It’s Time We Stop Being News Organizations

McKinsey’s “skill change” and “exposure” framing is useful as far as it goes. It tells you where pressure will land first. It describes the weather. It does not tell you how to build a ship. The failure is not in the analytics. The failure is in what enterprises do with the analytics. They treat it as content. They treat it as narrative. They treat it as news. In doing so, they avoid the only question that matters. Who, or what, has the legitimate right to act on the signal in a way that changes the outcome.

Exhibit. “Skill Change Index assesses how automation exposure varies across skills.” Vertical axis reflects relative exposure to automation. Horizontal axis ranks skills by percentile. Individual points represent skills, color-coded by category. The figure is optimized for describing

distribution of exposure across work, not for locating where outcome-shaping authority should reside. The answer to that question defines whether you are building an agent or a toy. If it cannot shape an outcome, it is not an agent. The newsroom inside the enterprise News organizations are optimized for describing reality. They compete on speed of interpretation, crispness of framing, and frequency of updates. They do not carry the obligation of making the described reality better. They do not own the bill when the description changes nothing. Inside most enterprises, the same pattern now runs through operating reviews, strategy sessions, and AI discussions. Charts move faster than decisions. Updates arrive faster than authority. The organization knows more and controls less. Smart people become fluent narrators of drift. This is not an insult. It is a diagnosis. The drift is understandable because narration is safe and action is costly. Narration can be justified. Action can be blamed. Narration can be shared without consequence. Action changes the world, then demands that someone own what changed. That is why the AI era is not simply a technical change. It is a legitimacy crisis. AI collapses the cost of execution, then exposes the part of the enterprise that was already fragile. Decision rights, permission boundaries, escalation logic, and accountability. The modern enterprise has learned to treat those as culture topics or governance topics, which means they become committee topics, which means they become calendar topics. They rarely become architecture topics. Architecture is where controllability lives. Without controllability, your “AI strategy” becomes a faster form of narration. If you think that sounds harsh, consider what most of the market is calling agents today. A system that drafts an email, answers a question, routes a ticket, writes SQL, or summarizes a report is useful. It may cut minutes. It may reduce keystrokes. It may create a feeling of progress. But unless it can act in a bounded way to change system state and close a loop to an outcome, it is not an agent. It is a tool. Tools are not new. Tools do not reorder responsibility. Agency does. McKinsey is measuring exposure. We have to measure responsibility. A skill exposure chart is built on a sensible observation. Some tasks can be absorbed by software. Some tasks are harder to absorb. Some human activities appear safer. In a labor market context, that helps people understand where pressure may show up.

Enterprises are not paid for understanding pressure. Enterprises are paid for producing outcomes. Outcomes are shaped by decisions that alter the future state of the system. Those decisions are not evenly distributed. They sit at specific nodes where a choice changes throughput, quality, service, risk, cash conversion, and option value. The organization that wins in the AI era will not be the one that cataloged skills correctly. It will be the one that placed authority correctly. That is the first correction to the mainstream conversation. Skill is a property of a person. Advantage is a property of a system. The second correction is more severe. Many things that are hard to automate are still low value. They are hard to automate because they are poorly structured, politically protected, or entangled with weak accountability. Corporate bureaucracy fits that description. So do many coordination rituals that exist primarily because the underlying system cannot move information and decisions with integrity. If you treat “hard to automate” as “high value,” you will preserve the wrong work. You will defend meeting load as leadership. You will defend approval chains as safety. You will defend process as governance. You will defend reporting as control. Then you will wonder why the organization knows more and still cannot act. What an agent is, and why that definition is a knife Your definition is the only one that matters because it is operational and falsifiable. An agent must be able to shape an outcome. Otherwise, it is not an agent. That sentence looks simple until you follow it into the places where enterprises actually live. “Shape an outcome” is not the same as “produce an output.” It is not the same as “recommend an action.” It is not the same as “complete a workflow.” It means the system has the legitimate ability to act in the world of the enterprise, not just in the world of text. That immediately forces a second question. What makes that ability legitimate. Legitimacy is not a moral add on. It is not branding. It is what separates agency from chaos. Legitimacy requires bounded authority, explicit permission, and traceable accountability. The agent must be allowed to act inside a defined boundary. Someone must be able to say what that boundary is, why it exists, and what happens when the boundary is reached. Without that, you do not have agency. You have a system that may act, but has no rightful place to act. That is the third question. How do we know the action was justified. This is where most agent talk collapses into theater. A system that acts without producing a governable justification record will not be trusted for the decisions that matter, and it should not be. The justification record is not the same as exposing raw internal chain of thought. It is the

operational artifact that ties reasoning to legitimacy. It is what allows audit. It is what allows escalation. It is what allows a board, a regulator, a customer, or a plant manager to ask, why did the system do that, and get an answer that can be tested. An agent, in the enterprise sense, must therefore meet three requirements that cannot be faked. It must be able to shape an outcome. It must have legitimate authority to act within a boundary. It must produce governable justification tied to evidence, assumptions, constraints, and thresholds. If any one of these is missing, the system is not an agent. It is assistance. It can still be valuable, but it is not the architectural reorder people are claiming. Legitimacy is not branding. It is bounded authority plus accountability. You cannot have an agent without causal reasoning Now we reach the line you asked to be made explicit. You cannot have an agent without causal. The reason is plain. Outcome shaping is intervention. Intervention changes the world. A system that cannot reason about intervention cannot be trusted to intervene. Pattern prediction can tell you what often follows what. It can tell you what looks like success. It can tell you what resembles past instances. That is correlation. Correlation can advise. It cannot govern. Causal reasoning asks a different question. If we do this, what will it cause. If we change this constraint, what shifts downstream. If we remove this bottleneck, what new failure mode appears. If we grant authority at the edge, what new risk becomes possible. If we tighten escalation, what latency do we buy and what volatility do we avoid. Those are causal questions. They are also the only questions an agent must be able to handle if it is to act in a way that shapes an outcome, rather than merely generating a plausible response. This is where the mainstream “agent” narrative becomes dangerous. If an enterprise builds so called agents on pattern matching, then hands them authority, it will experience a phase of superficial success followed by sharp failure when novelty enters. Novelty is not rare. It is the natural state of operations. A supplier misses a shipment. A constraint moves. A rule conflicts with another rule. A new product variant introduces a new defect mode. A customer changes a tolerance. A competitor changes terms. A weather event alters inbound flow. The world refuses to be a dataset. An agent that cannot do causal reasoning will act confidently into that world anyway, because it will not know what it does not know. It will not distinguish the symptom from the mechanism. It will not know when to stop. It will not know when to escalate. It will not know when to ask for human authority.

That is not agency. That is acceleration of error. The strictness of the definition matters because it prevents category collapse. We stop calling every automation an agent. We stop calling every recommendation autonomy. We stop calling every draft a decision. We stop acting like a newsroom that praises the story, then pays for the mistake. The missing link is reasoning tied to legitimacy You asked for “chain of thought reasoning tied to legitimacy.” The point is right, and it needs to be expressed in enterprise language that a board can use. A legitimate agent must be governable. Governance requires reasons that can be examined. That does not mean exposing every internal token, every intermediate thought, every private scratch line. It means producing a decision record that can be audited without mysticism. An agent’s justification record should, at minimum, make explicit the outcome it is trying to shape, the evidence it used, the assumptions it made, the constraints it respected, the alternatives it considered, the rationale for the chosen action, and the thresholds that would trigger escalation or stop. Those elements are not academic. They are the mechanism by which legitimacy is preserved. They are also the mechanism by which enterprise trust becomes possible at scale. In practice, this is where “agents” become real. Not when they write better text. When they can act within a boundary and leave behind a trail that allows governance to remain intact. A system that can act without a governable justification record is not an agent you can trust. It is a liability that will be defended by saying it was “just following the model.” That is not a defense. That is abdication. Without causal reasoning and governable justification, autonomy becomes un-auditable risk. Perspective. The move from insight to cause At this point, the word perspective often gets used as decoration. That is not what is meant here. Perspective is the ability to see what is producing what. It is the ability to look at a metric and ask what system state created it. It is the ability to look at a problem and ask what boundary condition makes it repeat. It is the ability to look at a “skill exposure” chart and ask what it implies about where authority should live. Insight without perspective produces commentary. Commentary is cheap. It also becomes addictive because it feels productive without carrying risk.

Perspective produces reordering. Reordering is costly. It creates friction in the organization because it touches permission and status. It forces decisions that remove comfort. It also produces controllability, which is the only thing that matters when volatility arrives. This is why “stop being news organizations” is not a slogan. It is an operating requirement. A newsroom reads signals and publishes interpretation. An enterprise must read signals and act. The cost of not acting is not theoretical. It is time. It is variance. It is cash. It is reputation. It is option value that decays while the organization holds meetings about it. The work that matters is not “less exposed.” It is more responsible. McKinsey is right to say that routine execution work is increasingly absorbable by software. That is the easy part. Every executive can nod at that. The hard part is what that reality does to the enterprise’s internal economy of status. When execution becomes cheap, the organization must move human authority away from execution nodes and toward responsibility nodes. That includes the design of permission. The design of incentives. The design of escalation. The design of risk boundaries. The design of what evidence is required for what kind of action. The design of how conflicts are resolved when objectives collide. Those are not soft topics. They are cash topics. They determine whether the enterprise can act at the speed of the signal, or whether it will remain stuck in a world where every signal becomes a meeting, every meeting becomes a deck, and every deck becomes a story the organization tells itself to explain why it did not act. The irony is that many organizations treat governance as the reason they cannot act. They describe governance as delay. They describe governance as diligence. They describe governance as safety. Each word sounds responsible. What is actually happening is that the enterprise has converted fear of consequence into permanent architecture. It is paying for time as if time were free.

Exhibit. “Decision Leverage Index reframes automation from skill exposure to outcome leverage.” Vertical axis reflects Outcome Leverage. Horizontal axis reflects Execution Automation Potential. Bubble size indicates relative capital impact. The “Legitimate Authority Layer” calls out that legitimate agents require outcomes, causal reasoning, bounded permission, escalation rules, and governable reasoning, and that the highest leverage work is the architecture of those constraints rather than the execution of tasks. A counterargument that must be faced There is a serious counterargument, and it deserves more than a dismissal. Some leaders will say that what is being described as “news behavior” is simply prudent risk management. They will say that speed creates errors, errors create harm, and harm creates legal and reputational exposure. They will say that strong governance must slow action, not accelerate it. That argument is not wrong. It is incomplete. The choice is not between speed and governance. The choice is between governance as ritual and governance as architecture. Ritual governance produces delay without clarity. Architectural governance produces bounded autonomy with audit. The reason speed becomes dangerous is not speed itself. Speed becomes dangerous when authority is unbounded and reasoning is un-auditable. If you build agents with causal reasoning, bounded permission, and governable justification, you can increase speed while reducing risk because you reduce random action and increase traceable action.

In other words, legitimacy is what makes speed safe. This is the crucial point. The enterprise is not being asked to become reckless. It is being asked to stop confusing delay with safety and meetings with control. A world class operation has never been slow because it is safe. It is safe because its controls are designed, tested, and enforced. That is as true in decision systems as it is in physical systems. Two diagnostic paragraphs a board can read aloud If you want to know whether your firm is acting like a newsroom, ask this in an executive meeting and do not rescue anyone from the answer. When a meaningful operational signal appears, who has the legitimate right to act, and what is the maximum time allowed between signal and intervention before the firm pays for delay in cash, customer trust, or risk exposure? Ask a second question that is more precise. In your current operating model, how many decisions require a meeting because nobody owns bounded authority, and how many meetings exist primarily to manufacture legitimacy after the fact because the system cannot produce a governable decision record at the moment of action? If those questions feel politically expensive, that is because they touch the real structure of the organization. That is also why they are useful. A falsifiable prediction Here is a prediction that should be embarrassing if wrong. Over the next cycle of enterprise AI deployments, the firms that treat “agents” as tools that draft and recommend will report high activity and low outcome change, while the firms that define agents as legitimate outcome shapers with causal reasoning, bounded permission, and governable justification will show measurable reduction in decision latency and faster recovery from operational exceptions. This is falsifiable. It will show up in the record. It will show up in cycle times, exception resolution times, and the frequency with which humans must intervene through meetings to keep the system coherent. If the prediction fails, it means one of two things. Either agents can create controllability without causality and legitimacy, which would rewrite what we know about governance. Or enterprises are so politically constrained that they will not grant bounded authority even when the architecture exists. Both outcomes are worth confronting because they define where real constraint lives. Where the McKinsey framing can be improved The mainstream “exposure” framing tends to treat leadership and management as safer categories. It assumes that because a behavior is interpersonal, it is protected.

That is not what will happen. AI will remove the need for human routing and human status narration. Any role that exists primarily to move information between layers will be absorbed. That includes many activities currently labeled as management. What remains is not leadership as theater. What remains is leadership as architecture. Leadership becomes the craft of boundary setting and consequence ownership. It becomes the craft of designing incentives that cause behavior without constant supervision. It becomes the craft of deciding what evidence is required for what action, then building systems that enforce that rule without turning every day into a courtroom. If leadership does not become that, it becomes commentary, and commentary will be automated too. Automation removes human routing. Agency remains where accountability cannot be outsourced. Why causal is the dividing line between “assistant” and “agent” A system that cannot do causal reasoning can still be useful. It can draft and summarize. It can reduce cognitive load. It can support a human operator. It can make execution cheaper. But it cannot be trusted to shape outcomes because it cannot distinguish intervention from association. The minute you hand such a system authority, it becomes a generator of plausible mistakes. Those mistakes will not be random. They will be systematic. They will be the consequence of treating patterns as mechanisms. Enterprises already suffer from this error in human form. They confuse symptoms for causes. They copy the visible traits of high performers and expect the outcome to follow. They copy dashboards and expect control to follow. They copy processes and expect quality to follow. They adopt the appearance of capability and miss the structure that produced it. Causal reasoning is what stops that. Causal reasoning forces the system to ask what produces what. It forces explicit assumptions. It forces counterfactual thinking, not as philosophy, but as an operational requirement. What happens if we intervene. What happens if we do not. What changes downstream. What new risk becomes possible. An agent without causal is a tool that sounds confident. A legitimate agent is a system that can act, explain why, and be governed.

The end of the newsroom habit

If you accept this model, the phrase “stop being news organizations” becomes sharp. It means you stop treating AI as content. You stop treating charts as control. You stop treating awareness as governance. It means you make the hard trade. You convert governance from ritual to architecture. You move authority closer to the edge where time is being spent, but you do it with boundaries that preserve legitimacy. You demand causality where intervention is required. You demand justification where accountability must hold. You design agents as legitimate outcome shapers, not as clever assistants. This is not a technology bet. It is a seriousness test. The firms that pass will not be the ones with the best dashboards. They will be the ones whose decision systems behave like engineered systems. Signal produces action. Action is bounded. Reason is recorded. Escalation is explicit. Outcomes are measured. The loop closes. The firms that fail will not fail because they did not adopt AI. They will fail because they adopted it the way newsrooms adopt stories, then called that leadership. The future will be built by organizations that can see cause, then act with legitimacy. The rest will keep publishing headlines inside their own walls. References This article draws on the McKinsey Global Institute’s late 2025 work on skills and automation exposure as a descriptive baseline, then tests it against the stricter requirement of agency as legitimate outcome shaping, which comes from Michael Carroll’s teaching and writing on agency, decision latency versus permission latency, permission as the architecture of risk, and causal reasoning as the difference between symptom tracking and controllable systems, as well as foundational ballast from W. Edwards Deming’s Out of the Crisis in 1986 on management systems and variance, Herbert Simon’s work on bounded rationality and organizational decision limits, Judea Pearl’s causal framework on intervention and counterfactual reasoning, James March on decisions and organizational behavior, and the practical governance literature that ties accountability, audit, and authority to real operational and financial consequence.