AI has crossed a threshold. Organizations are no longer simply deploying artificial intelligence as a tool that helps people do their jobs. They are deploying AI as an actor that initiates, executes, and reports back—moving through decision loops that leaders used to own. The output is faster, more consistent, and often more polished than what any individual could produce alone. The judgment behind it belongs to no one.
Consider “Elena,” chief revenue officer of a midmarket B2B software company, who deployed an agentic AI system to manage pipeline forecasting and deal prioritization. The agent created a weekly list of recommended actions for her regional VPs. Forecast accuracy improved. Elena presented the deployment as a win. The agent was not a generic tool. It had been trained on three years of proprietary pipeline data—every deal won, lost, and recovered. The C-suite was fully committed. The VPs had watched it call outcomes they could not have predicted. Their confidence in its recommendations was well earned. That is what made it so hard to see what was being lost.
Six months later, the company lost three enterprise deals the agent had ranked as low-priority—deals Elena’s VPs would have pursued on instinct, based on relationship signals no database captures: a champion in a family-run business whose word carried more weight than her title suggested; a nominally small client making critical introductions that opened larger account opportunities; and a pilot that looked, by the numbers, like the smallest deal in the pipeline, but was in fact the account’s trust-building on-ramp before any seven-figure commitment ever got approved.
When Elena tried to reconstruct what happened, neither she nor her VPs could explain the agent’s scoring logic. They had been approving its recommendations for two quarters without interrogating the assumptions. What eroded was not effort—her team was working harder than ever—but the intentionality and forethought that made them effective leaders.
Carly, as professor and practitioner, and Jenny, as organizational transformation advisor and executive coach, see the same pattern repeatedly, and the right lens to fix it comes from Albert Bandura, Stanford professor and psychologist whose work on self-efficacy and human agency reshaped how we understand human motivation and learning. He identified four properties that make people agents of their own behavior rather than passive responders to it—intentionality, forethought, self-reactiveness, and self-reflectiveness. Agentic AI, left unchecked, erodes these.
1. Restore Intentionality—make the Human the Author, Not the Approver
The most common failure mode in agentic AI environments is a gradual role shift. Professionals who once generated ideas and analysis begin to rely on AI for initial outputs. They review and approve rather than creating.
To counter this shift, leaders need to reintroduce intentionality into the workflow. Before engaging with AI on any meaningful task, individuals should articulate their objective, their perspective, and how they plan to use AI to support their thinking. Had Elena’s VPs been required to form their own read on which deals to pursue before the agent surfaced its rankings, the agent’s scoring would have been an input, not a verdict.
Require team members to answer three questions before engaging AI on any meaningful task: What am I trying to accomplish? What is my initial point of view? How is the AI serving my goal rather than defining it? This prebrief takes five minutes and preserves the orientation that separates a thinking professional from a processing function.
Make “What were you trying to accomplish?” a standard question in output reviews—not as a challenge, but as a genuine inquiry into whether intentionality preceded delegation.
2. Rebuild Forethought—expect Before You Inspect
Forethought is the ability to anticipate outcomes and form hypotheses before results are available. AI tools that generate immediate outputs eliminate the perceived need to think ahead; but without forethought, individuals lose the ability to assess whether an output is reasonable. When a professional can generate a complete competitive analysis in 40 seconds, the pressure to form her own hypothesis first effectively disappears.
Leaders can address this with one norm: Before reviewing any AI output, articulate what you expect to see. That hypothesis creates the benchmark needed to evaluate it honestly. Elena’s lost deals were not unforeseeable—a VP who had formed their own view of which accounts were strategically critical would have caught the misalignment before it became a missed quarter. Research on AI and workplace judgment notes that organizations that fail to protect “stretch experiences”—moments of genuine intellectual challenge—risk ending up with managers who have never done the underlying work and thin leadership pipelines as a result.
In high-stakes deliverables, require a one-paragraph prebrief capturing each professional’s own hypothesis before AI output is reviewed.
3. Protect Self-Reactiveness—Design Friction That Develops Judgment
Self-reactiveness is the ability to monitor and regulate one’s own thinking in real time—to notice disengagement, question assumptions, and adjust course. In agentic AI environments, this capacity is undermined by automation bias: the tendency to accept machine-generated outputs without sufficient scrutiny. Research shows that uncritical acceptance of AI recommendations is the default behavior, not the exception, unless structural prompts for critical evaluation are explicitly built into the workflow.
Add healthy friction, the productive discomfort of being challenged, to your AI workflows by auditing them for moments where human judgment is required and consequential. Introduce structured reviews, decision checkpoints, or requirements to justify key assumptions.
4. Institutionalize Self-Reflectiveness—Make Growth Structural
Self-reflectiveness is the ability to step back and ask not just “did we perform well” but “is the way we are operating making us more or less capable over time?” It is the most overlooked dimension of agency, and with agentic AI it is the most urgent.
This is what Chris Argyris called double-loop learning—not just “did we achieve the goal?” but “should we be pursuing this goal at all, and is the way we operate building or eroding our capacity to pursue it well?” With agentic AI, single-loop thinking is the default: Did the forecast improve? Did the campaign launch? Double-loop reflection—Are we becoming less capable of knowing whether these outputs are right?—has to be deliberately designed in.
In an AI-augmented organization, this is the question every leader must ask: Are my people developing through their work with AI, or are they being gradually substituted by it?
To address this, embed reflection into the operating rhythm—in one-on-ones and retrospectives that ask not just what was produced but how: “Where did you engage deeply?” and “Where did you rely on AI without fully understanding the output?” McKinsey’s research suggests that the workers who will sustain value in AI-integrated organizations are those who ask better questions, interpret results, guide machines, and exercise judgment. That capacity requires deliberate investment—not just time freed up.
Design for human jdugement
Reinvest efficiency gains deliberately and visibly. If AI agents free up 30% of a team’s time, name what that time is for. Efficiency redirected into more output is not development.
Elena’s lost deals are not strategy failures. They are agency failures—ones that unfolded without a single moment of obvious error, inside organizations that believed their AI deployments were working.
The difference between agentic AI that builds capability and agentic AI that erodes it is not the technology. It is whether leaders have deliberately designed the conditions under which human judgment is still required, exercised, and developed. Bandura called those conditions the foundation of human agency. They are also the conditions that AI cannot replicate.
The leaders who do that work will find their organizations get sharper the more their agents operate. The ones who do not will discover that their best people have become highly efficient auditors of decisions they no longer know how to make.
