The Leadership Cost of an Agreeable Workforce

A team that works well together is an asset. A team that cannot challenge its own conclusions is a risk. With AI now contributing to analysis, recommendations and decisions, leaders need to know which kind of collaboration they have built.

In September, students at HIM Business School ran commodity trading companies through a live market simulation. Their scores revealed a pattern: 89 for providing support, 84 for coordinating and 63 for negotiating.

These are capable students demonstrating skills their education and future employers value. The question is whether we give them equally strong reasons to question, negotiate and hold their ground.

Our May 2025 employer research described a similar graduate profile: comfortable cooperating, less comfortable confronting disagreement, and inclined to resolve conflict through an intermediary.

For executives, this is a question about the incentives shaping the workforce they are recruiting. What happens when employees who have learned to seek approval start working with machines that can reward them with agreement?

We recruit for agreement and promote for judgement

Our research at HIM maps internship advertisements against ESCO, the European Commission’s classification of occupations and skills. The dataset described here includes 30,987 postings collected since March 2026, covering 86 cities, 69 countries and more than 5,000 companies.

Across 9,036 finance internship postings, skills associated with agreement appeared in 77%. Skills associated with disagreement, including negotiating and critical thinking, appeared in 10%.

Advertisements reveal what employers explicitly request; they cannot capture everything an employer values. Some of this imbalance also reflects junior roles. Nevertheless, the signal to applicants is clear.

Postings mentioning AI listed an average of 3.4 additional skills compared with the same employers’ other postings. None of those additional skills were interpersonal.

We recruit for agreement. We promote for judgement.

That is the tension. Early careers reward responsiveness and fitting in. Senior responsibility requires people to challenge assumptions, negotiate competing interests and deliver unwelcome conclusions. We should not assume that the transition happens automatically.

DDI’s assessment of more than 70,000 manager candidates found that 49% did not demonstrate effective conflict management. This is already a leadership capability gap.

A 2012 study by Timothy Judge, Beth Livingston and Charlice Hurst also found a negative relationship between agreeableness and income, stronger for men. It does not prove that disagreement produces good judgement. It does expose a mismatch between the behaviors we encourage and those the workplace may reward.

The machine learned it from us

In April 2025, OpenAI rolled back a GPT-4o update after it became excessively agreeable. Its explanation identified too much emphasis on short-term user feedback, including signals such as thumbs-up and thumbs-down.

An answer can earn approval because it is useful. It can also earn approval because it confirms what somebody wanted to hear.

Research evaluating 11 AI models found that they preserved users’ desired self-image around 45 percentage points more often than human responses in the advice settings studied. Related research published in Science found that even a single interaction with sycophantic AI could strengthen participants’ conviction that they were right while reducing their willingness to repair interpersonal conflict.

These findings concern particular models and interpersonal advice, rather than every business application. But they identify a mechanism leaders should recognize: agreement can increase confidence without improving the underlying judgement or output.

We have long known how difficult it is to challenge someone who controls our appraisal or promotion. AI introduces another version of the problem: an adviser whose responses may be shaped by our preference for approval.

The machine did not invent that preference. It learned to accommodate it.

The real risk is correlated error

One agreeable colleague may leave an assumption unchallenged. Across an organization, repeated reliance on similar AI tools can leave the same assumption unchallenged in several departments.

Imagine a director, a finance team and an external adviser each using the same model to assess a proposal. Their conclusions look independent. Yet similar prompts, assumptions, and model tendencies may lead them towards the same blind spot.

The concern is not that AI is wrong. The concern is that everyone may become wrong in the same way.

This is a risk scenario, not a claim that AI invariably produces identical answers. The practical question is whether apparently separate reviews provide genuinely independent scrutiny.

Giving agents authority to act adds another exposure. Stanford’s research on automated negotiation found that stronger agents could exploit weaker ones. In one example, an agent acting for a buyer hoping to spend $500 agreed to a $900 purchase.

That example demonstrates a failure to respect a constraint; it does not establish sycophancy as its cause. For an executive, however, the consequence is concrete: delegation requires enforceable limits.

As decision volume increases, human review can become a formality. Someone still owns the budget, authorizes the transaction and answers for the result.

What the skills of agreement cost

The longer-term cost may be the loss of the work through which people learn judgement.

In BCG’s study of 70 C-suite leaders and senior executives, almost 90% cited overreliance on AI outputs without stress testing or challenge. The research also identified concerns about junior development as AI absorbs formative analytical tasks.

Drafting an imperfect analysis, defending its assumptions and revising it after criticism can be slow. It is also how people discover what they understand.

If AI removes those repetitions, organizations need to replace the learning deliberately. Otherwise, they risk expecting independent judgement from employees who have had fewer opportunities to practice it.

The same question applies to experienced managers. Time saved has value. But when speed becomes the dominant measure of performance, checking the answer starts to look like a delay.

Leaders determine whether that delay is treated as responsible work.

How to install doubt

Telling people to “think critically” leaves too much to chance. Constructive challenge needs a place in the workflow, an accountable owner, and permission to change the outcome.

Descartes offers a useful discipline: question what appears certain, break the problem into parts, work through them systematically, and review what may have been missed. His thought experiment asked what would survive a powerful deceiver making falsehood appear true.

Descartes needed a thought experiment. You have a subscription.

AI need not be deceptive for that discipline to help. Fluency alone can make a weak argument feel settled.

Begin with a rule:

No AI-generated recommendation should enter a file without the counter-case beside it.

The person relying on the recommendation should record the strongest reason it might fail, the evidence that would change their mind and any relevant information the model has not received. A ten-minute exercise is a practical starting point; higher-stakes decisions require more scrutiny.

Ask the model to argue the opposing case, then test both arguments against evidence. A generated objection is useful rehearsal, but it is not independent verification.

Make room for knowledge outside the prompt: a customer’s hesitation, a supplier’s capacity constraints, an operational problem that has not reached the dashboard. Human value includes knowing which missing fact could overturn an otherwise persuasive answer. Challenge also matters when it uncovers faulty reasoning or unreliable sources.

For consequential decisions, obtain an initial assessment before sharing the AI recommendation. For agents, enforce spending limits and approval thresholds through system controls.

Above all, reward the person who identifies the flaw. A counter-case requirement will become paperwork if challenging the preferred answer damages someone’s prospects.

Make room for critical thinking in the workplace

Our analysis suggests that AI requirements could appear in one in five finance internship postings within six months, compared with roughly one in eight in the dataset described here. That is a projection, not a settled outcome. It reinforces the need to develop judgement alongside technical adoption.

At the beginning of each semester, I offer my students a challenge: prove me wrong and receive 100 for the course. The purpose is to make disagreement worth attempting, and to require evidence strong enough to sustain it.

Employers need their own version of that invitation.

Business schools, including HIM, share responsibility for preparing people to cooperate and to challenge. Executives share responsibility for ensuring both behaviors remain possible once graduates enter the workforce.

When answers are cheap, leadership comes down to knowing which question is worth asking. That takes people who can spot what’s missing, show why it matters to the business, and speak up before agreement hardens into a decision. Dissent may be what defines us as human: a machine can object when asked, but only a person can risk something by doing it.

Frequently asked questions

What is AI sycophancy?

AI sycophancy is the tendency of an AI model to tell users what they want to hear: agreeing with their views, praising their ideas or accepting their mistakes instead of giving an accurate, balanced answer. OpenAI described its rolled-back GPT-4o update as 'overly flattering or agreeable'.

What is agreeableness?

Agreeableness is one of the Big Five personality traits, alongside openness, conscientiousness, extraversion and neuroticism. People high in agreeableness tend to be cooperative, trusting and kind, and are more inclined to avoid conflict.

What is groupthink?

Groupthink is the tendency of members of a close-knit group to accept what they see as the group consensus, whether or not they believe it is correct. The concept is most closely associated with psychologist Irving Janis, who studied it in policy decisions in the 1970s.

Does using AI reduce critical thinking?

A 2025 Microsoft Research and Carnegie Mellon University survey of 319 knowledge workers found that higher confidence in generative AI was associated with less critical thinking effort, while higher self-confidence was associated with more. The findings are self-reported.

What is psychological safety at work?

Psychological safety is a shared belief that a team is safe for interpersonal risk-taking, such as speaking up, asking questions or admitting mistakes, without fear of embarrassment or punishment. The term was defined by Harvard Business School professor Amy Edmondson in 1999.

Which skills do employers value most?

In the World Economic Forum's Future of Jobs Report 2025, analytical thinking is the top core skill, considered essential by seven out of ten companies, followed by resilience, flexibility and agility, and leadership and social influence.