model generation, and researchers at the University of Göttingen found that , with their voting protocols outperforming other decision structures. However, potentially the most important finding cuts against the hype. In a 2026 paper, , the authors demonstrated that AI model errors are highly correlated both within and across model families. When three instances of the same model agree, it doesn’t mean they’re right, rather it means they may share the same blind spots. Aggregation increases consensus faster than it increases truth.

The future of AI-assisted executive decision-making may look less like a single brilliant oracle and more like a room full of advisors that may often disagree because that’s how the best decisions have always been made.

This finding cuts both ways for practitioners like ¶¶ŇőłÉÄę enterprise architect Zafar Khan and his two AI advisors, Adrian and Elara, that were built on the same underlying model but differentiated by their analytical frameworks rather than their architecture. The divergence they produce is real and visible. For example, the analysis the two AI advisors did on a deal undertaken by Eaton Corp., in particular generated genuinely different conclusions because the two advisors were oriented towards different priorities.

Yet, research suggests that same-model divergence, while effective, has a ceiling. Prompt-driven personas can ask different questions, but they share the same training, the same blind spots, and the same failure modes. Khan is candid about this, noting that his current system is in the “very early” stages and is not a finished product. The value right now, he says, isn’t that Adrian and Elara are equivalent to truly independent minds, it’s that even a first-generation version of structured disagreement can identify insights that a single advisor would miss. It’s a large stride rather than an arrival at the ultimate destination.

The future of AI advisory is in the past

The principle behind this diverging analysis concept isn’t new. Indeed, it might be one of the oldest ideas in institutional design, rediscovered independently by many institutions that had to make decisions under uncertainty. Socrates built a philosophical method around cross-examination; Pope Sixtus V formalized opposition by creating the Devil’s Advocate in 1587; and the RAND Corporation operationalized it during the Cold War with the Delphi Method, using structured anonymous iteration to prevent groupthink.

The through-line across two millennia is simply that the best decision-making systems don’t minimize disagreement, rather, they engineer it.

Today, the developer community now uses production-grade code review tools to assign architecture, security, and functionality analysis to separate agents, using majority voting for routine decisions and unanimous consent for irreversible ones. And what Khan has built and what Perplexity, Microsoft’s Agent Framework, and a growing ecosystem of multi-agent tools are now pursuing, are the latest iterations of the simple concept: Internal conflict is not a system failure, it is a design requirement.

The question is no longer “whether”

Khan’s vision for what should sit at the decision table is specific — five AI advisors spanning technology, finance, regulation, workforce, and geopolitical risk. Each applies its own analytical framework, with the human executive responsible for integration and final judgment. The guardrails are three: i) transparency about what data the system uses; ii) verifiability that sources are legitimate; and iii) human accountability at every decision point.

“The race towards AGI [artificial general intelligence] is moving faster,” Khan acknowledges, adding that the human needs to be in the loop in order to bring AI to work in a governance fashion and an ethical way.

“I want to show the interaction between human and AI advisor, how they’re thinking through the problem together,” he explains. “Where the human judgment covers the analysis and where it diverges.” In other words, when the AI advisors agree, that’s your green light. When they diverge, that’s the conversation your board should be having.

The future of AI-assisted executive decision-making may look less like a single brilliant oracle and more like a room full of advisors that may often disagree because that’s how the best decisions have always been made. The technology to build that room now exists; however, the question is whether today’s leaders have the discipline to listen when the room argues back.

For more on AI transformation in the professional services market, you can download the Thomson Reuters Institute’sĚý2026 AI in Professional Services Report

Key takeaways:

• • • Prompt engineering pays off — Tax professionals who master clear, contextualized AI instructions see immediate gains in output quality and speed.

    • AI doesn’t replace professional responsibility — Every output that carries your name requires your verification and your judgment.

    • Link learning to a real problem — The most effective way to build needed skills is to focus on your current workflow, not to chase every new tool as it emerges.

For tax professionals, technical excellence used to be enough. Know the code, understand the cases, apply the rules correctly — that was the job, and it was sufficient. It isn’t anymore. Not because the technical knowledge matters less, but because the professionals competing for the same work increasingly bring other talents to the table, such as the ability to do in an hour what used to take a day; to provide insights from data that would have taken a week to compile manually; and to deliver polished, well-reasoned analysis at a pace that wasn’t possible five years ago.

This rarified capability doesn’t come from intelligence or experience alone; rather, it comes from skills — specific, learnable, practical skills.

The data bears this out. Improving efficiency through technology has been the top strategic priority for firms for three consecutive years, with 44% of firm leaders citing it as their primary focus, according to the Thomson Reuters Institute’s . Indeed, 47% of tax professionals surveyed said investing in AI should now be a top priority — and yet, 18% of firms still use no automation at all.

This gap between intention and capability is real, and it sits squarely with the individual tax professional.

The skills most needed by today’s tax professionals

To help close this gap and improve tax professionals’ overall work value, there are several specific skills that demand attention, including:

Prompt engineering: The skill nobody takes seriously until they see what it does

The name doesn’t help — but set that aside, because the underlying skill is straightforward: giving your AI tools clear, precise, well-contextualized instructions that produce outputs that are worth using.

Most people start badly when approaching a blank AI screen. They type something vague, get something generic, and conclude the tool isn’t useful. That conclusion is wrong, because it was the instructions given, the prompt, that was the problem. Specify the entity type, jurisdiction, tax year, audience, and format. Then tell the tool what you need and why. The difference in output quality is not marginal.

Of course, it’s important to remember that AI will tell you things that are wrong with complete confidence. It will cite an amended provision, apply a rule from the wrong jurisdiction, or construct a plausible analysis on a flawed premise — all without flagging any of it. The professional responsibility to catch it remains entirely upon the user. That’s not a flaw in the tool; it’s a reminder that expertise isn’t being replaced here — it’s being put to better use.

Data literacy: The capability gap most tax professionals don’t know they have

Tax work is data work. Today, what has changed is the expectations around the volume and complexity that professionals are now required to handle, interpret, and present, often with fewer resources than a decade ago.

Advanced spreadsheet proficiency is the starting point, and the emphasis on advanced is deliberate. The features that most professionals have never explored are precisely the ones that separate those who spend three hours processing data from those who spend 20 minutes. The ability to build visual dashboards that communicate tax data clearly — effective tax rates, provision variances, deferred movements, and more — is increasingly an expectation in corporate environments rather than a differentiator. For those professionals who handle large datasets or complex scenario modeling, even a foundational understanding of represents a significant capability uplift.

The Tax Professionals Report found that 57% of firm leaders cited getting better use out of existing technology as their top investment priority — more than those planning to buy new systems. The problem, in other words, isn’t the tools; it’s having the skills and the understanding to use them.

Workflow automation: Reclaiming time from work that shouldn’t exist

Look at any tax workflow closely and you’ll find steps that are repetitive, rule-based, and time-consuming — not because they require a tax professional’s skilled judgment, but because nobody has stopped to ask whether these routine tasks could be done differently.

Again, the harder part of improving your skill set as a tax professional isn’t learning the tools; rather, it’s developing the habit of process analysis, a way of thinking that will allow you (among other things) to distinguish between steps that require genuine expertise and steps that are simply consuming time.

AI judgment: Knowing what to trust and what to verify

This is the skill that determines whether AI makes you more effective or creates problems you didn’t anticipate. This means validating outputs against primary sources before they reach a client. It means recognizing that AI reflects training data that may be outdated or jurisdiction-specific in ways that aren’t readily apparent in the output. And it means knowing when a task is too nuanced or too high stakes for AI to handle reliably.

Professional responsibility does not transfer to the tool itself. If an AI-generated analysis carries your name, it is your analysis.

Communicating and staying current

As routine tax compliance work becomes more automated, the premium on communication rises sharply. The Tax Professionals Report found that three-quarters of clients now strongly desire advisory services beyond tax preparation from their outside tax professional — yet most tax firms still derive their greatest profits from simple tax return preparation.

Those professionals who can close that gap are those who can translate technical work into clear, confident guidance that their clients can act on.

Going forward, the tools will keep changing. Identify the problem in your current workflow that costs the most time, find the skill that addresses it, and build from there. The professionals who will define the next decade will combine this deep technical knowledge with the ability to work faster, more clearly, and more adaptively than those who came before them. That combination is not yet common, but it’s also not out of reach.

For more on how tax professionals are navigating technological change, visit the or download the full 2025 State of Tax Professionals Report

Key takeaways:

• • • Strategic collaboration is becoming a defining strength for the region — Latin American organizations are realizing that progress in AI accelerates when they combine forces by linking industry expertise, academic talent, and public‑sector support.

    • AI initiatives rooted in real local challenges are gaining global relevance — By developing solutions grounded in the region’s own structural needs, whether in infrastructure, finance, agriculture, education, or mobility, many LatAm firms are producing technologies that are both highly impactful and naturally scalable.

    • Demonstrating clear outcomes is becoming fundamental — Organizations that show concrete operational improvements, measurable efficiencies, or stronger customer outcomes are strengthening their position with investors and partners.

In recent years, Latin America has experienced significant growth in investments related to AI, accounting for . This is strikingly low given that the region makes up around 6.6% of global GDP, highlighting the region’s opportunities to scale AI initiatives even further. Although there are notable differences among countries, Mexico and Brazil — the two largest LatAm economies — stand out for their volume of AI projects and funding, followed by other nations such as Chile, Colombia, and Argentina.

By recognizing the region’s strengths — which include cost-effective operations, access to data, clean energy, and public support — the region’s businesses can better position themselves and design strategies to draw in international investors that may be increasingly seeking promising locations for AI development.

Lessons from LatAm’s AI success stories

Latin America has produced remarkable AI success stories that can serve as models to build confidence among investors. These cases — involving companies that attracted substantial investment and achieved growth — demonstrate valuable best practices that range from technological innovation to working with governments and corporations. Some of these best practices include:

Building strategic alliances

The journey of innovation rarely unfolds in isolation. At times, the presence of large, established companies, whether local industry leaders or multinationals, has served as a catalyst for AI projects. The experience of that specializes in AI-powered agricultural irrigation, proves it. Now, Kilimo is partnering with EdgeConneX, a data center company based in the United States, on a community .

Academia, too, can be woven into this narrative. Collaborations with research centers or universities offer scientific credibility and connect ventures with emerging talent. In Mexico, AI startups often originate within university settings — such as computer vision projects from the National Autonomous University of Mexico (UNAM), for instance — and maintain agreements that sustain ongoing innovation and technical progress even with modest resources. And academic validations, whether in published papers or conference accolades, tend to resonate with foreign investors. Indeed, the emergence of this ecosystem that features early corporate clients and academic mentors frequently lends a distinctive appeal for those seeking investment.

Focusing on local problems with global impact

Within Latin America, certain issues prove especially relevant in situations in which AI solutions intersect with sectors renowned for regional strengths, such as fintech and financial inclusion, agrotech optimizing agriculture, and foodtech drawing on local ingredients. The experience of Chilean food startup NotCo — in which and subsequently exported — suggests how innovations rooted in local context may generate broader attention.

By addressing needs in urban transport, education, mining and related areas, local LatAm companies can provide access to homegrown data and users, which can further refine technology and open pathways for investors into similar emerging markets. When AI solutions respond to genuine pain points rather than mere novelty, momentum often builds more quickly, and the model finds validation among that evaluate investments.

Showing results and AI ROI early on

Questions linger for many executives . Evidence of clear metrics like cost savings, sales growth, or error reduction can prove persuasive, especially when complemented by success stories from local clients.

Recent studies show that companies ; and such figures tend to reassure those considering investment by illustrating tangible improvements. Testimonials or independent validations, such as a university study, can further illuminate achievements.

The act of quantifying impact — whether in efficiency, revenue, or other relevant KPIs — has a way of transforming perceptions from uncertainty toward clarity.

Leveraging government incentives and collaborations

Many Latin American nations have put forth support programs for AI and tech projects, such as non-repayable funds, soft loans, and tax benefits for innovation illustrated in , , , or the .

Public financing, when present, often acts as a stamp of validation for private investors. For example, this trust extended to Brazilian startups receiving Finep support for AI health projects, which in turn can shift perceptions for foreign ventures capitals. Engagement in government pilots, such as smart city initiatives or solutions for ministries, provides valuable exposure. In such contexts, public-private partnerships and incentives seem to act as quiet levers for growth and legitimacy.

Seeking smart and diversified financing

Financial strategies in Latin America have been shaped by the interplay of local and foreign capital. Local funds often bring insights and patience, while foreign funds may offer larger investments and global scaling experience. Ownership dilution sometimes accompanies the arrival of strategic investors, whose networks can prove invaluable, such as . Programs like 500 Startups, Y Combinator, MassChallenge, and international competitions have ushered LatAm AI startups such as Heru, Rappi, Bitso, and Clip into new rounds of capital following increased exposure.

Efficiency in capital management, which can be demonstrated with lean burn rates and milestone achievement with limited resources, signals an ability to execute within the realities of LatAm, which may enhance the appeal for future investments. The cultivation of relationships and responsible stewardship of capital frequently matters as much as the funds themselves, suggesting that the value of mentorship, contacts, and reputation is often intertwined with deepening financial support.

Unlocking AI Investment

By applying these principles, Latin American companies have achieved a better position to attract AI investments to their projects and help position the region as a viable destination for technology capital. These recent experiences show that when a LatAm company combines innovation, talent, and strategy — while communicating its story well — it can win over global and local investors alike. Each of the best practices noted above is based on real lessons: international alliances (NotCo with US funds), leveraging incentives (Brazilian companies funded by Finep), talent formation (Santander and Microsoft programs), focus on ROI (successful use cases that convince boards), and more.

Latin America has challenges but also unique advantages. Companies that manage to navigate this environment intelligently will increase their chances of securing the financing needed to innovate and grow. By doing so, they will contribute to a virtuous circle in which each new success attracts more investment to the region and opens doors for the next generation of LatAm AI ventures.

You can find more about the challenges and opportunities in the Latin American region here

Key takeaways:

• • • AI threatens traditional lawyer development — As AI automates entry-level legal tasks like research and writing that historically has honed legal judgment skills, the profession faces a crisis in how new lawyers will develop such judgment abilities.

    • The profession can’t agree on what constitutes “legal judgment” — Unlike other professions, there is no agreed-upon definition of legal judgment or clear standards for when AI should be used.

    • Implementation requires unprecedented coordination and funding — A legal education fund as a proposed solution would require a small percentage of legal services revenue and coordinated action across law schools, legal employers, and state regulators.

This is the second of a two-part blog series that looks at how lawyer training needs to evolve in the age of AI. The first part of this series looked at how lawyers can keep their skills relevant amid AI utilization.

The key skills that comprise legal judgment have received mixed reviews, according to a recent white paper from the Thomson Reuters Institute that advocated for cultivating practice-ready lawyers. The white paper was based on feedback from thousands of experienced lawyers, judges, and law students and raises questions about how legal judgment forms when AI assistance is used for task completion.

notes that calls for “… to accelerate the development of legal judgment early in lawyers’ careers.”

The challenge is that each part of the profession — law schools, employers, state supreme courts (as regulators) — have distinctly separate responsibilities. That means, that in the age of AI, coordination across the entire legal profession is needed, especially as AI reduces the availability of traditional first jobs.

Furlong points out that there is no consensus for what legal judgment is or any agreed upon standards for in what instances AI should be used in legal. To bring clarity to these issues, the white paper proposed a profession-wide model that integrates three critical elements: i) work-based learning that’s modeled on medical residencies; ii) micro-skill decomposition of legal judgment; and iii) AI-as-thinking-partner throughout pedagogy.

Three pillars for an AI-era lawyer formation system

Not surprisingly, overreliance on AI can erode critical analysis and solid legal judgment skills. Addressing these concerns requires a comprehensive reimagining of how lawyers are educated and trained. One solution lies in three interconnected pillars that together form a cohesive system for developing legal judgment in an AI-integrated world.

Pillar 1: Integrate work experience into legal education

Core skills such as legal research, writing, and document review help develop legal judgment; yet these skills could collapse once AI assumes such tasks. The Brookings Institution recently proposed to preserve entry-level professional development in an AI era. This parallels the TRI white paper’s calls for mandatory supervised postgraduate practice as a key part of legal licensure.

While implementing a full residency model presents challenges, several law schools have already pioneered approaches that demonstrate the viability of work-integrated legal education that, if scaled appropriately, could improve new lawyer practice and judgment skills. For example, Northeastern Law School guarantees all students nearly before graduation through four quarter-length legal positions. The program integrates supervised practice into the curriculum so graduates can gain substantial hands-on experience alongside their classroom instruction.

Also, program offers an alternative pathway to bar admission through practice-based assessment rather than the traditional bar exam. The program demonstrates that competency can be evaluated through supervised experiential learning.

Pillar 2: Decompose legal judgment into teachable micro-skills

The legal profession needs to come to a common definition of legal judgment and develop its components to teach the concept effectively. “We can’t teach what we can’t describe,” Furlong says. To develop legal judgment, the profession must define its components, including:

• • • Pattern recognition — The ability to identify when different fact patterns are related to similar legal frameworks and distinguish when superficially similar cases are legally distinct.
    • Strategic calibration and proportionality — This means understanding what level of effort, precision, and risk each matter requires and matching responses to the stakes involved.
    • Reasoning through uncertainty — This is the capacity to make defensible decisions and provide sound counsel even when the law is ambiguous, unsettled, or silent on an issue.
    • Source evaluation and authority weighting — This includes knowing which legal authorities are most suitable and being able to assess their persuasive value.
    • Ethical judgment under pressure — This means spotting conflicts, confidentiality issues, and duty-of-candor moments while maintaining competence and knowing when to escalate beyond expertise.

Breaking down legal judgment into these discrete components makes it possible to design targeted teaching interventions. For example, , former law professor and executive director of , suggests we back into AI-assisted workflows by requiring a short verification log (detailing sources checked, changes made, and why); running attack-the-draft drills (find missing authority, weak inferences, and jurisdictional mismatch); and preserving slow work as formative work (citation chaining, updating, and adversarial research memos).

With judgment skills clearly defined and work experience integrated into training, the profession must then tackle how AI itself should be incorporated into lawyer development.

Pillar 3: AI-as-thinking-partner throughout a lawyer’s career

Warnings that are mounting. The legal profession must provide clear standards for in what instances and how AI should be used, with training in verification and judgment skills. Overreliance on AI could compromise lawyers’ capacity to fulfill their fiduciary duties to clients.

A phased approach in the introduction of AI in legal work helps protect critical thinking while building AI competency. For example, in Year 1, law students could complete core legal reasoning exercises without AI assistance in order to better develop their analytical muscles. In Year 2, students use AI as a research assistant with mandatory verification protocols that teach students to check outputs against authoritative sources. Finally, in Year 3, residencies can immerse students in real-world AI workflows under proper supervision and while providing feedback.

These three pillars form a coherent vision for lawyer formation in the AI era. However, the most well-designed system faces the obstacle of funding.

The challenge of who pays

Perhaps the most difficult part of any overhaul is the cost. The medical residency model works because — up to $15 billion-plus annually — for teaching young medical students to be doctors. Legal education has no equivalent. Without addressing funding, however, even the best reforms will fail.

One idea is to establish a legal education fund that’s supported by an assessment of a small percentage of the legal industry’s gross legal services revenue (while exempting solo practitioners and firms with less than $500,000 in annual revenue). These funds could be used to subsidize thousands of supervised residency placements, fund law school curriculum development, support bar exam alternative assessments, and provide employer training and supervision stipends.

The challenge is that each part of the profession — law schools, employers, state supreme courts — have distinctly separate responsibilities, and that means coordination across the entire legal profession is needed.

This proposal, of course, would require unprecedented coordination and financial commitment from the legal profession. Skeptics might argue that market forces can solve this problem, or that firms will simply create new training pathways, or that AI will prove less disruptive than feared. However, waiting for market forces risks a lost generation of lawyers. The medical profession already when the medical industry’s voluntary reform failed. Only later did coordinated regulatory intervention produce the consistent quality standards the medical industry sees now.

What is clear is that inaction is resulting in degradation of lawyering skills. “Maybe… we need catastrophic external intervention to bring about the wholesale changes we can’t manage from the inside,” Furlong suggests.

However, the question is whether the legal profession will wait for a crisis to force change or act proactively to make the needed changes now, before the crisis hits.

You can learn more about the impact of AI on professional services organizations at TRI’s upcoming 2026 Future of AI & Technology Forum here

Key highlights:

• • • Lawyers excel at semantic legal work while AI excels in syntactic tasks — Syntactic work (document generation, pattern recognition) is where AI excels, but semantic work involving exercising independent judgment, reflecting on consequences, and fulfilling fiduciary duties remains uniquely human.

    • Fiduciary duty as the core of legal relevance — What distinguishes lawyers isn’t justĚýwhatthey do, butĚýhow and whyĚýthey do it. The fiduciary relationship demands human understanding of context, balances competing interests, recognizes unstated concerns, and exercises discretion.

    • 5 hours to deepen or diminish — The five hours lawyers are expected to gain each week by using AI can either accelerate professional obsolescence or deepen lawyers’ relevance, depending on what they do with it.

This is the first of a two-part blog series that looks at how lawyers can keep their skills relevant in the age of AI

Lawyers expect to gain a full five hours per week of worktime due to the efficiency derived from AI use, according to the ¶¶ŇőłÉÄę 2025 Future of Professionals Report. Yet the fear of job loss among lawyers is rising, as those viewing AI as a threat or somewhat of a threat grew from to almost two-thirds (65%) of those surveyed, according to the Thomson Reuters Institute’s 2026 AI in Professional Services Report.

Many in the legal profession are asking how lawyers are uniquely valuable at a time when machines can process legal information faster and cheaper. The answer lies in understanding the difference between what AI does in processing legal information and what humans do in exercising legal judgment, says , Founding Director of the .

Defining 2 levels of legal work

Understanding what makes lawyers particularlyĚýmeaningfulĚýin this current AI moment requires distinguishing between two different levels of legal work in an environment in which AI-enabled information systems are compressing humanity and legal judgment into data points and draining away the storytelling and moral nuance that ground both. According to Lee, these different levels involve the syntactic and the semantic:

• • • Syntactic — Lawyers process information, generate documents, and recognize patterns at the syntactic level, meaning those tasks in which AI excels and delivers promised efficiency gains. “The danger is that we will use this efficiency merely to generate more syntactic volume,” Lee explains, adding that this will result in faster processing of more documents at greater speeds. “If we do that, we will have automated ourselves out of a profession.”
    • Semantic — The semantic aspect of lawyering highlights the irreducible skills of the legal practice, which include exercising independent legal judgment, reflecting on consequences, demonstrating care for clients, and fulfilling fiduciary duties.

This distinction between the semantic level is inherent within the practice of law definition, Lee says, pointing out that many jurisdictions distinguish between “providing legal information” (not practicing law) and “exercising independent legal judgment” (the essence of legal practice).

He also rightly contends that the existential risk facing lawyers is not in AI completing legal tasks, but rather the temptation to reduce lawyers’ role to verifying machine output and processing legal information. Conflating these two concepts is a challenge for the legal profession and requires increasing the appreciation for the craft of legal reasoning and judgment.

Making this more difficult is that the current information age complicates this picture by challenging society’s assumptions about reality, consciousness, and the moral meaning of human life — all at an exponential rate, Lee says. Similarly, AI and information systems threaten to reduce everything, including human beings and law itself, to processable data by stripping away the narratives and meanings that define humanity, he adds.

Semantic qualities of legal judgment

The question of what makes lawyers especially relevant in the AI era is mainly answered in how and why they do what they do, rather than in what they do. For example, Lee points to skills around executing their fiduciary duty and ensuring legitimacy and meaning as key characteristics of lawyers’ semantic qualities.

Fiduciary duty — When a client seeks legal counsel, it’s legal judgment — not information processing — that the client wants. Lawyers, as part of their fiduciary duty to their clients, demonstrate human and legal understanding of the unique context of each case and the consequences of various legal paths forward. This bond of trust between attorney and client demands reflection, consideration, care, and proper purpose.

The fiduciary duty of the lawyer to the client requires balancing competing interests, recognizing unstated concerns, and exercising discretion in ways that honor both the letter and spirit of the law. At the heart of this balance is legal reasoning and professional judgment, which often involves navigating the critical gap between legal rules as written and their meaningful application to human circumstances.

Legitimacy and meaning — Beyond the fiduciary of care exercised in individual client relationships, lawyers serve a broader purpose in their role to safeguard law’s connection to the narratives of justice and human dignity that legitimize its authority. Indeed, lawyers maintain the connection between law and its humanistic foundations, so that the narratives that give legal authority its legitimacy depend on this connection. “The artwork that one associates with the law (in law schools and courtrooms) connects actions and legal judgment of attorneys to the mythic meaning of justice, equality, and the rule of law,” Lee explains.

How to deepen appreciation for the special relevance of lawyers

The five hours that lawyers said they expect to gain each week through AI-driven efficiency represents a choice point for the profession. These hours can either accelerate lawyers’ obsolescence or deepen their relevance. To ensure the latter, Lee advises lawyers and legal institutions to examine ways to put those hours to good use by, for example:

Collaborating on apprenticeships — Bar associations, practicing lawyers, legal service providers, and law schools should consider apprenticeship models that teach professional norms and values through mentorship that allow law students to learn the craft of legal reasoning through guided practice.

Recommitting more fully to legal service — Law firms and in-house counsel must reclaim humanistic awareness as central to their professional identity. The efficiency gains from AI should be reinvested into semantic work, which include counseling clients, exercising moral judgment, and fulfilling fiduciary duties with greater care and reflection.

Improving legal education — Law schools must return to the humanistic formation of lawyers, echoing the vision of the pre-2007 , before economic pressures reduced legal education to producing commercially exploitable graduates. In addition, AI ethics must be integrated systemically across the curriculum into doctrinal courses rather than being confined to elective courses.

Looking ahead

The five hours gained through AI represent a defining choice for the legal profession. The special relevance of lawyers in the AI age lies precisely in the human components and semantics aspects of lawyering.

In the concluding part of this blog series, we look at how the legal profession needs to rethink how it trains lawyers in order to prevent AI from eroding legal judgment skills