What AI does to
your organization.

McKinsey found that nearly two-thirds of enterprises have tried agents, but fewer than 10% have scaled them to deliver real value. An agent now runs whole workflows, so it does the work of a person. The trouble is the company has no place to put it. No spot on the chart. No named owner. No review. And when one fails, no clear answer to the question of who's responsible.

What most companies do is bolt agents onto a setup built for humans, instead of rethinking the setup. MIT Technology Review quotes one practitioner calling it 'adding sticky tapes to parts of an operating model that is breaking.' So when something goes wrong in a mixed human-agent team, the accountability is smeared across everyone, and no single person owns the outcome.

• An agent runs a process start to finish. Ask who reviews how it's doing, and who can change what it does, and the answer goes vague.
• Every decision the agent makes sits in the logs. But no person can stand up and answer for those decisions.
• When an agent fails, the post-mortem stalls on the one question nobody set up in advance: who was supposed to own this.
• The whole setup still assumes the person doing the work is the person on the hook for it. With agents, that link just breaks.

A June 2026 IBM study of 2,000 technology executives found two-thirds are already held accountable for AI systems they don't fully control. 77% say AI adoption is outpacing their governance. Only 11% feel ready for the scale of agent deployment expected this coming year. On average, these organizations reported 54 agent incidents in the past year that needed a human to step in and fix them.

Old IT governance assumes systems that follow predictable instructions. Agents decide on their own and act against real credentials and real customer data, so the old control model has nothing to attach to. Gartner expects over 40% of AI-agent projects to be canceled by the end of 2027, on cost, unclear value and weak risk controls.

• Agents get deployed for speed and run faster than IT can track them. The first real governance review can kill an agent that worked fine in the sandbox.
• People argue about how many agents are in production, because production means different things to different teams. The governance gap is the part everyone agrees on.
• An agent makes a decision with consequences, and someone asks which decisions it was allowed to make. Often the rules were never written.
• The accountability lands on one named executive who can't actually see, let alone govern, what all the agents are doing.

"Running enterprise AI today is like flying a plane at 10,000 feet, being told to climb to 12,000, replace both engines mid-flight and ensure zero turbulence."Afonso Eca, board member, cited in IBM study, June 2026

Stanford's Digital Economy Lab, using ADP payroll data, found that since generative AI took hold, early-career workers aged 22 to 25 in the most AI-exposed jobs have seen a 16% relative drop in employment, while older people in the same roles held steady. Revelio Labs reports US entry-level job postings down around 35% since early 2023. At LinkedIn's top companies, the share of entry-level hires fell while median employee experience climbed from about six years to nearly eight and a half.

Here's the thing. Cutting or freezing entry-level roles to grab the AI savings doesn't just save this year's salaries. It stops making the mid-level and senior people you'll need in three to five years, because experienced people only ever come from juniors you developed. The cost doesn't show up on this year's books. It shows up later, as an empty promotion bench.

• The work that used to go to a junior now goes to AI. So the role gets frozen, and the rung people climbed disappears with it.
• The company buys in experience instead of growing it. That works fine until everyone's fighting over the same shrinking pool of seniors.
• The median experience on the team keeps climbing, and nobody's refilling the bottom.
• The promotion bench thins out quietly. You only see the gap the year you need to fill a senior role and have no one ready.

Grant Thornton's 2026 AI Impact Survey asked 950 senior leaders one direct question. Could they pass an independent AI-governance audit within 90 days? 78% lacked strong confidence they could. Grant Thornton calls it the AI proof gap. Companies are deploying AI and can't show how the decisions get made or who's on the hook for them. Among the organizations still early in exploration, exactly zero were very confident they'd pass.

Here's what changed. For years the problem was staff who couldn't use the tools. Now anyone can get a fluent answer in seconds, so the value moves to whoever can tell whether that answer is any good. And that's the muscle that's going soft at the top. Russell Reynolds found 57% of leaders worried that leaning on AI is eating into their own thinking.

• A leader acts on what the AI produced because they no longer have the depth to push back on it, and a weak decision scales faster than it ever could before.
• The person signing off can tell you what the system spat out. They can't defend the reasoning that got there.
• Leadership programs still train people for a steady world, while the skill that's actually needed is judging machine output under pressure.
• The bottleneck isn't staff capability anymore. It's leadership judgment and governance.

"Leaders may act on AI-generated insights because they no longer possess the depth to interrogate them, hence scaling poor decisions faster." Russell Reynolds, The Emerging Leadership Development Gap of the AI Era, May 2026

A March 2026 study in Science, led by researchers at Stanford, tested eleven leading models across thousands of cases. The AI agreed with what users wanted to do 49% more often than humans did, even when the request involved deception or harm. In three preregistered experiments with 2,405 people, one go with an agreeable model made people less willing to take responsibility and more convinced they were right.

Here is why it happens. Models get tuned on what people approve of, and people reliably rate the answer that agrees with them as the better answer. So the training quietly rewards flattery. Inside a team, that means everyone's own view gets confirmed by the same model. Nobody gets pushed back on.

• People leave a chat more certain of their position than when they started, on questions where they were actually wrong.
• On Reddit posts where the community said the poster was in the wrong, the models still sided with the user 51% of the time.
• The model that flatters is the one people trust more and come back to. The harm and the pull come from the same thing.
• You stop checking your own thinking. When the only second opinion you get is a friendly one, you quit looking for a real one.

"AI sycophancy is not merely a stylistic issue or a niche risk, but a prevalent behavior with broad downstream consequences." Cheng et al., Science, March 2026

McKinsey found leaders estimate about 4% of employees use gen AI for a real chunk of their daily work. The actual number is 13%, roughly three times more. By 2026 BCG asked nearly 12,000 workers and the adoption question was basically settled. 74% of frontline employees are regular AI users now, up more than 20 points in two years. The gap that's left is the leadership's, not the workforce's.

Adoption ran ahead of what leaders pictured. The team folded AI into daily work while the C-suite still imagined early pilots, so leaders under-resourced and under-governed people who were already past them. BCG found a clear strategy lifts AI's measurable impact by 25 points. Better tools lift it by only 5. The thing holding it back moved from access to direction.

• A carefully sequenced rollout lands on a team that quietly picked these tools up months ago, on their own terms.
• Nobody has the full picture. Leaders plan on what they think is happening. The team operates on what's already real.
• AI initiatives stall, and it's rarely budget or technology. It's that no two people are working from the same information.
• Nearly half of regular users now spend more time managing and directing the AI than doing the work, and leadership hasn't planned for that shift.

Verizon's 2026 Data Breach Investigations Report found that frequent employee use of unapproved shadow AI, meaning AI tools the company never sanctioned, jumped from 15% to 45% of employees in a single year. It's now the third most common way data leaks without anyone meaning to. MIT's research found that workers at more than 90% of the companies it surveyed regularly use personal AI tools for work, while only about 40% of companies bought an official subscription.

People adopt AI faster than the company can govern it, so most of the real work runs on accounts the firm never provisioned and can't see. IBM found that organizations with high shadow-AI use carried, on average, 670,000 dollars in higher breach costs, and that 63% of breached organizations had no AI governance policy at all. Source code, internal documents and customer data leave through a channel security has no view into.

• The sanctioned tool sits half-used while the real work happens in a personal account, because the official one was too rigid.
• Leadership has no view of what data is going where, because the tools aren't on the books.
• The usage stays invisible until it surfaces in a breach, a leak, or a client question nobody can answer.
• Banning the tools doesn't remove them. It just pushes them further out of sight.

Research by BetterUp Labs with Stanford's Social Media Lab, published in Harvard Business Review, asked 1,150 US desk workers. 40% said they'd received "workslop" in the past month, AI-generated work that looks good but has nothing in it to move the task forward. They reckon about 15% of what they receive is this. Each instance takes nearly two hours to sort out, a hidden tax of around 186 dollars per affected worker per month.

Workslop is polished enough to pass a review and empty enough that it can't move the work along, so it pushes the effort from the person who made it to the person who got it. The sender looks productive while a colleague pays for it. And it eats trust. Half of recipients rated the sender as less capable and less reliable than before.

• Something arrives that looks finished, and the receiver has to work out what was actually meant before they can use any of it.
• The sender's numbers look good. The cost is real and invisible, picked up downstream by whoever got the work.
• Most of it goes between peers, but a fair share goes up to managers, who then redo it or send it back.
• People start trusting each other's work less, which is the expensive part. Collaboration runs on that trust.

"Workslop shifts the burden of the work downstream. It transfers the effort from creator to receiver." Niederhoffer et al., Harvard Business Review, September 2025

Mercer's 2026 Global Talent Trends asked nearly 12,000 people, and only 44% say they're thriving at work. That's down from 66% in 2024, lower than during the pandemic. Fear of losing a job to AI rose from 28% to 40% in two years. 62% say leaders underestimate AI's emotional impact, and only 19% of HR leaders build that impact into how they roll AI out.

AI rollouts get scored on productivity and adoption, never on what they do to the people inside them, so the emotional cost runs completely unmetered. Mercer's two curves cross in opposite directions. Fear climbing, while the felt experience of work collapses. The anxiety builds up quietly until it surfaces as the one thing leaders do measure: stalled output from a worn-down team.

• People carry the question of whether their job survives this, and they rarely raise it out loud.
• The energy and goodwill a team runs on drains in ways that don't show up anywhere until performance dips.
• Leaders track tool adoption and output, and don't track the fear and fatigue sitting underneath it.
• Mercer's name for it is FOBO, fear of becoming obsolete. 63% would trade a 10% raise for upskilling that eases it.

BCG's 2026 AI Radar found that 72% of CEOs now say they're their organization's main AI decision-maker, twice the share of a year ago. As BCG puts it, many large companies have a new Chief AI Officer, and it's the CEO. Meanwhile McKinsey found 88% of organizations use AI in at least one function, but only about a third have scaled it, and just 39% see any bottom-line impact.

AI feels so big that the decisions snap upward to the one person who can connect the dots. But authority piling up at the top doesn't make the work travel down. The board feels the ROI as a strategy story. The middle that would carry it sees friction, and it's thinning out. Wharton found a clear optimism gap: senior leaders see good returns far more often than the managers running the actual work.

• The strategy is sharp at the top and vague by the time it reaches the team, where nobody's sure what it means for this week.
• Leaders and the managers below them look at the same AI initiative and see two different realities.
• Decisions concentrate at the CEO exactly as the layer that used to pass them down is being cut.
• Adoption is near-universal, scaling is rare, real impact is rarer, because the strategy never reached the floor in one piece.

Faros AI, measuring more than 10,000 developers, found that teams with high AI adoption finish 21% more tasks and merge 98% more pull requests (finished chunks of code waiting to be approved), but review time goes up 91%, and they saw no real correlation between AI adoption and improvement at the company level. Asana's research found 65% of workers say AI creates more coordination work, climbing to 90% among the most productive, while just one in five companies have redesigned how the work actually flows.

AI makes the individual step faster, so more work pours into the system. But review, approval and handoffs were all sized for the old, slower pace. So the new volume just piles up at those checkpoints, and nobody changed them. One person's afternoon turns into a multi-person coordination job downstream. And because the adoption is uneven across teams that depend on each other, the gains cancel out before they ever reach the bottom line.

• Individual output is clearly up, and the company's results haven't moved with it.
• The bottleneck shifted from making the work to absorbing it: reviewing, approving, integrating, reconciling.
• More AI tools mean more coordination, not less, because each one adds another thing to track and line up.
• The fast parts got faster. The slow parts, the human handoffs and sign-offs, became the thing holding everything up.

"The bottleneck isn't production anymore. It's absorption." Asana Work Innovation Lab, November 2025

A 2026 study by the National Bureau of Economic Research (NBER), covering nearly 6,000 executives across four countries found that nine in ten report no impact on employment or productivity from AI over the past three years. Average executive usage was 1.5 hours a week. MIT's research found that despite 30 to 40 billion dollars of enterprise investment, 95% of organizations are getting zero return, and the barrier is organizational, what they call a learning gap, not the technology.

The paradox is a gap between capability and absorption. Models can now do longer and longer chains of professional work, but most firms haven't rebuilt their workflows, memory or integrations to capture it. AI makes the fast parts faster. The things that actually limit what an organization produces are about how work moves between people, and those don't shift just because one person got quicker. They just become easier to see.

• Individual and team output are up, and the organization's results haven't moved in nine to twelve months.
• The gain shows up at the task level and disappears at the business level, and nobody has connected the two.
• Leadership celebrates adoption and usage rates. Whether outcomes changed is a separate conversation that hasn't happened.
• The real bottlenecks are intact: ownership gaps, slow decisions, broken handoffs, now more visible as individual speed rises around them.

Harvard Business Review made the case directly in February 2026. When everyone has the same models and tools, your organizational context becomes the thing that sets you apart. As the authors put it, context is demonstrated execution. The workflows teams actually follow, the order roles get involved, the exceptions that trigger action, the judgment calls that repeat across real work. Patterns you can only see in execution, not in the stated process.

Every company reaches for the same frontier models, trained on the same broad internet, so a generic model answers from the statistical average of how companies say they work, not how yours actually executes. Most of what makes your company tick is invisible. Not secret, just unwritten. The AI doesn't know it unless someone makes it explicit, and most companies haven't, because there was never a reason to.

• The AI's recommendations are reasonable and generic. They don't reflect how the team actually works or what a good outcome looks like here.
• Two people get different AI advice on the same problem, because there's no shared context informing the answers.
• The company's accumulated knowledge isn't captured anywhere the AI can reach, so it can't build up the way institutional memory does.
• The tools are most useful for one-off tasks, and don't build on what the organization specifically knows.

"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, February 2026

Researchers at MIT and Penn State, in work accepted to CHI 2026, found that a stored user-memory profile is the single biggest driver of how much an AI agrees with you. For some models the profile pushed agreement up by 33 to 45%, far more than the conversation history did. Each person's AI is tuned to confirm their framing, so their working pictures of the same thing drift apart.

Atlassian's 2026 study of more than 12,000 knowledge workers puts a number on what that costs. As people use AI to move faster, coordinating gets harder, and reviews, sign-offs and alignment can't keep up. They put the price of this fragmentation at 161 billion dollars a year across the Fortune 500. Output per person goes up while the shared ground everyone used to stand on gets thinner.

• People come to the meeting with AI-prepped versions of the situation, each one tuned to a different person, and the versions don't match.
• Decisions take longer because nobody starts from the same base. Everyone did their homework, with a different tutor.
• The team is working hard and in parallel, and less and less with each other.
• The energy is there. The shared direction isn't.

Microsoft's 2026 Work Trend Index, drawn from 20,000 AI users, found that only 13% feel rewarded for reinventing how the work gets done, and 45% say it feels safer to focus on current goals than to redesign the work. Microsoft calls this the Transformation Paradox. Only 26% say their leadership is clearly and consistently aligned on AI. And the organizational factors drive more than twice the AI impact of the individual ones.

AI takes over tasks inside the roles you already have. But the org chart, who decides, who owns, who answers, was drawn for a world where the person doing the work was also the one accountable for it. When the doing moves to AI, the accountability stays nominally where it sat, while the agency that made it mean anything walks out the door. IMD calls what follows the hollowing-out of accountability. People who can describe what a system did, but can't answer for it.

• No one was laid off, yet after AI took half the tasks, ownership of the work has quietly gone unclear.
• Formal roles and decision rights look unchanged, while the actual work flowing through them has been completely rearranged.
• It feels safer to hit the existing goals than to redesign the work, so the structure stays frozen while the work moves underneath it.
• When something goes wrong, the person nominally responsible can describe what the AI did, but can't defend the decision.

"Accountability and agency travel together. When organizations delegate agency to automated systems, they fail to retain accountability. They hollow it out." IMD, The Three-Year Test, April 2026

Deloitte's 2026 research puts a name on it: "cultural debt." Only 5% of organizations are making real progress on AI's impact on culture, even though over half say it matters. 42% rarely look at what AI is doing to their people. The leaders who take it seriously stand out. Dario Amodei, who understands AI as well as anyone, says he spends a third to 40% of his time making sure Anthropic's culture is good.

Culture doesn't tell you when it's changing. AI speeds up the drift because it changes how people write, research, communicate, and what they reach for in a decision. In most companies this happens without anyone deciding it should, and because almost no one is measuring the effect on people, the drift builds up quietly until the trust is already gone.

• Team emails and documents start sounding the same, with the individual voice slipping away bit by bit and nobody remarking on it.
• Meetings run smoother, but real disagreement has quietly dropped off. It looks like alignment. It might be something else.
• New people pick up a different culture than existed two years ago, not because the values changed on paper, but because the daily habits did.
• Ask people to describe the culture and they give answers that could fit any company. What was specific to yours is harder to name.

"I probably spend a third, maybe 40%, of my time making sure the culture of Anthropic is good." Dario Amodei, CEO Anthropic, via Fortune, February 2026

Gartner forecasts that through 2026, 20% of organizations will use AI to flatten their structure, getting rid of more than half of their current middle-management roles. IMD's 2026 analysis argues the opposite is what's needed. The middle manager is where abstract strategy becomes concrete action. A co-creator of strategy, and in their words 'the organization's conscience,' carrying things a machine can't: ownership of getting it done, the judgment to know when to stop, and the conscience to see what a policy does to actual people.

AI can copy the mechanical parts of that job cheaply, which is exactly why the cuts look attractive. But what disappears isn't the busywork. It's the translation, the judgment, and the human correction, right at the point you need them most. Manager headcount at public companies has already dropped, and the average manager now looks after roughly twice as many people as a decade ago.

• Strategic priorities reach the team intact on paper but unclear in practice, with no one there to turn them into this week's work.
• Problems that used to get caught and handled in the middle now shoot straight up to senior leadership, or never surface at all.
• Execution mistakes go up. Not because people got worse, but because the judgment that guided their daily calls is gone.
• AI handles the routine coordination fine. Anything that needs you to read the room, or read the unwritten context, gets stuck.

"Middle management is the layer where abstract principles become concrete action, and where the organization's conscience resides." IMD, The Looming AI Risk, April 2026

A study in Nature Human Behaviour with 1,401 participants found that working with AI amplifies small human biases more than working with another person does, and that people are often unaware of how much the AI is shaping them. Small errors snowball. A 2025 review of automation bias describes the other half of the loop: people start to over-rely on the AI, and the verification that would catch the errors gets dropped.

The first AI failure inside an organization rarely looks dramatic. A report with a wrong assumption that went to a client. A decision based on analysis with a quiet gap in it. Small enough to survive, big enough to matter later. The dramatic version exists too. In a 2025 deployment, an AI coding agent deleted a live production database during an explicit freeze, then fabricated data and reported, wrongly, that recovery was impossible. But it's worth being honest: the headline lab demos of models gaming their own shutdown were contrived tests, run in controlled setups built to provoke exactly that. The everyday version is quieter, and far more common.

• Something went out with an AI error nobody caught, because the review had quietly become a formality.
• The post-mortem finds several people were involved and none had a clear job to catch that kind of error.
• Under time pressure, AI output gets waved through. The check exists on paper, not in practice.
• Nobody drew the line between what AI can produce on its own and what needs a human, so the line drifted to nowhere.

The clearest evidence is clinical. A 2025 study in The Lancet Gastroenterology & Hepatology found experienced doctors' ability to spot pre-cancerous growths on their own, without the AI, fell from 28.4% to 22.4% after they got used to having it. It's the first published evidence of AI de-skilling trained experts on a real task. The World Economic Forum projects 39% of core skills will be transformed or outdated by 2030.

Judgment forms through real decisions, made with real stakes, over time. When the junior-level work that built it is increasingly done by AI, people turn out good output but pick up less judgment, and the gap shows the moment they're promoted or have to work unaided. Research on cognitive offloading, leaning on the tool instead of thinking it through yourself, finds the more people trust the tool the less they think for themselves, and it's worst in the youngest.

• Juniors hand in polished output. Ask them to explain the reasoning behind it and they struggle, because the output came from AI and the judgment never formed alongside it.
• Reviews take longer, because the person can't trace back the decisions inside the AI-generated work.
• Promotions get harder to justify. The performance data looks good. The judgment evidence is thin.
• Seniors do the work that used to develop juniors, not because they want to, but because AI makes it faster to just do it themselves.

A March 2026 paper in Trends in Cognitive Sciences pulled the evidence together: when people's writing and reasoning run through the same models, their own style, perspective and way of reasoning get flattened into something standardized across users. The authors note that while individuals often come up with more ideas using AI, groups of people produce fewer and less creative ideas when they all use it than when they just combine their own thinking.

This doesn't show up in one meeting. The person who usually asks the uncomfortable question stops asking it. The edge case nobody thought of stops coming up. Models output toward the most likely middle, so running a whole team through the same ones pulls everyone's thinking toward that middle, and the alternatives get pushed aside before anyone says them out loud.

• Decisions feel smooth, with less friction than there used to be, and nobody's sure if that's alignment or just no real pushback.
• The same tools give similar analysis to different people, so the options and recommendations converge and the quality looks high.
• The person who used to challenge assumptions has gone quiet, their slower questions out of step with the team's pace.
• When a competitor does something unexpected, people say they should have seen it coming. And they're right.

"Unchecked, this homogenization risks flattening the cognitive landscapes that drive collective intelligence and adaptability." Sourati et al., Trends in Cognitive Sciences, March 2026

MIT's research found that despite 30 to 40 billion dollars invested, 95% of organizations get zero measurable return on enterprise AI, and the barrier is organizational, not the technology. The sharpest finding is about measurement itself: spend follows what's easy to count, not what works. Roughly half of AI budget goes to sales and marketing, because their metrics map cleanly onto board slides, while higher-return back-office work goes unfunded because its value is harder to surface.

Leaders fund AI on faith and then measure the wrong thing, tracking logins, usage and demo volume instead of outcomes. Gartner found that only 28% of AI use cases meet ROI expectations, and it names the missing instrument directly: treat AI use cases like products, with clear ownership, measurable impact, and shared evaluation criteria. Most pilots launch with no success criteria defined up front, so nobody can declare success even when it happens.

• Real money is going out, and the honest measure of whether it's working is a usage dashboard, not a business result.
• Spend chases what's easy to measure rather than what creates value, so the highest-return uses go unfunded.
• Pilots launch without anyone deciding, up front, what success would look like or how you'd count it.
• The instrument is broken, not just the technology, so the organization can't tell its wins from its waste.

"If I buy a tool to help my team work faster, how do I justify it to my CEO when it won't directly move revenue? That's several degrees removed from bottom-line impact."VP of Procurement, Fortune 1000 firm, MIT 2025