Leading in the Age of AI: Discernment Is the Work

This is Part 3 of Thayer Leadership’s “Leading in the Age of AI” series. In Part 1, France Hoang, J.D., argued that AI does not replace leadership; it exposes it. In Part 2, MG (Ret.) Keith Thurgood took that argument into practice, offering the DEPTH model and three immediate actions leaders can take to close what he calls “the missing middle” between AI potential and organizational results. He named five human capacities no algorithm can replicate, and he put discernment first on that list.

Artificial intelligence has fundamentally changed what it means to do the work. But it has not changed what it means to lead.

If anything, AI has made leadership more visible. It exposes whether an organization can define the right problems, validate the answers it receives, and build judgment in people before speed outruns wisdom.

Discernment deserves a closer look because in an AI-enabled organization it is not simply one leadership capacity among several. It is where the work now lives. The leader’s job, day in and day out, is to build discernment in every room at every level.

Generation Got Easier. Discernment Got Harder.

Consider any piece of knowledge work in four phases:

1. define the problem
2. design an approach
3. create the output
4. validate the result

For most of professional history, human effort was heaviest in the two middle phases. Design and create were the hill you had to climb. Schools trained you for that climb. Companies promoted the people who could make it.

Generation is what AI does best: fast, fluent, and nearly free. The climb becomes a drop, and a finished-looking draft appears in seconds. The ease reads as arrival. There is output, so the organization mistakes motion for progress.

But the work did not vanish. It moved. It shifted to the two ends:

• defining the right problem before you begin
• validating the result before you call it finished

Colonel Chris Lowrance, Associate Professor and Deputy Department Head for the Department of Electrical Engineering & Computer Science (EECS) at the United States Military Academy at West Point, first framed this inversion, and I have come to call the shape it produces the Lowrance Curve: a valley where a hill used to be. The middle sinks toward zero. The two rims rise. Both rims are discernment.

This is not an academic distinction. A 2024 RAND analysis found that AI projects fail at roughly twice the rate of information-technology projects that do not involve AI. The leading cause was not weak models. It was organizations misunderstanding or miscommunicating the problem they set out to solve. Stanford and BetterUp researchers have given the downstream artifact a name: workslop — AI output polished enough to pass as finished but hollow enough that whoever receives it has to redo the work.

Most leaders have seen this already. It is the market analysis that looks polished but never asks whether the target customer exists. It is the policy memo that summarizes sources but misses the legal constraint. It is the sales email campaign that sounds fluent but misunderstands the buyer. It is the code that runs but solves the wrong problem.

Workslop is not a technology failure. It is a discernment failure. Most leaders reading this have received it, sent it, or approved it in the last thirty days.

The Trap: Acquisition Is Not Adoption

When leaders sense this failure, most reach for the oldest instinct there is: when the journey feels too slow, ask for a faster horse. AI looks like the answer. Buy more licenses. Mandate adoption. Send the memo. Change nothing else and expect to arrive sooner.

But AI is not a faster horse. It is a car; a genuinely new kind of machine that can reach places no horse could ever go. And a car is only as good as the road beneath it. Put a sports car on a rutted cow path, and it is not faster. At best, it rattles to the same place. At worst, it sinks in the mud.

Most AI transformations skip straight to the car. Fast machines on old roads. University systems hand AI to half a million students with no training and get panicked faculty returning to blue books. Well-meaning companies buy a thousand chatbot seats, touch no workflow, and wonder why they are still stuck in the mud.

Acquisition is not adoption. Buying a thousand licenses is acquisition. Building the discernment to use them well is adoption. A subscription without support is a car without a road.

Discernment cannot be purchased. It cannot be licensed. It cannot be prompted into existence. It is built deliberately, at every level of the organization, or it is not there at all. That is the leader’s job now.

What Leaders Must Actually Build

Tools alone will not close the gap. Coaching, workflows, and training will.

For leaders, that means three moves:

1. coaching people to frame problems well
2. redesigning workflows so validation is unavoidable
3. training people in ways that protect the struggle required to build judgment

It must become an organizational system: coached in people, embedded in workflows, reinforced through training, and all protected by trust.

1. Coach the framing

The single most consequential decision in any AI-enabled workflow is how the problem gets framed. Frame it well and even a mediocre model can produce useful work. Frame it poorly and the best model on the market produces confident nonsense.

In most organizations, framing authority still sits at the top. Senior leaders define the problem; everyone else executes. AI puts that model under real strain. If only senior leaders frame and think, then the organization loses the distributed judgment it needs to keep up with the pace of change.

The move is to push problem-definition down, not up. Teach frontline employees to interrogate the ask before they open the AI chatbot. Make “what problem are we actually solving?” a standing human-asked question among team members at the start of every project, every meeting, every workflow. Reward the person who reframes a bad question more than the person who answers it fastest.

This is coaching. Not coaching in the soft sense, but in the leader-development sense: teaching people how to think before asking them to move faster.

2. Build validation into the workflow

If framing is the near rim of the Lowrance Curve, validation is the far rim. It is also the rim most organizations skip.

An AI system that checks its own work is not validating. It confirms that the output matches the specification it was given. Real validation asks harder questions.

• Does this actually answer the real problem?
• What did it assume?
• What did it miss?
• Whose experience is not represented in the data?
• What happens if it is wrong?

Those are human questions. They require someone accountable, by name, for the answer. Leaders should install a validation gate in every workflow where AI touches consequential work. No AI-assisted output ships without a human who owns it, and that human’s job is not to rubber-stamp the draft. It is to interrogate it.

In practice, this might mean a required “AI Assumptions” slide in every strategy deck, a named human owner for every AI-assisted deliverable, or a rule that no AI-generated code, analysis, legal summary, or customer communication moves forward without human peer review.

The question is not, “Did AI produce this?” The question is, “Who validated it, against what standard, and what did they check?”

That is workflow design. It makes discernment unavoidable.

3. Train by protecting the struggle

This is the move most leaders resist, because it looks like inefficiency. It is not. It is investment.

Learning science calls it desirable difficulty: the effort that feels wasteful in the moment is exactly what builds durable skill and judgment. Generation used to be how juniors learned to discern. You produced for years, and the doing of it slowly taught you to judge. AI now does the generating, the very rung careers used to start on.

If leaders let their people skip that climb entirely, they are not accelerating development. They are stealing the apprenticeship. The junior who never wrestles with a bad first draft never learns why a good one is good. The analyst who never builds the model by hand never develops the intuition to know when the model is lying. The student who never struggles to organize an argument never develops the internal standard to evaluate one.

Protecting the struggle does not mean banning AI. It means being deliberate about when to hand generation to the machine and when to make people do the reps. Some work should be accelerated. Some work should be practiced. Leaders have to know the difference.

That is training. Not training as a one-time module on prompting, but training as the deliberate development of human judgment.

Leaders Must Model the Standard

If a CEO accepts polished AI output without asking what problem it answered, what assumptions it made, and who validated it, the organization learns that speed matters more than judgment. If a senior leader rewards volume over accuracy, the organization learns to produce more workslop faster. If a manager quietly fixes AI-generated errors without naming the failure mode, the team loses the lesson.

But the opposite is also true. When leaders pause to interrogate the work, name the uncertainty, ask who validated the result, and reward the person who catches the flaw, the organization learns that discernment is not bureaucracy. It is excellence.

The leader’s behavior becomes the standard. The standard becomes the culture. The culture becomes the road.

Trust Is the Precondition

None of these moves work without trust.

Discernment collapses the moment people fear that using AI means giving their hard-earned work away, exposing sensitive data, or creating a record that will later be used against them. People will not experiment honestly if they feel exposed. They will not use approved tools if they do not trust them. They will not disclose mistakes if they believe those mistakes will be weaponized.

In testimony to Congress on small business and AI on July 14, 2026, I made this point plainly: no owner should have to fear that using AI means giving her work away. The same principle applies inside every organization. No employee should have to guess whether using AI will compromise their work, their customer, their company, or themselves.

Leaders own the trust environment. That means being explicit about which tools are safe to use and which are not. It means establishing clear standards for data protection and non-training on organizational data. It means explaining how AI-assisted work will be evaluated. It means making the guardrails visible enough that people do not have to choose between freezing and going underground.

If your people cannot tell where the guardrails are, they will either avoid the tools or use them in ways you cannot see. Neither builds discernment.

The Bar and the Road

AI has changed what it means to do the work. It has not changed what it means to lead. Leaders still define reality, build people, create trust, and hold the standard. What has changed is where that leadership must now show up: before the prompt, after the output, and throughout the organization.

Part 1 of this article series argued that technology raises the bar for leadership. Part 2 gave us the framework and the playbook for crossing it. This part 3 carries the through-line one step further.

The bar is discernment, and the leader’s job is to build it: in themselves, in their people, and into the road the organization drives on.

AI hands everyone a faster car. The leaders who matter in this decade will be the ones who build the better road: shared judgment, distributed framing, disciplined validation, protected struggle, and the trust that lets all of it hold together.