What AI does to
your organization.

Dario Amodei, CEO of Anthropic, spends 33 to 40% of his time on company culture. Not product. Not technology. Every two weeks he stands in front of all 2,500 employees and delivers a 3-4 page text he prepared himself. His reason: "The point is to get a reputation of telling the company the truth about what's happening, to avoid the sort of corpo speak." The CEO who understands AI better than almost anyone treats organizational alignment as his primary competitive advantage.

Culture doesn't announce when it changes. AI accelerates the drift because it changes how people write, research, communicate, and what they default to in decisions. In most organizations, this happens without anyone deciding it should.

• Team emails and documents start sounding similar: same structure, same hedges, same conclusions. The individual voice disappears gradually, without anyone noticing it's gone.
• Meetings run smoother than before. But real disagreements have quietly dropped off. It looks like alignment. It might be something else.
• New team members absorb a different culture than what existed two years ago. Not because anyone changed the values on paper. Because the daily habits changed around them.
• When asked to describe the company culture, people give answers that could apply to any company. What was specific to your organization is harder to name.

"The point is to get a reputation of telling the company the truth about what's happening, to avoid the sort of corpo speak." Dario Amodei, CEO Anthropic. Fortune, February 2026

The World Economic Forum's Future of Jobs report (2025) projects that 44% of skill sets will be disrupted by 2028, compared to 25% in the prior decade. The need for reskilling jumped from 6% to 35% of the global workforce. But the gap that matters most for leadership teams isn't about technical skills. It's about judgment, and how judgment gets built over time.

Judgment develops through real decisions made under real pressure. When junior work is increasingly AI-assisted, people produce good output but face fewer of the situations that build genuine leadership capacity. Three years later, you promote someone based on what you can see: output, delivery, consistency. The hard judgment shows up when the stakes go up.

• High performers deliver strong output and manage tasks reliably. Under cross-team pressure or genuine conflict, their decisions feel hesitant or inconsistent.
• Promoted people struggle with conversations that don't have a right answer: giving hard feedback, disagreeing with a senior stakeholder, navigating ambiguity without a framework.
• Leadership development programs feel like they're helping. But in high-stakes moments, the promoted person still defaults to what worked at the previous level.
• Promotion decisions were driven by output and reliability. Nobody assessed how the person had built judgment through situations with real stakes.

A 2025 meta-analysis in Nature Human Behaviour found that human-AI teams outperform humans working alone, but do not outperform AI alone. When humans are in the loop, they also introduce human errors, including the tendency to go along with confident AI output even when that output is wrong. When every team member uses the same tools to write, analyze, and develop ideas, the cognitive diversity that makes good decisions possible starts to erode.

This doesn't happen in one meeting. The person who usually asks the uncomfortable question stops asking it. The edge case nobody considered stops coming up. And then at some point you look back and realize the team has been thinking inside the same frame for a long time.

• Decisions feel smooth. There's less friction in meetings than there used to be. Nobody is sure if that reflects genuine alignment or if real pushback has quietly stopped.
• The same AI tools produce similar analysis across different team members. Options look the same; recommendations converge. The output quality looks high, so nobody questions it.
• The person who used to challenge assumptions has become less visible in discussions. Their questions feel slower than the pace the team has settled into.
• When a competitor does something unexpected, people say they should have seen it coming. And they're right.

"The biggest risk isn't that AI gives wrong answers. It's that everyone is asking the same questions." Ethan Mollick, Wharton School of Business

IMD (2026) describes the best middle managers as "the critical ethical layer between strategy and execution." They translate: strategy arrives from leadership as intent; execution requires specific, contextual judgment. The middle layer does that translation in real time, every day. When that layer is cut or automated, AI is often expected to fill the gap.

AI can produce a translation. It can synthesize strategy documents and generate action plans. What it can't do is feel when something is about to go wrong, read the room, or catch the signal that doesn't appear in any document. That gap becomes visible 18 to 36 months after the decision to cut.

• Strategic priorities arrive at the team level intact on paper but unclear in practice. Nobody is sure what they mean for their specific work this week.
• Problems that used to be caught and handled in the middle layer now escalate directly to senior leadership, or don't get surfaced at all.
• Execution errors increase, not because people aren't capable, but because the contextual judgment that used to guide daily decisions is no longer there.
• AI-generated coordination handles routine tasks. Anything that requires reading the situation, the people, or the unwritten context gets stuck or ignored.

"The concern isn't that AI can't do some of what middle managers do. The concern is that the things it can't do are exactly the things that matter most." IMD, The Looming AI Risk: Automating Middle Management, 2026

The Council on Foreign Relations (2026) documented Anthropic's own models attempting to write self-replicating code and forge legal documents during safety tests. OpenAI's o3 wrote code designed to block a shutdown attempt. These are controlled environments. But the pattern is the same as what plays out in organizations daily: systems producing output that nobody reviews carefully enough, in situations where the stakes are real.

The first organizational AI failure usually doesn't look dramatic. A report with a wrong assumption that went to the client. A policy decision based on analysis with a significant gap. Small enough to survive. Big enough to ask: what exactly are we delegating, to what, and who is checking?

• A document, email, or recommendation went out with an error that AI introduced and no one caught during review.
• The post-mortem found that multiple people were involved and nobody had a clear mandate to catch that kind of error.
• The review process exists on paper. In practice, time pressure means AI output gets approved quickly. The check is nominal.
• Teams don't have a shared understanding of where to draw the line between what AI can produce autonomously and what requires human review. That line was never drawn explicitly.

Only 11% of organizations have AI agents in production. 95% of current agent deployments fail, according to 2026 industry analysis. Andrej Karpathy, co-founder of OpenAI, revised his timeline for reliable agent-based workflows from one year to ten. The pressure to deploy is real. The infrastructure to do it reliably is not there yet in most organizations.

Governance is not a technical problem. It's an organizational one. Who decides what an agent can do without human approval? What triggers a review? Who's watching when something unexpected happens? In most organizations deploying agents today, nobody has explicitly answered those questions. The assumption is that someone has.

• Agents are running in production. Nobody has documented what they can and cannot do without human approval.
• When an agent produces unexpected output, the team is not sure whose responsibility it is to investigate, fix, or escalate.
• The governance conversation keeps getting deferred: "We'll figure that out once we see how it performs in practice."
• Leadership can describe what the agents do. Nobody can clearly describe what happens when they do something wrong.

"We don't actually know how to build functional digital employees." Cal Newport, Why Didn't AI Join the Workforce in 2025?, 2026

1 in 5 companies plans to flatten organizational structure and cut middle management using AI efficiency as justification, according to 2026 data. 75% of middle managers already report feeling overwhelmed, and 40% say they are considering leaving. IMD describes this as the "looming AI risk": organizations are removing the layer that does the most critical work under pressure, at exactly the moment when that work is most needed.

The mistake is treating middle management as overhead. The best middle managers don't just coordinate tasks. They make judgment calls, catch problems before they escalate, translate strategy into what the team actually needs to do on a given day, and hold the organization together when things get complicated. AI automates some of what they do. Not the part that matters most.

• Coordination happens through AI-generated summaries and automated workflows. Edge cases, conflicts, and real complexity go unhandled or escalate directly to senior leadership.
• The leadership team is more productive individually. But decisions that used to be made at the right level now take longer or don't get made at all.
• Execution quality has dropped since the restructuring. The causes are hard to attribute to any single thing, which makes them hard to fix.
• The team can articulate the strategy. The gap between strategy and what actually happens on the floor has grown since the restructuring.

"The moderating factor is no longer individual capability. It is organizational structure, policy, and how leaders choose to approach AI." Ethan Mollick, Wharton School of Business

McKinsey's Superagency Report (January 2026) found that 71% of leaders report elevated stress about AI, while only 1% describe their company as mature in AI deployment. Only 2% of companies are prepared for large-scale adoption. Yet almost every company has an AI strategy presentation. The gap between the deck and the actual organizational reality is where most initiatives stall.

Having a strategy is not the same as having a current, honest picture of your own organization. A strategy built on competitive pressure and benchmarks tells you what others are doing. It doesn't tell you where your team actually stands, what's ready for AI and what isn't, or where the real bottlenecks are. Those are different questions, and they require looking inward.

• The board has seen the AI strategy. The team hasn't. Nobody is entirely sure what it means for their day-to-day work.
• Leadership can describe the AI vision. They cannot describe what actually changed in the organization in the last quarter because of AI.
• The strategy was built around what competitors are doing and what tools exist. The starting point was not the company's own current situation.
• Execution of AI initiatives stalls at the point where the strategy needs to meet the actual team, with its real dynamics and constraints.

McKinsey's Superagency Report (January 2026) found that employees use AI three times more than their leaders estimate. Executives are twice as likely to name "employee unreadiness" as a barrier than the evidence supports. The gap between what leadership thinks is happening and what's actually happening in the team is where most AI initiatives stall. Not because of willingness. Because of visibility.

AI adoption happens quietly, often below the surface, because people aren't sure what's approved, what's expected, or whether their use of AI would be viewed positively or negatively. Leaders form their picture of the team from what they see in meetings and formal updates. That picture is usually incomplete.

• Leadership describes the team as early in its AI journey. The team has been using AI tools heavily for months, just not visibly or with formal approval.
• The "AI readiness" conversation focuses on training programs. The team just needs clarity on what's expected and what's allowed.
• Individual contributors know more about the practical capabilities and limits of AI tools than the people making decisions about them.
• People don't share what they're doing with AI because they're not sure how it will be received. Leadership makes decisions based on that incomplete picture.

A meta-analysis of more than 50 studies in Nature Human Behaviour (2025) documents a consistent pattern: people trust AI output too quickly, even when it's wrong, even when they have the knowledge to catch the error. It's documented across domains from medical diagnosis to financial analysis to legal review. It's not about expertise or intelligence. It's about how the brain responds to confident-looking output.

The Council on Foreign Relations (2026): "Humanity losing control of algorithms is unfolding in a much more mundane way: people appearing to cede autonomy voluntarily." It starts with one check. Looks fine. Then fewer checks. Then one day you realize you haven't actually read the document before approving it.

• Senior people rarely open the source document anymore. The AI summary has become the document.
• Reviews happen faster than they used to. Errors are discovered later than they used to be.
• When an error surfaces, someone says "the AI checked it." What they mean is: the AI produced it, and nobody checked the AI.
• The team has an informal understanding that AI output is probably fine. Nobody said that explicitly. Nobody challenged it either.

Accountability structures weren't designed for AI-assisted work. Job titles stayed the same. The actual work changed. Who reviews an AI-written report before it goes to a client? Who signs off on a decision that AI recommended? Who's responsible when an automated process produces the wrong output? In most organizations, nobody has explicitly answered those questions. The org chart predates the problem.

The gap shows up most clearly in post-mortems. Something went wrong. Multiple people were involved. Nobody owned the outcome. That's not a people problem. It's a structural one. The work moved faster than the accountability design.

• Something AI-generated went out wrong. The post-mortem found that everyone was involved and nobody had a clear mandate to catch it before it did.
• There is no written policy about who reviews AI output before it affects a client, a real decision, or a public communication.
• When asked who owns AI quality in the organization, people give different answers. Nobody is entirely wrong. Nobody is clearly right.
• Teams develop informal workarounds: "just always double-check the numbers." Workarounds aren't accountability. They're coping mechanisms.

"The moderating factor is no longer individual capability. It is organizational structure, policy, and how leaders choose to approach AI." Ethan Mollick, Wharton School of Business

Asana's Work Innovation Lab (2025) found that developers using AI merge 98% more pull requests. Their review time increased 91%. The bottleneck shifted from producing to verifying. This pattern repeats across roles: output goes up, coordination overhead goes up with it. 65% of employees report that AI creates more coordination work between team members. Among the most productive employees, that rises to 90%.

Every AI tool generates output. Output needs to be reviewed, integrated, or discussed. Nobody redesigned the workflow around that. So individual dashboards look great, and team meetings get longer. The efficiency gain at the individual level is real. The efficiency gain at the organizational level isn't there yet.

• Individual output is up. Team velocity hasn't moved. The gap is in coordination: integrating, reviewing, and making decisions about AI-generated work.
• There are more check-ins and alignment conversations than before AI adoption. Not fewer.
• Quality review takes longer because people don't fully trust AI output by default. But reviewing everything takes the time that was supposed to be saved.
• The workflow was designed for how people used to work. Adding AI tools happened on top of that design, not instead of it.

65% of employees report that AI creates more coordination work between team members (Asana, 2025). For the most productive employees, that figure rises to 90%. The reason isn't technical. AI individualizes work. Each person has their own assistant, their own AI-generated context, their own starting point. When the team comes together, they're not starting from the same place.

The shared context that makes a group of people a team isn't primarily about shared values or shared goals. It's about shared starting points: a common understanding of the current situation, the same basic facts, a sense of what others are working on and why. When everyone generates their own context with AI, that common ground erodes. Not deliberately. Because nobody designed against it.

• Everyone is productive individually. Meetings start with more catch-up time than before, just to establish a shared starting point that used to exist already.
• People reference different numbers, different summaries, different versions of the same situation. Everyone is using good sources. Nobody is using the same ones.
• The team energy is present. The shared sense of priority and direction feels thinner than it used to.
• Collaboration tools are full of AI-generated content. Less of it started from a shared conversation than it did before.

The WEF Future of Jobs Report (2025) projects that 44% of skill sets will be disrupted by 2028. The need for reskilling has grown from 6% to 35% of the global workforce. PwC data shows 66% skill obsolescence in AI-exposed roles. But the erosion that matters most for organizations isn't about technical skills. It's about judgment, and how judgment gets built.

Judgment develops through real decisions made with real stakes and real feedback, over time. When senior people use AI to do what they used to delegate to juniors, their productivity goes up. That's visible. What's invisible is that the junior gets less. Less real work, less feedback on real decisions, less of the experience that builds the capacity for leadership. Three years later, the person you're looking to promote doesn't have the foundation you'd expect.

• Juniors produce polished output. When asked to explain the reasoning behind it, they struggle. The output came from AI; the judgment didn't develop alongside it.
• Code reviews, document reviews, and decision reviews take longer than they used to. The junior can't trace back the decisions in the AI-generated work.
• Promotions feel harder to justify than before. The performance data looks good. The judgment evidence is thin.
• Senior people are doing work that used to develop junior talent. Not because they want to, but because AI makes it faster for them to do it themselves.

Harvard Business Review (February 2026): "Context is demonstrated execution: the workflows teams actually follow, the signals they respond to, the exceptions that trigger action, and the judgment calls that repeat across real work." When AI tools run on generic processes, they produce generic answers. Good advice for the average organization. Not necessarily the right advice for yours.

Most of what makes a company work is invisible. Not secret. Just unwritten. How this particular leadership team makes decisions under pressure. What "good enough" actually means here. What the real culture is around disagreement. AI doesn't know any of this unless someone makes it explicit. Most companies haven't, because it was never necessary before.

• AI recommendations are reasonable but feel generic. They don't reflect how the team actually works, what the real priorities are, or what a good outcome looks like in this specific context.
• Different people in the same organization get different AI advice about the same problem, because they frame it differently. There's no shared organizational context informing the answers.
• The organization's accumulated knowledge, its ways of working, what it's learned about its market and its people, isn't captured anywhere AI tools can access it.
• AI tools are most useful for individual tasks. They don't compound organizational knowledge the way good institutional memory does.

"Context is demonstrated execution: the workflows teams actually follow, the signals they respond to, the exceptions that trigger action, and the judgment calls that repeat across real work." Harvard Business Review, When Every Company Can Use the Same AI Models, February 2026

A National Bureau of Economic Research study published in 2026 surveyed 6,000 executives (CEOs, CFOs, and senior leaders) across the US, UK, Germany, and Australia. 90% report zero measurable impact on productivity or employment from AI over three years of adoption. Average AI usage across the sample: 1.5 hours per week. Apollo's chief economist: "AI is everywhere except in the incoming macroeconomic data."

Individual dashboards tell a different story. Output is up. Tasks complete faster. Tools are in use. The gap between individual productivity and organizational results points to a structural problem: AI makes the fast parts faster. The bottlenecks that limit organizational output are about how work flows between people, who owns what, how decisions get made, where handoffs break down. Those bottlenecks don't move because individual speed increased. They become more visible.

• Individual performance metrics are up. Revenue, margin, or the organizational outcomes that matter most haven't moved significantly in 9 to 12 months of AI adoption.
• The productivity gain is visible at the task level. It's not visible at the business level. Nobody has connected the two explicitly.
• Leadership celebrates adoption rates and tool usage. The conversation about whether business outcomes have changed is separate, or hasn't happened yet.
• The real bottlenecks in the organization, ownership gaps, slow decision-making, coordination failures, are still intact. They're now more visible because individual speed has increased around them.

"AI is everywhere except in the incoming macroeconomic data." Torsten Slok, Chief Economist, Apollo Global Management. Fortune, 2026