Integration Capacity Analysis
integrationcapacity.org

Sovereign Alignment

Why AI gives organizations so little, and why the answer is alignment, not more capacity

Wim Van Laere  ·  August 2026

Summary

Organizations everywhere see the same thing. Each employee works faster with AI. The organization as a whole barely improves.

This paper explains why, and what to do about it. The explanation rests on one distinction. Capacity is how much a system can do. Direction, or orientation, is what it works toward. The two are independent — it is the old duality of efficiency and effectiveness. AI adds capacity only. It has no direction of its own. It follows the direction of whoever uses it. So when an organization adds AI without aligning direction, the extra output does not become value. It becomes duplicated work, material nobody reads, new bottlenecks — friction. The question this paper puts at the center is simple: as AI increases the capability of every part of the organization, does the organization still act as one?

The cure, therefore, is not more or less AI. The cure is radical alignment of orientation. Radical means: at the root and at every level. Every person, every team, every department pointed at the same declared purpose, and the distance between that purpose and actual behavior being measured all the time.

From this thesis, four requirements follow, and they organize the rest of the paper.

1. Alignment needs a reference at every level. Not only the organization, but each department, each team, and each coworker — internal or external — states its own charter. Equally important, the same gets done for each AI agent and each major system in the organization. Then everything can be measured twice: against its own charter, and against the level it contributes to. Drift becomes visible early, at the level where it starts, and can be corrected precisely.

2. Alignment includes the outside. Direction does not stop at the mission statement. It must hold against the value for customers, for society, and for the geopolitical and economic environment the organization operates in. This is how an organization can test whether its direction is still right — not by conviction, but by measurement against the levels above it.

3. Knowing how to use AI matters more than the AI tool itself. The same tool can push work toward the right goals or away from them; it can raise a person’s ability or lower it. The difference sits in the user: a clear goal, and a clear view of his own strengths and weaknesses. Organizations must recruit and train for this — it changes what a valuable profile looks like.

4. AI requires a culture of feedback and truth. AI is jagged — very good at some tasks, surprisingly bad at neighboring ones, without signaling the difference — and no feedback from the real world reaches it through the tool itself. So the organization must organize what the tool cannot give: feedback from reality, and the routine finding, checking, and evaluating of truth — about meaning, about alignment, about results. Without that culture, two documented drifts set in: each user is confirmed in his own view instead of corrected, and the thinking of all users grows more alike. Both remove the friction that catches errors.

The same holds for governments and states, which regulate AI while using it themselves. One word of caution: sovereignty here does not mean owning the technology or storing the data at home. It means keeping authority over your own direction. Sovereignty, in the age of AI, means two testable things: has the organization set an explicit — and as far as possible public — direction, and can it show that it still controls that direction in practice?

This is not only a diagnosis. An instrument exists that measures it — Integration Capacity Analysis, Section 3.3 — and its measurements are published.

The paper ends by separating what is proven, what is validated, and what is still theory.

1. The problem: the gains are real, the results are not

Two public debates about AI are running at the same time. They sound different. They are about the same problem.

The first debate is about money. In almost every organization, employees say AI makes their own work much faster. Yet the results of the organization barely move. This is now well documented. Atlassian surveyed more than twelve thousand knowledge workers and 173 Fortune 1000 executives in 2026. 89 percent of the executives said AI increases the speed of individual work. Only 6 percent were sure they could point to a clear organization-wide return. The same study put a price on the gap: fragmented, uncoordinated AI adoption costs the Fortune 500 an estimated 161 billion dollars per year. McKinsey surveyed more than ten thousand senior executives in fifteen countries in the same year. 88 percent of their organizations are experimenting with AI. Fewer than one in five has seen it meaningfully affect the bottom line.

The second debate is about control. Organizations and governments hand over more and more work and decisions to AI systems — mostly, as the adoption surveys consistently show, to raise productivity: the same work, faster and cheaper. But nobody checks whether the sum of all these handovers still follows the direction the organization — or the country — has chosen. That is the worry behind the European AI Act and similar laws: for high-risk systems, they require that a human can still see, understand, and if needed stop what the systems decide. This debate uses one word all the time and never defines it: sovereignty.

The claim of this paper: these are not two problems. It is one problem, seen at two levels. To show this, we need two concepts, and we need to keep them strictly apart.

2. The causes

The problem has two causes. The first is structural and sits in the organization: capacity and direction are different things, and AI adds only capacity. The second is human and sits in each single use of the tool: AI’s competence is jagged, the border of that competence is invisible, and the skill that decides the outcome — the user’s — is nowhere measured. A third condition makes both causes worse: the tool brings no feedback from reality. Together they produce three distinct forms of damage.

2.1 First cause: capacity without direction.

Capacity and direction are two different things. Capacity is how much a system can produce and process. Direction is what the system works toward: its declared purpose, or something else. The two are independent, and this fits the traditional duality of efficiency versus effectiveness. A system can have huge capacity and the wrong direction. A system can have modest capacity and the right direction. Raising the one does nothing for the other. This single distinction carries the whole paper. And it has one direct consequence: AI multiplies whatever organizational condition already exists. Misalignment becomes amplified friction; alignment becomes amplified capability.

AI is capacity without direction. AI multiplies what an organization can produce: drafts, analyses, reports, decisions. It also multiplies how much material flows between people, “agents” and teams. (An AI agent can be given goals. But those goals are assigned by someone. No AI system decides what the organization is for.) It adds nothing to the one thing that turns production and flows into results: a shared understanding of what all the activity is for. The framework behind this paper describes every organization on three strands: what it has and produces (the material strand, B), what its parts exchange with each other (the interaction strand, I), and what really keeps it coherent and together (the meaning strand, Φ). AI speeds up the first two strands. The third it can store, repeat, and even spread — but it cannot carry it. Carrying a purpose means committing to it and answering for it, and that remains with people, level by level, or the purpose is not carried at all. An AI system can store a mission statement. It cannot mean it.

More capacity without shared direction creates extra work instead of value. Every extra document still has to be read. Every extra analysis still has to be checked. Every extra decision still has to fit the decisions of others. Two documented mechanisms show how the extra volume becomes coordination cost. First, the work moves instead of shrinking: writing is now cheap, reading is still expensive. An employee saves three hours by letting AI draft a report; fifteen colleagues then spend half an hour each reading it. The cost moved from one desk to fifteen desks. This receiving end has been measured: a 2025 Harvard Business Review study with Stanford found that a worker who receives a low-quality AI-produced piece spends on average nearly two hours interpreting, correcting, or redoing it. Atlassian’s 161-billion-dollar estimate is the price tag of the same fragmentation at Fortune-500 scale. Second, the workflow stays the same: most organizations use AI to speed up their existing processes, and almost none redesigns the work around the new tool. Electricity had the same history. Factories that only replaced the steam engine gained nothing for decades. The gains came when they redesigned the factory floor around electric motors.

On top of this, the gains that do exist stay invisible. An employee who finishes a six-hour task in forty-five minutes gets more work, or a smaller team next year. So he keeps the gain to himself, and the organization never sees it. This is not extra work; it is missing information. It is part of why so few executives can point to any organization-wide gain. The repair is cultural, and Section 3.5 names it.

This is why alignment needs a reference at every level: friction of this kind only becomes visible — and correctable — when each level has something explicit to measure its actual behavior against. Without that reference, the friction has nowhere to show up, and it simply grows.

2.2 Second cause: a jagged tool, an unmeasured skill.

Section 2.1 described what goes wrong between people when capacity grows. This section is about what goes wrong within a single use of the tool. There, two things decide the outcome: whether the task lies inside the tool’s real competence, and whether the user can see where that competence ends. Almost no organization can tell the two apart, because it measures neither the tool’s contribution nor the user’s skill. What we know is this.

AI is jagged. It is very good at some tasks and surprisingly bad at neighboring tasks, and the border between the two is invisible — the tool itself never signals it. A field experiment by Harvard Business School and Boston Consulting Group gave 758 consultants access to GPT-4. On tasks inside the tool’s area of competence, they completed about 12 percent more work, 25 percent faster, at measurably higher quality. On tasks just outside it, the same consultants were 19 percentage points less likely to reach a correct answer. Same tool, same people, opposite results. The difference was whether the task sat inside or outside a border nobody could see.

Who gains most? In the same experiment, the weakest performers gained the most — inside the tool’s competence, AI worked as a leveller. The reading this paper gives of the whole pattern — Section 6 marks it as a reading, consistent with the evidence but not yet separately proven — is that the deciding skill is direction in miniature: a user with a clear goal gives the tool concrete instructions, checks the output against that goal, and rejects what does not serve it. If that is right, the people who create value with AI are not necessarily the brightest technical minds. They are the people who can state clearly what value they want to add, translate it into concrete instructions, and judge the result — people who understand their field at the conceptual level, not only the technical one.

This is why recruiting and training must change. The tool can be bought. The skill cannot — it has to be selected for and built, and today it is neither measured nor named in most job profiles.

2.3 Third cause: a tool cut off from reality.

AI brings no feedback from reality. A model learns from recorded text and data, and in use it mirrors its user; the consequences of its own answers never reach it. Two documented drifts follow.

First, AI confirms the user’s worldview. A wrong idea normally dies against resistance: colleagues push back, reality resists. AI removes that resistance. A Stanford-led study published in Science in 2026 tested eleven popular AI models on thousands of real situations. The models endorsed the user’s behavior about half again as often as humans did. In situations where human consensus said the user was clearly in the wrong, the models still sided with the user in about half the cases — humans almost never did. And after one agreeable conversation, users were more convinced they were right and trusted the AI more. The same mechanism appears in professional settings: clinicians in one study trusted AI advice more when it confirmed what they already believed (Bashkirova and Krpan, 2024).

Second, AI levels down the diversity of thinking. Writers who start from AI suggestions produce work that is more similar to each other’s (Doshi and Hauser, 2024) — individual quality can rise while the diversity of the whole falls.

Note that these two drifts pull in opposite directions — one locks each user deeper into his own view, the other pulls all users toward the same middle. That is not a contradiction; they are two sides of the same missing resistance. A wrong idea used to meet two correctives: reality pushed back, and colleagues saw differently. The tool supplies neither — it flatters the individual and averages the collective. Both correctives must therefore be rebuilt by the organization itself: measure alignment with customers, society, and the wider environment, and make the finding and checking of truth a routine. That is what requirements 2 and 4 of the summary exist for, and Sections 3.2 and 3.5 build them out.

2.4 The damage: three forms, not to be confused.

When capacity grows and direction does not, an organization gets hurt in three distinct ways.

• Friction. The parts work against each other. Levels drift apart, or the organization drifts away from its customers and the world it operates in. Visible as double work, rework, and meetings about meetings — the coordination cost measured above.
• Hollowing. The organization starts serving its own measures instead of its purpose. Activity stays high, output stays high, the numbers may even improve — while the substance underneath drains away. Hollowing has a second face: everything pulls to the middle. Diversity makes place for conformity, and innovation for conservation — the organization keeps repeating what it already does, only more of it. From the outside it looks healthy, until the moment it snaps.
• The say-do gap. The organization’s actual conduct drifts away from what it says about itself. This gap can be measured directly, by comparing claims with recorded behavior.

One distinction matters here: not all friction is unhealthy. Disagreement between people who share a goal is healthy friction — it is how errors get caught and better ideas win, and it must be protected. Unhealthy friction is the friction of misaligned direction: the parts pull against each other and nothing comes out of it. The first is a sign of life; the second is the damage.

These are three different problems. An organization can have any one of them without the other two, and each needs a different repair. That is why they must be measured separately — which is what the instrument in Section 3 does.

2.5 None of this corrects itself.

The reason is simple: the instruments organizations use all measure activity and output. A dashboard full of green numbers answers the question “how much are we doing?” It never answers the question “do we still act as one, and what do we answer to?” As long as only the first question is measured, success itself protects the problem: nobody investigates what looks like it is working. The missing measurement is therefore not a detail. It is the heart of the matter, and the subject of the next section.

3. The answer: radical alignment of orientation

If direction is the deciding variable, then the cure is to align it. Radically means: at the root, at every level, and measured instead of assumed. Five elements turn this from a slogan into a working discipline.

3.1 A charter: something real to align to. An organization cannot be aligned “in general.” It is aligned to a reference, and in most organizations that reference is incomplete. The mission statement is written for the outside world. The values describe how people want to work together — they are a real part of the organization’s meaning, but they hollow out easily when nobody compares them with daily behavior. The strategy changes with the market. A charter brings these together in a form that can be checked. It states: what this organization is, what it wants to become, what it stands for, what it refuses under all circumstances, and how it accounts for itself. Every sentence must pass one test: can we compare our actual behavior with this sentence and see the difference? A sentence too vague to compare with behavior does not belong in a charter. Two habits keep a charter alive. First, the distance between charter and actual behavior is measured openly, and anyone in the organization may point at it without being punished. Second, the charter itself stays alive: it evolves with the organization, its market, and society — openly, with old versions kept — so it never freezes into a museum piece.

3.2 A charter at every level — including the AI. This is the radical part. Not only the organization has a charter. Each department, each team, and each coworker — internal or external — states one. And equally: each AI agent and each major system gets one. For an AI agent, the charter states what it is for, what it may decide on its own, what it must hand to a human, and what it must never do. From that moment the agent can be measured like any team: does it do what its charter says, and does its charter still serve the level above it? This is lighter than it sounds. A charter at team level is a page at most. For an AI agent it is not even a document in the human sense: it is the agent’s operating constraints made explicit — in a form that both a human and the system itself can check. And the bureaucratic version of this idea — thousands of static documents nobody reads — is exactly what the measurement discipline prevents: a charter nobody compares with behavior fails the test of Section 3.1 and is removed.

Every element is then measured twice. First: does this element do what its own charter says? Second: does its charter still serve the charter of the level it contributes to? A team can hit every target and still drift away from what its division is for, causing gradually more friction or silently hollowing out the organization from the inside. You only see that drift when you take both measurements separately and compare them. Because every person, every unit, and every agent has its own reference, drift becomes visible early — at the level where it starts — and can be corrected precisely, before it spreads.

Moreover, the measurement does not stop at the top of the organization. The organization itself is measured against its customers and against the society it works in, including the zeitgeist and the geopolitical and economic environment. That answers the question every strategy debate circles around — how do we know our direction is still right? — not with conviction, but with measurement against the levels above. As such, an AI project is never judged only on its outputs. It is judged on whether its purpose still serves the purpose above it, all the way up.

3.3 The instrument: measuring from conduct, not from words. The instrument behind this paper, Integration Capacity Analysis (ICA), was developed by the author on the basis of thirty years of turnaround and leadership practice. It is proprietary, its protocol is fixed in advance, and its track record is public: a battery of 31 historical cases scored under blind conditions (Section 6), published structural reads of countries, an economic sector, and the European Union, and dated forward calls published before the outcome at integrationcapacity.org. The instrument takes its measurements from conduct and structure — from what a system actually does — not from surveys or self-description. In one sentence: it measures whether a system still acts as one, and what it answers to. Concretely, for each level it answers the questions no activity dashboard answers. Do the parts still pull together, or only appear to? When a decision is taken, is it executed — or nodded at and quietly ignored? Do resources turn into coordinated capability, or do they pile up while the ability to use them drains away? Is the direction that is actually lived the one that is declared — and how wide is that gap? And is apparent health real, or propped up from outside?

The sharpest public case shows what that means in practice. In March 2023, Silicon Valley Bank collapsed in forty-eight hours. Until the end, its standard indicators looked healthy: profitable, officially “well-capitalized.” The instrument was run on this bank and on a bank that faced the same interest-rate shock and stood, JPMorgan Chase — using only records dated from before the event, with the scoring rules frozen in advance. It placed the first bank deep in the danger band and held the second in the healthy band. This is a worked demonstration, not a prediction: the cases were chosen after the outcomes were known, and the published version says so. What it shows is the thing that matters here: the measurement separates a hollow system from a sound one on information that was available beforehand. It reads what actually holds a system up, not what the system produces. The full demonstration — four famous failures, each paired with a survivor of the same shock — is published at integrationcapacity.org.

3.4 The manager: from managing output to managing alignment. Purpose crosses levels through people. The person at each crossing point is the manager, and his job changes. The old job was managing output: set targets, check volume. Machines now produce, and even check, much of the output. What machines cannot do is keep a unit aligned. So the manager’s core work becomes threefold: alignment inside his unit; alignment between his unit and the organization; and the charter friction of both — of the unit as a whole, and of the people, instruments, and AI agents that are part of it. In both directions: upward, he condenses what his teams are really doing and really facing into one honest picture for the level above; downward, he translates the direction from above into terms that fit each team’s real situation. This does not mean output stops mattering. It means the manager works the way agile management works: in short cycles — align, produce, check the result against the purpose, adjust. Efficiency remains; effectiveness leads. And the design rule for AI follows: let AI do the mechanical part of coordination — collecting, formatting, routing — so the manager has time for the part no tool can do: knowing his people, and holding the whole to its purpose.

3.5 A culture of feedback and truth. None of the above works in a culture that punishes truth. So the deepest element is cultural. An organization that wants alignment must become an organization that collects feedback from reality, and finds, checks, and evaluates truth — about its results, about its direction, about itself — more than anything else. In practice this means six things.

Recruit and train for judgment. Using AI well is a skill, and its parts can be tested: checking an answer against its source before trusting it; trying to break an answer before accepting it; judging whether an output is fit for its purpose; keeping a question open when the evidence is not in yet; and changing one’s mind when the evidence changes. Hire for these abilities, and train them.

Work experimentally. Treat important decisions — including every AI adoption — as experiments: a named owner, an expected result stated in advance, a check afterwards, and a change of course when the facts say so. This is how truth-finding becomes routine instead of heroism.

Protect disagreement. AI pushes everyone toward the same answers, because it mirrors its users and never touches reality itself. Growing sameness feels comfortable, and it is dangerous: it erodes the diversity of thinking that correction and innovation depend on. So keep several independent views alive on every important question, and make it safe to bring bad news — a person who reports a problem must never pay a price for it.

Reward surfaced gains. Here the incentive mismatch of Section 2 is cured. An employee who shows a time gain, or a manager who reduces coordination cost while keeping his unit aligned, must be rewarded for it — not punished with more work or a smaller team. As long as gains are punished, they will stay hidden, whatever the tooling.

Keep the craft alive. If AI writes every first draft, nobody learns to draft — and in ten years nobody can judge the machine’s work. So decide, on purpose, where people still practice without AI. The loop to avoid is simple: automation reduces practice; less practice means less competence; less competence means weaker oversight; and weaker oversight invites more automation.

Build a common worldview. Write down, explicitly, how the organization believes its world works: its market, its customers, what creates value and what destroys it. The best-known example is an investment firm: Ray Dalio’s Bridgewater wrote its worldview down as principles and tests decisions against them. The point is not that the written worldview is right. The point is that only a written, shared worldview can be tested against reality, corrected, and used to judge whether the measurements — and the efforts — still make sense.

This culture is not decoration around the measurement. It is what the measurement is for: the measurement finds the gaps; the culture closes them.

4. The state: the same problem one level up

Governments are organizations too. Everything above — capacity and direction, charters, drift, measurement — applies to a ministry or a regulator exactly as it applies to a firm. And governments sit in this problem twice: they adopt AI in their own work, and at the same time they oversee everyone else’s adoption of it.

The law demands something there is no standard way to check. The EU AI Act requires, for high-risk AI systems, detailed logging, continuous risk management, and — the core clause — effective human oversight: a person must be able to monitor the system, intervene in it, and stop it. Now consider how that could ever be verified. Not by looking inside the machine: today’s AI systems are, in the relevant part, black boxes. Their full reasoning cannot be inspected, so “the human understands what the system is doing” cannot be certified at the level of the model. It can only be checked at the level of the organization around the system: does the organization state what the system is for and what it must never do; does information about what the system actually does reach a person with authority; does that person demonstrably decide; and does the whole still follow the direction the organization has declared — toward its own purpose, and within what society and law demand? In other words: oversight of AI can only be measured as alignment. Today no regulator has a standard way to make that check on a company — and none can make it for itself either. The ambition of the law is right. The instrument to enforce it is missing.

Sovereignty, defined. Everyone uses the word; almost nobody defines it. Here is a definition with two testable parts. An organization — or a state — is sovereign when it has set an explicit, and as far as possible public, direction; and when it can show that it still controls that direction in practice. AI puts pressure on the second part from two sides. From outside: dependence. Most organizations rent their AI from a handful of suppliers. The supplier sets the price, the terms, and the pace of change; the customer adapts. That is a fact of the current market, not a judgment. But dependence that is not priced and not managed slowly becomes control from outside. And note what sovereignty here does not require: you do not have to own the model to stay sovereign. You have to keep control over what the model is allowed to do, for what purpose, and who can change that. From inside: delegation. Decision after decision is handed to AI systems, because it is more efficient. No single handover changes the direction of the organization. Thousands of small handovers together can — without anyone ever having decided it. So the sovereignty question is not “ban AI or adopt AI,” and not “domestic or foreign.” It is: who sets the direction here, in practice — the charter, the suppliers, or the sum of all the small handovers nobody looks at together?

That question can be answered with the measurement of Section 3.3. And this is why the discipline in this paper is called sovereign alignment. Alignment is the work: everything pointed at the purpose. Sovereign is the condition the work protects: the organization still directs itself.

What law can do, and what it cannot. Law can require what can be written down and checked: logs, reports, procedures. Direction cannot be required by law. No statute can make an organization mean its purpose; direction lives in people. This is why organizations with perfect compliance departments can still lose their purpose, and why every new regulation produces a new layer of paper obedience. What law can usefully require are the conditions under which organizations correct themselves. That a critical finding from an auditor or a court is answered by a traceable decision — not only by a polished report about the finding. That auditors are independent in fact, which is decided by who appoints them and who pays them. That people who report problems are protected. And that important public judgments are not all produced through the same AI layer — because a state whose ministries, media, and markets all think through the same models has traded many instruments for one, without deciding to. In short: law can keep correction possible. People still have to do the directing.

5. The opportunity: what alignment pays

Aligned teams cooperate cheaply. When two teams work toward the same purpose, they need less from each other: less checking, less negotiating, fewer documents written to cover oneself. Even their disagreements are useful, because both sides argue about how to reach the same goal — healthy friction, the culture of Section 3.5 at work. This is the real prize of the AI age. It is not efficiency; efficiency makes the parts faster. The prize is that the parts, now enormously capable, work as one whole. Whether this also holds for whole societies is theory, and we mark it as theory. But the most productive periods in history — the rebuilding after 1945 is the clearest case — show exactly this pattern: not better people, but less loss between people, all working together for the improvement of the future.

The human gain is the deepest one. For a century, most office workers have spent most of their day on exactly the work AI now does almost for free: drafting, formatting, collecting, routing, reporting. Suppose AI genuinely takes over much of that layer — not as an excuse to cut staff, but with the freed time captured and put to use. Then human work shifts to what only humans carry: judging, deciding, providing meaning, and holding a team to its purpose. Two conditions from Section 3.5 decide whether this really happens: the freed time must be made visible and rewarded, and people must keep learning the craft, so the next generation can still judge the machine’s work. Where both hold, the levelling of Section 2.2 becomes a property of the whole organization: people at every level of talent become able to create real value with the tool.

The purpose itself sets the ceiling. Does it matter what the purpose is? Yes, and the reason is practical, not moral. Ask where a purpose ends. Some purposes end in something finite: the share price, market dominance, GDP, the survival of the institution itself. Such a purpose stops working the day it is reached — and then the organization starts circling around its own past. Other purposes point beyond the organization, at a good it serves but does not own. Such a purpose survives every leadership change, and there is always a next step toward it. The test is simple, and every board can run it today: ask what society would lose if this organization stopped existing tomorrow. If the honest answer is “nothing another supplier could not replace,” the purpose ends in the organization itself — and so does the energy it can generate. If something real would be lost, the purpose points beyond the organization, and it can keep pulling the organization forward.

One direction, many eyes. One objection deserves an answer before anyone raises it: “shared direction” has been the sales pitch of every unifying ideology, and it usually ended in forced sameness. This discipline closes that road itself. Forced sameness is a failure of alignment, not a form of it: an organization where everyone thinks the same loses the ability to see its own mistakes, and that ability is the heart of the whole method. So the goal carries its own guard: one direction, many eyes. Shared purpose is what makes different views workable; different views are what keep the shared purpose honest. A project that builds the first by destroying the second is not an ambitious version of this program. It is a failure of it.

6. What is proven, what is validated, what is theory

The claims in this paper do not all have the same strength. Separating them is part of the method.

Documented by others: the gap between individual gains and organizational results (Atlassian 2026; McKinsey 2026); the receiving-end cost of low-quality AI output (BetterUp Labs and Stanford, 2025); the jagged profile of AI — strong inside its area of competence, harmful just outside it — and the levelling effect inside the frontier (Dell’Acqua and colleagues, 758 consultants); AI siding with its users and the effect of that on users (Cheng and colleagues, Science, 2026); confirmation bias in AI-assisted decisions (Bashkirova and Krpan, 2024); the growing sameness of AI-assisted output (Doshi and Hauser, 2024).

Validated at face level (the classification fits the historical record, and independent raters agree on it): the framework’s classification of what happens to a system under this kind of strain. Four outcomes: renewal (the system reorganizes itself around the new capability, at a higher level), held descent (no crisis, no renewal — a permanently strained plateau; the likely default for an organization that changes nothing), capture (the system starts serving something smaller than its purpose — a supplier’s roadmap, a cost target that became the real goal), and dissolution. Tested on 31 historical cases, in two blind rating rounds with seven independent raters; agreement rose from κ = 0.66 to κ = 0.72 as the rules were sharpened, and no case ever forced a fifth category. As stated on the public record: the raters are AI models and the case briefs were written in-house; human raters and independent briefs are the next step.

Theory, internally consistent, with outcome validation as the next step: the capacity-versus-direction explanation of the problem; the reading that direction-skill is what decides who gains from AI (Section 2.2); the claim that alignment pays at the scale of societies; the claim that the purpose itself sets the ceiling. The instrument has a fixed protocol, a built-in check that two independent measurements must agree, and published results.

The instrument exists and is in use; its published measurements are at integrationcapacity.org. The next step is applying it where alignment decides value: inside organizations, at every level, against every charter, continuously.

Sources

Atlassian, State of Teams 2026 (double-blind survey, 12,035 knowledge workers and 173 Fortune 1000 executives, fielded January–February 2026; 89% / 6% figures; $161bn estimated annual cost of fragmented AI adoption across the Fortune 500). McKinsey & Company, The State of Organizations 2026 (survey of more than 10,000 senior executives across fifteen countries; 88% experimenting with AI; 81% report no meaningful bottom-line gains). Deloitte, The State of AI in the Enterprise 2026 (survey of more than 3,000 executives; productivity gains reported as widespread, while only 30 percent of organizations redesign key processes around AI). F. Dell’Acqua, E. McFowland III, E. R. Mollick, H. Lifshitz-Assaf, K. Kellogg, S. Rajendran, L. Krayer, F. Candelon, K. R. Lakhani, “Navigating the Jagged Technological Frontier” (HBS/BCG field experiment, 758 consultants; published in Organization Science, 2025). M. Cheng, C. Lee, P. Khadpe, S. Yu, D. Han, D. Jurafsky, “Sycophantic AI decreases prosocial intentions and promotes dependence,” Science 391 (2026): eleven models; ~49% more affirmation than humans; sided with users in about half of clearly-wrong cases where humans almost never did. A. Bashkirova and D. Krpan, “Confirmation bias in AI-assisted decision-making,” Computers in Human Behavior: Artificial Humans 2 (2024). A. Doshi and O. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 10 (2024). BetterUp Labs and Stanford Social Media Lab, “AI-Generated ‘Workslop’ Is Destroying Productivity,” Harvard Business Review, September 2025 (recipients spend on average 1h56 per instance). EU AI Act, high-risk system provisions on logging, traceability, risk management, and human oversight (Article 14: high-risk systems must be designed so that natural persons can effectively oversee them — monitor, intervene, stop).