How agentic AI exposed the absence of deliberate judgment architecture in professional services—and why every organization should pay attention.
The work nobody talks about fondly—but everyone learned from
If you spent your early career at a professional services firm, you remember the work. Reconciling data at 11pm. Rebuilding slide decks because a partner changed one assumption. Running quality checks on deliverables you barely understood.
Nobody called it education. But something happened inside that repetition. You started noticing when numbers didn't feel right before you could articulate why. You learned to read a partner's reasoning by reverse-engineering their edits. Through a hundred small corrections, you discovered what "good" looked like.
The firms never designed this as a learning system. Junior staff did the heavy lifting, and somewhere along the way, they became professionals who could exercise judgment. The mechanism was invisible, and because it worked, nobody examined it. Until now.
What AI agents changed—and what they surfaced
In 2025, each of the Big 4—Deloitte, PwC, EY, and KPMG—launched agentic AI platforms capable of processing in seconds what junior professionals once spent days producing: document drafts, data reconciliations, quality checks, summaries, and preliminary recommendations.
A February 2026 Business Insider report captured the uncomfortable question this raises: how will junior professionals learn the work if they no longer do the work?
The answers from firm leaders were notably candid. KPMG's head of AI talent acknowledged that developing core skills alongside agentic tools remains an open challenge—and admitted he doesn't fully know the answer. The CEO of the American Accounting Association called it the profession's biggest unsolved problem. At Davos 2026, reporters observed that many senior leaders had given surprisingly little thought to how workforce development needs to evolve.
This honesty is worth respecting. But it invites a deeper look at what the question actually is. On the surface, it appears to be about AI removing learning opportunities. Read more carefully, and a different shape emerges. The real issue isn't that AI took something away. It's that AI made visible something that was never deliberately built.
The firms' responses: well-intentioned and genuinely useful
Before going further, the Big 4 are not ignoring the problem.
PwC has articulated an approach centered on teaching the "why" behind tasks. New hires complete an intensive AI onboarding program pairing technical AI skills with human skills—critical thinking, structured reasoning, better questioning. KPMG's leadership has argued that learning patterns must shift: junior staff should analyze agent outputs and understand how conclusions were reached. EY has suggested that AI support may actually accelerate development by giving junior professionals earlier exposure to strategic client conversations. Deloitte has emphasized ongoing investment in upskilling.
Each response has its own logic, and none is wrong. Teaching "why" builds cognitive foundations. Output verification develops critical thinking. Early strategic exposure broadens perspective. These are good-faith investments in helping junior professionals make better judgments.
But notice what they share in common. Every initiative is designed to improve the quality of judgment. That practice has a name: Decision Support (improving the inputs, analysis, and verification available to a decision-maker—distinct from the technology category of "decision support systems"). The question is whether Decision Support, however sophisticated, is sufficient when something more fundamental remains undefined.
Why "teach the why" is right—but incomplete
Consider what it means to "teach the why." In practice, this involves explaining the reasoning behind audit procedures, the logic of compliance frameworks, the purpose behind each engagement step. Valuable knowledge transfer. But knowledge of "why" is different from the experience of exercising judgment under ambiguity.
Understanding why a reconciliation matters is not the same as finding a discrepancy at midnight and deciding whether it's a rounding error or something to escalate. The first is education. The second is a judgment encounter—a moment where you must decide, with incomplete information, what to do next.
KPMG's proposal that junior staff analyze agent outputs is sensible. But analyzing an output requires a basis for comparison—the ability to imagine how the output could be wrong. That capacity is built through having done the underlying work, having seen the ways data breaks and assumptions distort conclusions.
EY's suggestion of earlier strategic access is valuable, but strategic judgment rests on a substrate of operational micro-judgments. Without that foundation, a junior professional in a strategic meeting may be present without being a participant.
Each approach improves the inputs available to a decision-maker. None addresses a prior question: where is judgment supposed to happen, and who is supposed to be exercising it?
The missing layer: where judgment is supposed to live
Here is the structural issue the Big 4 conversation has not yet articulated.
In the traditional apprenticeship model, junior professionals encountered judgment opportunities incidentally. Nobody planned which reconciliation would surface a meaningful discrepancy. These encounters were scattered within the volume of repetitive work, occurring frequently enough that most professionals accumulated a working base of judgment over time.
This was not a designed system. It was a fortunate accident. Repetitive work happened to contain small decision points—moments where someone had to assess, interpret, escalate, or let pass.
AI agents absorbed the repetitive work and the incidental judgment encounters embedded within it. What disappeared was not just task volume but the invisible scaffolding on which professional development had always depended—without anyone realizing it.
The problem isn't that AI disrupted a functioning development system. It's that AI revealed there was no system—just conditions that happened to produce growth.
Decision Design: designing the "where" of judgment
To make sense of this, it helps to distinguish three layers that are often conflated.
The first is decision-making itself. People make decisions constantly, in every role, at every level. This is not in question.
The second is Decision Support—the practice of improving the inputs, analysis, and frameworks available to decision-makers. Training programs, AI-generated recommendations, mentoring, the "teach the why" approach—all operate at this layer. Decision Support makes individual judgments better-informed. It is valuable and necessary.
The third layer is different in kind. It asks: where do decisions live in this organization? Who owns which judgments? What is the scope of responsibility attached to each decision point?
At Insynergy, we call this Decision Design™—the practice of deliberately defining where decisions live, who owns them, and the scope of responsibility attached to each.
Decision Design does not replace Decision Support. It provides the architecture that support needs in order to land. You can invest heavily in helping people make better judgments, but if you haven't defined where those judgments are supposed to occur—whose domain they belong to, what authority they carry—then the support has no address.
This is the gap the Big 4 are experiencing. They are investing seriously in support. But the location of judgment—especially for junior professionals—has never been explicitly designed. It was an emergent property of a work environment that no longer exists.
Decision Boundary: when the lines stay on paper but go hollow inside
A Decision Boundary™ defines which judgments belong to which role—the perimeter within which a professional is expected to exercise assessment and bear responsibility for the outcome.
In the pre-AI model, junior professionals had small but real decision boundaries. Is this data accurate? Does this structure hold together? Should this discrepancy be flagged? Narrow questions, but with genuine consequence. A missed anomaly could cascade. A poor deliverable would be caught and corrected. The boundaries were small, but substantive.
As professionals matured, boundaries expanded. Mid-level staff owned project-level judgments. Senior leaders owned strategic and risk decisions. The progression was intuitive: your boundary grew as your judgment proved reliable within it.
Now consider what happens when AI agents take over the tasks that defined junior boundaries. The reconciliation, the anomaly detection, the structural proposal—all handled by the agent. What remains for the junior professional is "review."
On paper, the junior still has a decision boundary. But to verify something meaningfully, you need to imagine how it could be wrong. That capacity is built through doing the work yourself. Without that substrate, review becomes ritual. The boundary remains formally assigned, but its interior is hollow. The junior is "making a decision" in the organizational sense—they have signed off—but the cognitive act of judgment may not have occurred.
AI didn't remove junior decision boundaries. It left the boundaries formally intact while emptying them of substance.
Judgment compression: when review becomes ceremony
Judgment compression is the collapse of a multi-step decision process into a single act of approval—where the sign-off survives but the reasoning behind it does not. The hollowing of decision boundaries produces this phenomenon, and it is worth examining closely.
Consider a traditional judgment sequence: examine raw data, identify a discrepancy, hypothesize its cause, assess its significance, decide whether to escalate. Each step involves a distinct cognitive act.
When an AI agent handles the early steps and presents a conclusion, the reviewer is left with only: approve or question. But questioning effectively depends on having internalized the earlier steps. Without that, the reviewer lacks the reference frame to evaluate whether the conclusion is sound.
The insidious quality of compression is that it feels like judgment. The reviewer looks at the output, considers it, approves it. Subjectively, this registers as a decision made. But the intermediate reasoning that would make it substantive has been bypassed.
PwC's "teach the why" is an attempt to counteract this—ensuring junior professionals understand reasoning behind conclusions they didn't produce. Sound instinct. But "why" as concept is categorically different from "why" as lived experience. You can explain to a junior auditor why a reconciliation procedure matters. That doesn't substitute for personally discovering the discrepancy it was designed to catch.
The compression is quiet. It doesn't produce visible failures immediately. It produces professionals who advance with formally intact records of review but progressively thinner judgment beneath them. The risk compounds as today's juniors become tomorrow's partners, carrying forward boundaries they were never substantively trained to fill.
What leaders can do next
This is not a training problem with a training solution. It is a design problem. The question is not "how do we teach better?" but "have we defined where judgment lives in our organization, and are those definitions still real?"
Three design questions can begin to clarify the path forward.
First: make decision boundaries explicit by role—not by assumption. In many organizations, junior boundaries are implicit. Everyone "knows" what a first-year analyst is responsible for, but that knowledge lives in culture, not architecture. When the work environment changes, implicit boundaries dissolve unnoticed. The design task: articulate, for each role, what specific judgments that role owns, what authority those judgments carry, and what feedback mechanisms confirm judgment is actually occurring.
Second: preserve or deliberately reconstruct the intermediate steps that build intuition. This doesn't mean banning AI or manufacturing busywork. It means identifying which steps in a judgment sequence are essential for developing evaluative capacity—and ensuring junior professionals experience them, even when an agent could handle them. The design question: in each workflow where an agent operates, which intermediate steps must a human walk through to develop the judgment needed to evaluate the agent's conclusions?
Third: audit where "review" has become ceremony rather than judgment. In any organization deploying AI agents, there are already processes where human review is formally required but substantively hollow. The sign-off is a compliance act, not a judgment act. The design task: identify these points and decide whether to redesign the boundary to make review substantive, or relocate accountability to where judgment actually resides.
The question returns to your organization
The Big 4's talent challenge is clarifying because these firms have articulated, with unusual candor, a problem most organizations have not yet named. When a firm's own AI talent lead openly admits to not having a complete answer, the signal is not weakness—it is the sound of an unnamed structural gap becoming audible for the first time.
The apprenticeship model professional services relied on was never designed. It was an emergent property of high-volume repetitive work, within which judgment encounters happened to be distributed frequently enough to develop professionals. AI agents removed the repetitive work, and with it, the incidental architecture of professional development.
The firms' responses—better training, output verification, early strategic exposure—are genuine and useful. But they operate at the layer of Decision Support. They improve judgment quality without addressing a prior question: where is judgment located, who owns it, and is there real substance inside each boundary?
This is not a Big 4 problem. Any organization that relied on "learning by doing"—OJT, apprenticeship, "watching and absorbing"—rests on the same undesigned foundation. When the conditions that produced growth change, the belief offers no fallback.
Decision-making is happening. Decision Support is improving. But decision boundaries—the lines that define where judgment lives and whether it is substantively exercised—are, in many organizations, either undefined or quietly hollow.
In your organization, where junior professionals are said to be "making decisions"—what is actually happening inside that boundary? Is it judgment? Or is it the appearance of judgment, with the substance quietly compressed away?
The answer is not a training program. It is a design—one that makes boundaries visible, ownership explicit, and the location of judgment deliberate.
This article was prompted by Business Insider's February 2026 report on the Big 4's AI revolution has a problem: how junior staff actually learn (Polly Thompson). The analysis and frameworks presented here are those of Insynergy.