Misalignment at Speed.

Welcome to the age of Agentic AI.

Ninety-five percent of enterprise AI pilots will never deliver a return on investment, and the failure has nothing to do with the intelligence of the software.

The AI in Products Community Event

On May 5th, 2026, I stood at the back of a packed room at the AI in Products Community Event. I was there to speak. The honest reason I attend events like this one is straightforward: I go to share what I have seen in the field and to learn from others encountering the same structural problems. The hallway conversations between sessions told me more than any slide deck on the main stage.

A data architect from a retail conglomerate described spending eight months building an AI readiness assessment, only to watch the initiative die in a steering committee that never understood what it was approving. A product lead from a healthcare company confessed that her team’s “production” model was actually a series of manual Excel corrections disguised as automation. A platform engineer from a bank said the quiet part out loud: “We are not building AI systems. We are building PowerPoint decks about AI systems.”

These hallway conversations are the real quarterly business review of the enterprise AI industry. The official sessions present the success cases — the polished demos, the ROI claims, the before-and-after metrics. The hallways process the failures. The failures all sound the same.

The Quarterly Business Review as Power Theater

Picture the standard quarterly business review in a large financial services firm. Five years ago, this room held a specific kind of tension. Data engineers debated database architects over indexing strategies. Product managers defended feature scope against the capacity limits of the engineering team. The arguments were technical, granular, and grounded in the physical constraints of servers, latency, and human labor hours. The hierarchy in the room was visible. The person who understood the system’s limits held the most authority when a decision needed to be made.

Walk into that same conference room today. The vocabulary has changed. The screen at the front displays a slide titled “AI-Driven Operational Transformation.” The Chief Information Officer is speaking, but the language has shifted from how the system works to what the system will achieve. The data engineer sits in the back row, arms crossed, checking their phone. They have been rendered mute. The new hierarchy does not reward the person who understands the technical limits. It rewards the person who can narrate the algorithmic future.

The enterprise AI initiative has fundamentally restructured the sociology of this meeting. The old meeting required operational fluency. The new meeting requires rhetorical fluency. The CIO describes a future where machine learning models automate risk assessment. The product manager nods. The data engineer remains silent. The engineer knows the training data is fragmented across fourteen legacy databases with inconsistent schemas. The engineer knows the “AI-driven future” on the slide is structurally impossible given the current state of the data pipelines. The engineer holds the reality of the organization in their head, but the meeting is no longer structured to hear it.

This silence is the most important sociological data point in the modern enterprise. The 82% of leaders who believe their AI strategy is structurally coherent are looking at the slide deck. The 23% of individual contributors who actually agree are looking at the fragmented databases. Monday.com research confirms this fracture from a different angle: 45% of senior leaders believe change is being managed well, compared to only 23% of individual contributors. The meeting room has become a theater where structural coherence is performed rather than achieved.

The New Hierarchy of the Algorithmic Age

Every technology redistributes authority within an organization. Enterprise AI is displacing the architects of process and elevating the architects of narrative. This is the new hierarchy, and it operates through a mechanism best described as a legitimacy machine.

The legitimacy machine works by converting political decisions into algorithmic necessities. When a leadership team decides to automate a customer service workflow, they are making a political choice about which customer interactions deserve human empathy and which can be standardized. If they presented this choice to the organization as a political choice, it would invite debate. The legitimacy machine prevents this debate. The leadership team frames the automation as a “technical upgrade” or an “AI integration.” The decision is stripped of its political weight and presented as an act of engineering progress.

The people who benefit from this reframing sit at the top of the new hierarchy. They gain the authority to reshape human work without subjecting that reshaping to organizational scrutiny. The people who are harmed sit at the bottom. They lose discretion over their work and are reclassified as “data sources” or “human-in-the-loop validators.” Their expertise is not eliminated. It is extracted, codified into a training dataset, and owned by the central platform.

The power dynamics are visible in the architectural blueprints themselves. Consider the federated operating model promoted by modern data platform vendors. The documentation describes this as a technical pattern for organizing compute resources. Look past the technical language. The architecture is a new org chart written in infrastructure-as-code.

Central IT owns the platform account and the unified catalogs. They build the highway and set the speed limits. Data Engineering owns the pipelines and the intermediate data tables. They maintain the asphalt. Business Intelligence teams are granted read-only access to prevent them from breaking the underlying structures. Business Units are given create privileges inside their own isolated workspaces. They are allowed to drive, but only within a fenced-in parking lot.

The three-level namespace of a modern data catalog — Catalog, Schema, Table — is a caste system. The aristocracy controls the catalog. The merchant class controls the schema. The laborers interact with the table. This architectural choice determines who can see the entire enterprise and who is confined to a local, segmented view. When an organization implements this architecture without explicitly negotiating the power dynamics it encodes, it defaults to the existing informal power structures. The dominant business unit gets the best data. The marginalized department gets the fragmented tables.

The Dithering Zone as a Site of Lost Agency

When authority is extracted from the people doing the work, the work itself degrades. The research I presented at the event references a phenomenon called the “Dithering Zone” — a state where 71% of engineering teams generate over 250 requirement changes per cycle, and 58% face weekly change requests. This statistic is usually framed as a project management failure. Sociologically, it is a symptom of stolen agency.

The requirement change request is the only weapon available to a worker who has been excluded from the design process. Leadership defines the AI initiative at a high level of abstraction. The charter is signed. The budget is allocated. The engineering team is handed a set of requirements that have been filtered through three layers of management. The engineers read the requirements and immediately see the gaps. The requirements do not account for the edge cases in the legacy system. The requirements assume the data is clean. The requirements define a success metric that is trivially gameable.

The engineers cannot reject the requirements. The new hierarchy does not grant them the authority to say “no.” They can only submit a change request. The change request is a bureaucratic act of rebellion. It is a formal documentation of the gap between the leadership’s narrative and the operational reality. The Dithering Zone is low-intensity organizational warfare disguised as project management confusion.

The cost of this warfare is astronomical, and it compounds as it moves through the system. The NASA cost curve demonstrates this clearly. A defect discovered during the requirements phase costs one unit to fix. The same defect discovered in operations costs up to 1,500 units. The 1,499 unit difference is the financial penalty for excluding operational reality from the planning process. The leadership team saves thirty minutes in a steering committee by deferring a difficult structural question. The engineering team pays for that thirty minutes with months of rework.

The $1.75 billion requirements management market exists to mediate this conflict. Organizations purchase sophisticated traceability matrices and requirement lifecycle tools. The tools do not solve the structural coherence problem. They document the failure with greater precision. Projects using requirement traceability matrices show a 28% higher success rate. The baseline success rate is 31%. The ceiling with these tools is roughly 40%. In large enterprises, the success rate drops to 9%. The tools provide the illusion of control. They allow leadership to point to a completed matrix and claim the process was rigorous, even as the project collapses under the weight of 250 change requests per cycle.

Shadow IT as Rational Resistance

The sociologist Michel de Certeau distinguished between “strategies” and “tactics.” Strategies belong to the powerful. They are the formal plans, the architectural blueprints, the approved budgets. Tactics belong to the weak. They are the improvised workarounds, the unauthorized shortcuts, the informal practices that allow people to survive inside a system they do not control.

The formal AI strategy is the strategy. The ungoverned agentic sandbox is the tactic. The research notes a 45% shadow IT rate and a projected 29% cloud waste rate for 2025. These numbers are universally presented as failures of governance. They are failures of governance only if you assume the governed want to be governed.

Shadow IT is what happens when the formal architecture does not serve the operational need. A business unit needs a report to make a quarterly decision. The formal data pipeline has a six-month backlog. The business unit downloads a dataset to a local laptop, connects it to an unlicensed visualization tool, and builds the report in an afternoon. The central data team discovers this during an audit and classifies it as a security risk and a compliance violation.

From the perspective of the central data team, the business unit is undermining the enterprise architecture. From the perspective of the business unit, the enterprise architecture is a barricade. The business unit is not acting out of ignorance. They are acting out of rational self-preservation. They have a deadline. The formal system failed to meet it. They found a way to meet it themselves.

The 29% projected cloud waste is the financial tariff of this organizational cold war. The business unit spins up unoptimized compute instances without access to the centrally governed, highly optimized platform. The central platform is highly optimized precisely when it is locked down. It is locked down when the central team does not trust the business unit to use it correctly. The business unit proves the central team right by wasting resources in the shadow environment. The cycle reinforces itself.

Organizations attempt to solve this problem with stricter governance. They implement more approval chains. They deploy automated compliance scanners. The Standish Group data shows the result: a 70% transformation failure rate for initiatives using traditional manual approval chains. The BCG research corroborates this, showing a 35% goal-achievement rate for organizations relying on procedural governance. Adding more checkpoints to a broken process does not fix the process. It incentivizes the development of more sophisticated shadow processes.

Governance as Structural Design

The alternative to procedural governance is structural governance. Structural governance lives in the architecture itself. Microsoft’s Cloud Adoption Framework for 2025 forces AI workloads to inherit mature governance policies from day one. Databricks’ Unity Catalog propagates security policies automatically across namespaces. These are not documentation exercises. They are encoded rules that execute without human intervention.

The philosophical distinction is profound. Procedural governance asks, “Who approved this action?” Structural governance asks, “Is this action permitted by the system?” Procedural governance relies on human vigilance. Structural governance relies on architectural constraint. Human vigilance is inconsistent. Architectural constraint is absolute.

When an organization transitions from procedural to structural governance, the power dynamics shift again. The person who used to hold the stamp of approval loses their authority. The system now holds the authority. This transfer of power is deeply uncomfortable for the middle management class. Their entire organizational identity is built on being the node through which all decisions must pass. Removing the approval gate removes their reason for existing.

The resistance to structural governance does not come from the engineers. Engineers prefer clear, automated constraints. They want to know the boundaries of the system so they can operate within them confidently. The resistance comes from the management layer that is being bypassed. They will argue that automated governance is too rigid. They will argue that human judgment is required for edge cases. They will argue that the system needs a “flexibility layer,” which is a bureaucratic term for a loophole they can control.

Organizations that successfully implement structural governance see a dramatic improvement in outcomes. The research notes a 5.3x higher success rate for organizations that pair structural change with cultural investment. The cultural investment is not optional. Changing the architecture without changing the organizational incentives leads the middle management layer to find new ways to insert themselves into the workflow. They will create new committees. They will demand new review meetings. They will colonize the new structure with the old politics.

The Financialization of Structural Dysfunction

The market is beginning to price the coherence gap. Hyperscalers are projected to spend $675 billion on AI infrastructure in 2026, a 63% year-over-year increase. This spending is not driven by proven demand. It is driven by a speculative bubble fueled by executive anxiety. Sixty-one percent of senior leaders report feeling extreme pressure to prove ROI on AI investments immediately. Fifty-three percent of investors expect positive ROI within six months.

These expectations are structurally impossible for most organizations. McKinsey research indicates that workflow-first design is twice as likely to see significant returns. Strong data integration delivers a 10.3x ROI compared to 3.7x for poor connectivity. Most organizations do not have strong data integration. Gartner projects that 60% of AI projects lacking AI-ready data will be abandoned by 2026. The organizations spending aggressively on GPU infrastructure today are often the same organizations that have underfunded their data integration for a decade. They are buying sports cars before they have paved the driveway.

The financial penalty for this structural incoherence is becoming explicit. Citi research identified a 30 basis point credit spread penalty for companies perceived to have uncoordinated AI spending. The debt markets are telling corporate leadership that throwing money at AI without a structural foundation is a risk factor. The equity markets tell the opposite side of the story. Organizations with “dual leaders” on measurement and infrastructure returned 41.38% compared to the S&P 500’s 29.40%, a spread of approximately 1,200 basis points.

The 1,200 basis point reward is not a reward for having good AI. It is a reward for having a credible narrative about AI governance. The market cannot see the inside of the organization. It can only see the org chart, the press releases, and the financial filings. An organization that appoints a Chief AI Officer and a Chief Data Officer and gives them joint P&L responsibility is signaling to the market that it takes structural coherence seriously. Whether the actual plumbing works is invisible to the equity analyst.

This creates a perverse incentive structure. Leadership is rewarded for performing coherence, not for achieving it. The CIO builds a beautiful slide deck outlining the federated operating model. The equity analyst upgrades the stock. The CIO receives a bonus. In the back office, the data engineer is still manually reconciling data from fourteen legacy systems when the “federated operating model” remains a PowerPoint presentation rather than an architectural reality. The market rewards the slide deck. The worker pays for the gap.

Minimum Viable Clarity as an Act of Organizational Courage

The research I presented prescribes a response called “Minimum Viable Clarity.” The concept borrows the logic of the Minimum Viable Product and applies it to organizational structure. Instead of demanding complete coherence before starting work, the organization demands just enough structural agreement to identify who owns the “Yes” and what defines value for the next ninety days.

Sociologically, Minimum Viable Clarity is an act of organizational courage. It requires a leader to stand in front of their peers and admit to the boundaries of their own knowledge. The standard corporate behavior is to hide ambiguity behind abstraction. “We will leverage AI to improve customer engagement” is an abstraction that hides ambiguity. “We will use a classification model to route 30% of tier-one support tickets to an automated resolver, saving an estimated $400,000 in Q3” is an act of clarity. It is specific enough to be wrong. Being specific enough to be wrong is what makes it valuable.

The resistance to clarity is fierce. Clarity creates accountability. An abstract goal makes failure impossible to prove. A specific goal makes failure undeniable. The leader who advocates for abstraction is not being strategic. They are protecting themselves from accountability. The leader who demands specificity is taking a personal risk. They are tying their reputation to a measurable outcome.

Amazon’s Type 1 versus Type 2 decision framework is the operational expression of this courage. Type 1 decisions are irreversible. They require extensive deliberation. Type 2 decisions are reversible. They should be made quickly by the people closest to the work. The framework is elegant in theory. In practice, it requires leaders to relinquish control over Type 2 decisions. Most leaders refuse. They classify routine decisions as Type 1 to maintain their position in the approval chain. The framework becomes another tool for centralizing authority, the exact opposite of its intended purpose.

Implementing Minimum Viable Clarity means forcing the organization to make decisions at the “Last Responsible Moment.” This is the latest possible point in the process where a decision can still be made without causing delay. Making a decision at the last responsible moment maximizes the information available. Most organizations make decisions at the first possible moment, which minimizes the information available. Early decisions are made under conditions of maximum ignorance. They are almost always wrong. The wrong decision then constrains all subsequent work, leading to the 250 requirement changes per cycle documented in the research.

The Question of Silent Consent

RAND Corporation research estimates that 80% of AI projects fail, with only one in five root causes being purely technical. BCG data from late 2025 suggests only 5% of AI initiatives create substantial value at scale, with 60% generating no material value at all. MIT NANDA research indicates that procuring specialist vendors succeeds roughly 67% of the time, compared to 33% for internal builds.

These statistics describe an industry failing structurally, not technically. The models work. The compute scales. The failure sits in the space between the executive’s vision and the engineer’s terminal. It sits in the quarterly business review where the data engineer is silent. It sits in the shadow IT environment where the business unit builds the report the platform cannot provide. It sits in the Dithering Zone where the change request is the only available form of agency.

The eight structural failure modes — shallow basins, perception gaps, cascading drift, siloed architectures, governance forks, broken highways, documented icebergs, and financial shears — are not sequential steps in a project lifecycle. They are concentric circles of organizational dysfunction. The outermost circle is the financial penalty. The innermost circle is the silent engineer in the back row.

When an enterprise launches an AI initiative, it is conducting a test of its own organizational integrity. The technology does not care about the politics. The model will train on the data it is given, regardless of whether that data represents the enterprise as the leadership imagines it or the enterprise as the engineer experiences it. If the data is fragmented, the model will be fragmented. If the requirements are contradictory, the model will output contradictions. The machine reflects the organization back to itself with brutal, mechanical honesty.

I left the AI in Products Community Event with this thought circulating. Nobody has figured it out entirely, and we are all in it together. The 95% pilot failure rate confirms this is not individual incompetence spread across hundreds of organizations. It is a measure of how deep the structural problem runs. The hallway conversations I had in May showed the same walls, in the same order, for the same structural reasons. That shared failure is the only reliable ground to stand on.

The question for any leader standing at the front of a conference room — whether it is a quarterly business review or a community event stage — is not whether the technology is ready. The question is whether the silence in the room represents agreement or resignation.

You can find my full slide deck on my website, by the way. Happy to get your thoughts. Critical as well — as always.

More actionable advice over at https://mohammed-brueckner.com/publications and even more good reads and all-time classics from great authors at https://itbookhub.com