Palantir ставит на то, что искусственный интеллект позволит компаниям сократить цикл принятия решений с дней до секунд

In 1924, General Motors had a problem that would sound almost absurd to a CEO today: it didn't really know what was happening in its own business. Cars could be piling up on dealer lots while executives in Detroit were still working from weeks-old information. Alfred Sloan, who led GM from 1923 to 1946, later recalled that GM sometimes “knew nothing about the most recent five or six weeks” of its car sales. By the time those executives understood what customers were actually buying, the market had already moved. To attack that problem, GM created a system of making information move faster. Its automobile divisions began requiring reports from dealers every ten days: new orders, orders waiting to be filled, and new and used cars sitting on lots. Executives could see demand sooner. Production could adjust faster. Decisions that had been made through weeks of uncertainty could now be grounded in information merely days old. Ten days. That was a breakthrough. But here is the part of the GM story that relates to today. When information is slow, organizations need infrastructure to move it. Infrastructure meaning they needed people to collect it, people to reconcile it, people to turn it into reports, managers to interpret those reports, and additional layers to move information upward and decisions back down. Over time, much of that infrastructure became what we simply call management. That doesn't mean management exists only to move information. Managers develop people, resolve conflict, exercise judgment, build trust and make decisions under uncertainty. But a meaningful share of the corporate hierarchy we inherited was built around a practical problem: the people making decisions could not see what was happening fast enough. Today, that problem is disappearing. Information that once took weeks, and later days, to reach decision-makers can increasingly be reconciled, analyzed, and acted upon in seconds. If companies no longer need as many people and layers simply to collect information, interpret it, move it upward and send decisions back down, what happens to the organization built around doing exactly that? That is not primarily a jobs question. It is an organizational-design question. If AI changes how information moves and decisions get made, the issue isn't simply how many people a company needs. It is how the company itself should be built. Now put that problem inside a company today, a century later. Monday, 8:02 a.m. A divisional executive opens the 47-slide operating deck her team spent Friday assembling. There is only one problem: the system already knew everything in it. Before anyone opened PowerPoint, it had reconciled the weekend's results against inventory, flagged three anomalies, traced one to staffing and another to a supplier, modeled possible responses, executed the moves it was authorized to make, and put two decisions in front of her that actually required a human. Her meeting is shorter. Her team is smaller. The company is faster. So what's the problem? She hasn't lost her job. She's lost the work that taught her how to do it. She learned the business by building those decks. She learned which numbers mattered, which explanations didn’t hold up, where operating problems hid, and which questions senior executives asked when the numbers went wrong. She was even promoted three times because she became exceptionally good at turning information into judgment. Now, information arrives already assembled, analyzed, and increasingly accompanied by a recommendation. That leaves three questions that are much harder than whether AI will eliminate jobs. If management layers were built partly to solve an information problem, what happens when machines solve much of that problem differently? 1. If AI absorbs the work through which future managers learned, where will the next generation develop judgment? 2. And, if AI radically lowers the cost of coordinating information, people and decisions, how much of the corporation we inherited still needs to exist at all? 3. Most of the AI-and-work debate starts with the worker: Which jobs can a machine perform? I think that's the smaller question. The bigger one starts with the organization: Which parts of the company existed because humans couldn’t process information, coordinate activity, and make decisions efficiently enough without them? This is where Palantir becomes worth watching, and not for the reason it usually gets covered. OpenAI, Anthropic, and Google compete over intelligence. Palantir occupies different ground: organizational context, then decision, then action. Its Foundry, AIP, and Apollo stack is marketed as an enterprise operating system, but the load-bearing concept is the Ontology — an attempt to represent an organization's objects, relationships, logic, permissions, and actions in a form both humans and machines can operate against. For decades, enterprise software helped organizations record themselves. ERP recorded what a company owned. CRM recorded what customers did. Business intelligence explained what had happened. Generative AI answered questions people already knew to ask. Palantir is attempting something structurally different: to make enough of the organization computationally understandable that humans and machines can act against the same operational reality. Which produces the thesis worth arguing about. Palantir’s most consequential bet may not be that AI can replace workers. It is that AI can absorb enough information processing, coordination, and routine decision-making to change how much traditional management an organization needs. So, AI may flatten the corporation before it empties it. Nobel Prize-winning economist Ronald Coase famously asked why firms exist at all. One answer: some activities are cheaper to coordinate inside a company than through the market. If AI radically lowers the cost of coordination, it could eventually change not just how many management layers a company needs, but how much of the company needs to remain inside the company at all. That is a larger question than this piece can settle. The disruption inside the org chart is already visible. AI Is Coming for the Org Chart Two bodies of evidence are pointing in what look like opposite directions. Read together, they describe something more unsettling than either does alone. At the bottom, two large datasets point in a similar direction. Stanford's Digital Economy Lab, working with ADP payroll data covering millions of workers, finds employment among 22-to-25-year-olds in AI-exposed occupations running roughly 19% below where it would be had it tracked less-exposed peers — with no comparable gap for experienced workers. Separate résumé-based research covering tens of millions of workers finds junior employment declining at firms that adopt generative AI while senior employment holds steady. Both sets of authors are careful to describe these as patterns rather than causal estimates. The mechanism matters more than the magnitude. In both cases the adjustment runs primarily through reduced hiring rather than separations. Companies are not firing their apprentices. They are quietly no longer creating apprenticeships. That distinction is the whole argument, because entry-level work was never only production. Building the model, running the analysis, drafting the document, preparing the recommendation, assembling the operating review — these were the mechanisms through which people learned how an institution actually worked. The output was the visible product. The education was the point. In the middle, the picture is noisier but harder to dismiss. An analysis Revelio Labs conducted for Business Insider found that, at the white-collar hiring trough, openings were down 43% for middle managers and 57% for senior leaders against the 2022 boom — compared with 14% for junior roles. Gartner, in strategic predictions issued in late 2024, forecast that through 2026 one in five organizations would use AI to flatten their structures, eliminating more than half of current middle-management positions. That forecast window closes this year. The prediction is worth treating as a test, not a finding. The corporate examples are not subtle: Citi targeted a reduction from thirteen management layers to eight, and UPS eliminated roughly 12,000 positions during 2024. Amazon separately moved in 2024 to raise its ratio of individual contributors to managers by at least 15%, and announced roughly 14,000 corporate job cuts the following year. An honest caveat belongs here. Not one of those decisions can be cleanly attributed to artificial intelligence. Post-pandemic overhiring, interest rates, and ordinary cost discipline explain a great deal of it. The defensible claim is narrower and, I think, more interesting: AI is arriving precisely when corporations were already questioning how much hierarchy they need. It has the potential to turn a cost-cutting fashion into a structural redesign. The Rise of the Hollow Company For most of the last century, companies developed executives the same way: junior worker, experienced contributor, manager, senior leader. Each stage taught something the next one required. Junior work was apprenticeship disguised as production. Management was judgment training disguised as coordination. AI now exerts pressure on both stages at once. At the bottom, it absorbs the production work that doubled as apprenticeship. In the middle, it absorbs the information movement and coordination that justified the managerial layer. The near-term economics look excellent. Productivity rises. Layers fall. Spans widen. SG&A drops. Decisions accelerate. Revenue per employee improves. Every one of those numbers will look good in a board deck. What they will not show is the thing being consumed to produce them. The Hollow Company: an organization with sophisticated executives at the top and extraordinary intelligence throughout, but a steadily weakening mechanism for producing the humans capable of leading it next. Companies can improve today's margins while liquidating tomorrow's leadership bench. The company eliminating managers in 2026 still needs leaders in 2030, and it may have spent four years dismantling the developmental system it took for granted. There is a second-order problem layered on top, and it explains why this is so hard to fix from inside. The managers AI makes less necessary in their current form are the same managers companies need in order to implement it successfully. Their hesitation is not technophobia. It is arithmetic. Asking a director to deploy a system that eliminates information bottlenecks, widens spans, and pushes authority toward the front line is asking her to automate the specific activities that justify her position. Rational people move slowly under those conditions. That is a leadership problem wearing an IT problem's clothing, and it is why organizational transformation will prove harder than technological implementation at most companies. What Palantir Is Really Betting On Palantir's durable advantage may not be its AI at all. The company allows customers to run multiple underlying models and bring their own. That is a revealing design choice. It suggests the scarce asset isn't the intelligence — it's the organizational context surrounding the intelligence. Intelligence, then context, then decision, then action, then feedback. The strategic question is not who has the smartest model. Who owns the layer through which intelligence becomes organizational action? Customers appear increasingly willing to pay for that proposition. Palantir's second quarter of 2026 produced $1.94 billion in revenue, up 93% year over year, with U.S. commercial revenue up 149% to $764 million and remaining U.S. commercial deal value more than doubling to $6.24 billion. Valuation is a separate argument. Revenue at that trajectory is difficult to dismiss as experimental spending. None of which means Palantir has cornered anything. Microsoft has distribution no one can match and is pushing agents into the surface where work already happens. ServiceNow occupies workflow territory that matters enormously in the agentic era. Salesforce, SAP, and the hyperscalers are all moving toward related ground. Everyone will have agents. The harder question is who can build a trustworthy operational representation of how a consequential organization actually runs and then safely let machines act inside it. Palantir spent two decades working on versions of that problem across defense, intelligence, manufacturing, healthcare, and supply chains — environments where fragmented information and consequential decisions are the normal condition. The skeptic's case deserves its paragraph. Is the Ontology a moat or an extraordinary switching cost dressed as one? How much of the value depends on Palantir's forward-deployed engineers rather than the software? And what governance risk is created when an increasing share of a corporation's consequential decisions flows through a single computational layer? But there may be a deeper coherence here worth naming. Palantir's products repeatedly attack the same target: fragmentation, delay, bureaucratic handoffs, and the distance between knowing something and acting on it. That suggests the company did not merely build enterprise software. It built technology around a theory of institutions — that organizations become weaker when information is fragmented, authority is diffuse, and execution is slow. Read that way, Palantir may not be betting against managers at all. It may be betting against the organizational friction that made so much management necessary in the first place. The Apprenticeship AI Could Rebuild Here is the part I find genuinely surprising. If leadership development can no longer run through production, the answer cannot be preserving obsolete tasks because previous generations happened to learn from them. Nobody should build a financial model by hand in 2030 for pedagogical reasons. Development has to move from production apprenticeship to judgment apprenticeship — learning by interrogating the machine's recommendation, stress-testing its assumptions, identifying the context it lacks, comparing alternatives it discarded, studying exceptions, deciding when to override it, and owning the outcome. And judgment apprenticeship requires something organizations have never really possessed: a record of judgment. An operating layer that captures what the organization knew, what was recommended, which alternatives existed, what the human chose, what action followed, and what actually happened is producing a longitudinal dataset of institutional decision-making. Companies have always had this in fragments — in memory, in war stories, in the intuition of people who were there. The most valuable corporate training database of the AI era may not contain courses. It may continue decisions. Palantir's own documentation goes partway there. It describes an Ontology that captures decision lineage — when a decision was made, against which version of the enterprise data, through which application — and discusses aggregated decisions becoming material that improves the performance of both people and agents over time. What remains an inference is the human half: that such a record could become the structured basis for developing executive judgment, not only for tuning the system. That is an implication of the architecture, not a product being sold. But it is the most interesting thing the architecture makes imaginable: the infrastructure disrupting the old apprenticeship could also become unusually powerful infrastructure for building a better one. It also has a dark twin, and it needs to be named before someone else names it. A permanent record of organizational judgment is also the most sophisticated management surveillance system a corporation has ever had. Reconstruct what a manager knew, when she knew it, what the machine advised, what she chose instead, and what happened — and you have built a performance review with perfect hindsight attached. The pathology writes itself. A manager who knows every deviation from the machine's recommendation will be permanently reviewable does not make better decisions. She makes more defensible ones. The organization slowly optimizes for auditability and calls it discipline. Decision lineage should develop judgment, not police it. That requires actual rules — who can access the record, whether it enters performance evaluation, how hindsight bias is controlled, and whether managers retain protected space to disagree with a machine and be wrong. Companies that deploy the capability without answering those questions will get the surveillance system and lose the learning system. Meanwhile, the job itself changes shape. As AI absorbs reporting, monitoring, information movement, routine coordination, and predictable escalation, what remains is judgment, ambiguity, trust, persuasion, conflict, culture, ethics, motivation, talent development, and accountability. The more management AI can perform, the more human the remaining job of the manager becomes. Management doesn't disappear. Mediocre management becomes much harder to justify. That is AI's organizational last mile. Palantir may be building an operating system for the AI-native enterprise. But software cannot renegotiate authority, redesign careers, create trust, develop judgment, or decide what humans should remain accountable for. Somebody still has to build the human operating system around it — decision rights, incentives, spans, learning, accountability. Technology redesigns workflows faster than institutions can renegotiate any of that. Build the first without the second and you have simply automated yesterday's bureaucracy at considerable expense. Diagnosis is cheap. Here is what to measure before the next board meeting. Information time. Audit one representative week of your senior managers' calendars. What percentage went to obtaining, reconciling, packaging, and transmitting information rather than exercising judgment? Decision distance. Take one consequential operational decision and trace it from the first observable frontline signal to authorized action. Count the humans it passed through. Not the org chart — the actual path. Reporting latency. Take your last meaningful negative surprise. Find the date the underlying data made it knowable. Compare it to the date someone empowered to act recognized it. Then ask the question that should keep you up: how long did your last bad quarter sit inside your systems before anyone realized you had a problem? A century ago, ten days was revolutionary. What is your number? Human decision rights. List the ten highest-consequence recurring decisions in the enterprise. Classify each: AI recommends, AI executes after approval, AI executes autonomously, human decision required. If your leadership team cannot complete that exercise before agentic systems arrive, governance is already behind implementation. Monday, 8:03 a.m. Return to the executive we discussed earlier. She learned the business by building the deck. Her younger colleague never will. He is sitting beside her, not assembling information but interrogating a recommendation. He thinks one of its assumptions is wrong. She asks why. He explains. They argue about it for four minutes. She decides. He watches what happens next, and so does the system. She learned management by producing information. He is learning it by challenging intelligence. That is production apprenticeship becoming judgment apprenticeship, and it is probably better training than she got. It is also training for fewer people. Twelve people once participated, badly, in preparing that operating review. Two now participate deeply in the decision. Better apprenticeship for a narrower funnel is not a solved problem. It is a different one, and it belongs on the agenda of every CEO who is about to celebrate a flatter org chart. Across town, a competitor is opening slide one. General Motors transformed management by compressing an information gap measured in weeks toward one measured in days. Palantir is betting that AI can push meaningful parts of that gap toward seconds. The company that collapses a seven-day reporting cycle to seven seconds is not 30% more productive. It is operating on a different clock. By the time that advantage shows up in a competitor's margins, you may already be two organizational redesigns behind. Palantir's bet against management may ultimately be something more ambitious than a bet against managers: a bet that companies can finally separate leadership from bureaucracy. For a century, we paid for management partly because information was scarce, fragmented, and slow. If that stops being true, the question is not whether management survives AI. It is what management is for when information no longer needs a manager to move it.