What AI Actually Changes in Large Organizations — and What It Does Not

Start With What the Tools Cannot Change

Start with the numbers. In a 2025 study from Harvard Business Review by Jin Li, Feng Zhu, and Pascal Hua, 45% of executives reported that their AI investments came in below ROI expectations. Only 10% exceeded them [2]. The technology works. The pilots succeed. The disappointment shows up somewhere between the model that produces the answer and the human who has to decide what to do with it.

I have been spending most of my time inside that gap. The leaders I talk with do not have a tooling problem. They have a different kind of problem, and they usually feel it before they can name it. Their dashboards look fine. The cost-per-output line is dropping. And yet the value the deck promised is somewhere else, doing something else, refusing to land.

The default story is that better tools will fix this — better prompts, better change management, better training. The story the data actually tells is different. AI selectively removes one kind of constraint. Three others remain, and they are the ones now setting the speed of the system. Organizations that succeed at this redesign for what AI did not change, not for what it accelerated.

The question is which constraint is which.

The Constraint AI Actually Removes

What AI compresses, precisely, is coordination. Alex Farach’s 2026 arXiv paper on coordination-compressing capital puts it as cleanly as anyone has: “Agent capital does not make individual tasks more productive but makes the organization of tasks more efficient” [1].

The distinction is between doing a task faster and reducing the work needed to organize tasks. The middle of every large organization — the phases, the layers, the handoffs that exist because humans cannot hold the full picture in their heads — was built to manage the cost of coordinating cognitive work. AI dissolves a meaningful share of that cost. Entire management layers existed to solve a problem that just got cheaper.

This is not a thought experiment. Ewens and Giroud studied more than 3,100 public companies and found that firms have flattened hierarchies and widened spans of control following AI adoption. The first visible effect is acceleration: planning improves, analysis scales, coordination becomes easier. (“Making the organization of tasks more efficient” sounds like a sentence designed to put a board to sleep, until you realize it describes the entire reason your org chart looks the way it does.)

That is one layer of the system, and only one.

What Does Not Change: Physics, Politics, and the Question of Who Signs

Physics first. In a 2026 European Business Review analysis of smart manufacturing, the most consistent failure mode is a hardware-software mismatch: companies invest heavily in algorithmic optimization and discover that legacy machinery, facility limitations, and equipment wear cannot keep pace [7]. The plant has its own clock. Better models do not pour concrete faster. They do not shorten installation cycles or rewrite the physical lead time of a turbine. AI optimizes sequencing inside material reality; it does not amend it. The most counterintuitive finding here is that the organization that invested most in AI optimization is usually the first to discover that the ceiling is the factory floor, not the algorithm.

Politics second. Organizations are coalitions, not machines, and the same HBR study makes that visceral: programmers were 16 to 18 percent less likely to recommend AI access to their own teammates than to others. Managers protect team size because team size is authority. Junior employees outperforming veterans creates hierarchy tension that no model resolves. The Farach evidence sharpens the point: hierarchy flattens across those 3,100 companies, and the manager-worker wage gap universally widens. Identical technology, divergent political outcomes — depending on who controls the elasticity. AI destabilizes the political structure without resolving it.

(Programmers using AI to hoard AI access is probably the most measurably political act in modern tech. The irony is doing its own work; I am just pointing at it.)

Accountability third — and last because it is the heaviest. When the decision chain compresses, ownership becomes visible in a way it was not before. Mario Thomas documented in 2025 that 63% of organizations lack AI governance policies while 88% deploy AI in at least one function [3]. Capability is outpacing oversight at the speed of procurement. Meanwhile, the courts have started doing what governance has not: more than 120 court decisions involving AI-hallucinated citations have been recorded, with the rate escalating from roughly two per week in early 2024 to two or three per day by spring 2025. The legal system is concentrating what organizations have been busy diffusing.

That leaves an apparent contradiction: a shorter decision chain can still make responsibility harder to locate.

The Accountability Paradox

A common objection cuts the other way: AI does not concentrate accountability. It diffuses it. Opaque algorithms scatter responsibility across black-box systems instead of people, as Luca Collina, Mostafa Sayyadi, and Michael Provitera argue in the California Management Review [8]. When 63% of organizations deploy AI without governance, ownership becomes blurrier, not sharper, because the decision chain has been abstracted into a system most of the people running it do not fully understand. If accountability were really concentrating, ownership would be clearer post-AI. In most organizations, it is not.

Both dynamics are running at once, and that is the structural condition we are inside. AI creates the conditions for accountability concentration — fewer actors, compressed decision surfaces — and organizations default to diffusion because governance has not caught up. The court decisions are evidence of the gap, not the resolution. Courts are slowly forcing the concentration organizations have not voluntarily built.

Two cases make this concrete.

In late 2022, Air Canada deployed a customer-facing chatbot. Jake Moffatt, whose grandmother had just died, asked it about bereavement fares. The chatbot told him he could book at full price and apply for the discount retroactively within 90 days. That was wrong. When Moffatt sought the refund, Air Canada argued in tribunal that the chatbot was a “separate legal entity” responsible for its own actions. In February 2024, the British Columbia Civil Resolution Tribunal rejected the argument in essentially one sentence: a chatbot is still just part of Air Canada’s website. The airline paid $812 in damages. The financial penalty was trivial. The governance precedent was not. The American Bar Association and most major legal commentators now cite the ruling as foundational. Cognitive constraints compressed. Accountability constraints did not. Pretending otherwise produced legal liability.

OpenAI is the other shape of the same problem. The organization was deliberately built to concentrate accountability for safety in a board insulated from commercial pressure. In November 2023, the board exercised exactly that authority and fired Sam Altman. Within 72 hours, more than 90% of employees had signed a letter threatening to resign en masse and join Microsoft. Microsoft applied its leverage. Altman was reinstated within five days. The Harvard Law Review later called the pattern “amoral drift”: governance structures designed for accountability systematically drifting toward commercial optimization under stakeholder pressure [9]. The accountability mechanism was exercised correctly and then politically overridden. Legal authority without political infrastructure is structurally fragile.

The result of unresolved accountability gaps is initiative death. Thomas’s reporting also found that 42% of organizations are abandoning AI initiatives because the gap between what the system can do and who owns it cannot be closed with their current governance models, while 43% plan to replace entry-level roles with AI even as they reduce senior verification staff. The expertise pipeline thins on a five-to-ten-year horizon. More output, fewer reviewers, no named owner: a mathematical setup for failure at scale.

This is the part most strategy decks skip, and where I find the conversation with executives changes register. Once you stop arguing about whether AI is fast enough, you have to start arguing about whether anyone in the room is willing to sign for what it produces. That is a different meeting.

Where the Framework Breaks — and Where It Holds

Two objections deserve a direct response.

The first: the taxonomy misses emergent constraints. Verification bottlenecks are real. So is the expertise pipeline destruction implied by replacing junior analysts before the next generation of seniors has been built. Both are real. Both are also second-order interactions between removed and remaining constraints, not new categories. Verification bottlenecks arise because cognitive generation compresses while accountability sign-off does not — the boundary between a removed constraint and a remaining one. The pipeline gap is a governance choice, not a technological inevitability. The taxonomy is not a map of every AI effect. It is a map of the structural fault lines where the most consequential failures concentrate.

The second is sharper: a framework that tells leaders what does not change is descriptively interesting but strategically inert. That criticism is valid if the article stops at taxonomy. Capability-first frameworks tell you what to deploy. Constraint-first tells you what your governance has to absorb before you do. The inversion is the practical contribution, and DBS has already built the proof of concept.

The Constraint Inversion: A Diagnostic, Not a Dashboard

Most AI transformation frameworks start with capability: what can AI do for us? The Constraint Inversion Audit starts with the opposite question — what will AI not change, and are we structured for that? Four steps.

One. Classify your constraint portfolio. For each high-stakes workflow, ask: if we could process information infinitely fast, would this workflow get faster? If yes, it is cognitively bound and AI will compress it. If no, identify whether physics, politics, or accountability is the binding constraint. The common failure mode is classifying workflows as cognitive because that is where AI helps, not because it is accurate. A procurement process may look cognitive (analyzing vendor data) when it is actually political (department heads protecting preferred vendors). Misclassification accelerates the wrong part of the process and exposes the real bottleneck without resolving it.

Two. Map constraint migration. When cognitive work compresses, bottlenecks migrate. Simulate the move: if AI generated this deliverable in 10 minutes instead of 10 days, who would need to approve it, and how long would that take? The gap between generation speed and approval speed is the new bottleneck. Thomas’s analysis found 43% of organizations plan to replace entry-level reviewers with AI while reducing senior verification staff — more output, fewer reviewers, a mathematical mismatch.

Three. Stress-test governance against the remaining constraints. Define who signs off on AI-assisted decisions before deployment, not after. For political constraints, map the incentive conflicts — resource hoarding, hierarchy threats — and define escalation paths. For physics, validate that the physical infrastructure can absorb the throughput AI enables. Test against realistic stakeholder pressure, not paper compliance. That 63 / 88 governance gap [3] is what this step is built to close before it becomes a court case.

Four. Instrument the constraint boundary. Three measurements. Verification ratio: AI-generated outputs versus human-verified outputs per workflow. Decision latency delta: time from AI output to human sign-off. Stakeholder friction index: escalations, overrides, and blocked decisions at political constraint points. When the ratios drift, investigate whether output is outrunning review or decisions are waiting on unresolved conflicts.

DBS Bank is the proof of concept. Southeast Asia’s largest bank scaled AI across more than 1,500 models and 370 use cases between 2019 and 2024 — and built its PURE framework (Purposeful, Unsurprising, Respectful, Explainable) starting in 2019, before scaling reached critical mass. SGD 370 million in economic value from AI in 2023; more than SGD 750 million in 2024, per DBS public disclosures and HBR coverage. DBS did not succeed because its AI was better. It succeeded because its governance was earlier.

Before the next deployment

Three things to do before the next AI deployment.

Name the accountable person for every AI-augmented workflow before deployment, not after. If you cannot name them, the deployment is not ready.

Measure the gap between generation speed and approval speed. When AI output arrives faster than humans can verify it, you have built an accountability vacuum — and that vacuum is currently the leading reason organizations abandon AI initiatives.

Stop treating political resistance as a change management failure. Stakeholder self-interest is a structural constraint that requires structural solutions, not better communication.

The question is no longer how decisions are made. It is who is accountable when they are.

References

• [1] Farach, Alex. (2026). AI as Coordination-Compressing Capital: Task Reallocation, Organizational Redesign, and the Regime Fork. arXiv. https://arxiv.org/html/2602.16078v3
• [2] Li, Jin; Zhu, Feng; and Hua, Pascal. (2025). Overcoming the Organizational Barriers to AI Adoption. Harvard Business Review. https://hbr.org/2025/11/overcoming-the-organizational-barriers-to-ai-adoption
• [3] Thomas, Mario. (2025). The Accountability Gap: When AI Delegation Meets Human Responsibility. mariothomas.com. https://mariothomas.com/blog/ai-agency-accountability/
• [7] (2026). From Digital Transformation to Physical Constraints: The Overlooked Limits of Smart Manufacturing. The European Business Review. https://www.europeanbusinessreview.com/from-digital-transformation-to-physical-constraints-the-overlooked-limits-of-smart-manufacturing/
• [8] Collina, Luca; Sayyadi, Mostafa; and Provitera, Michael. (2023). Critical Issues About A.I. Accountability Answered. California Management Review. https://cmr.berkeley.edu/2023/11/critical-issues-about-a-i-accountability-answered/
• [9] (2025). Amoral Drift in AI Corporate Governance. Harvard Law Review. https://harvardlawreview.org/print/vol-138/amoral-drift-in-ai-corporate-governance/
• Air Canada chatbot ruling: Moffatt v. Air Canada, 2024 BCCRT 149. American Bar Association analysis. https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/
• DBS Bank AI governance and the PURE framework. https://www.dbs.com/artificial-intelligence-machine-learning/artificial-intelligence/dbs-ai-powered-digital-transformation.html
• The Conference Board (2026). AI and the C-Suite: Implications for CEO Strategy in 2026. https://www.conference-board.org/research/ced-policy-backgrounders/ai-and-the-c-suite-implications-for-ceo-strategy-in-2026
• Harvard Law School Forum on Corporate Governance (2026). Top 5 Corporate Governance Priorities for 2026. https://corpgov.law.harvard.edu/2026/04/07/top-5-corporate-governance-priorities-for-2026/