You Are Measuring the Wrong Thing

Speed is the metric most AI dashboards track. It is not the one that determines which side of the Superstar Economy divide a firm ends up on.

Executive Summary

Across the management pyramid, just 13% of management time goes to genuinely good decision-making. The rest is execution: writing, coordinating, processing, preparing. Most AI deployments accelerate that 87% — and leave the bottleneck untouched. Measuring AI performance by throughput is measuring the wrong thing. Three structural failures compound beneath any productivity metric: the decision bottleneck that AI does not fix, the 85–95% of executive intent that evaporates before it reaches execution, and the 42% of institutional knowledge that walks out the door with every senior departure. Conventional AI deployment makes each of these more expensive, not fewer. The organisations that understand this are building a structural advantage. The rest are building a faster version of the same problem.

The Boards Are Tracking the Wrong Number

Most AI governance reports published in the past two years measures roughly the same things: tasks automated, hours saved, headcount rationalised, output volume increased. These are real numbers. They are not the numbers that determine whether a firm ends up in the top quintile of the Superstar economy or spends another decade earning near-zero economic profit in the middle three.

The actual number — the one that separates firms that compound their advantage from firms that compound their inefficiencies at higher speed — is decision cadence: the rate at which genuine human judgement converts into committed, executed action across an organisation.

No AI dashboard currently tracks it. Most boards do not know what it is. That is precisely the problem.

The Bottleneck No One Is Measuring

Start with the data. Across the management pyramid, a weighted analysis of McKinsey's research on roughly 1,200 executives produces a conservative finding: approximately 30% of total management time goes to deliberating and deciding; approximately 70% goes to execution and administration (McKinsey, Decision Making in the Age of Urgency, 2019; McKinsey, Stop Wasting Your Most Precious Resource: Middle Managers, 2023).

That 70/30 split is the management reality in most organisations. It is also, emphatically, not a split between two equally tractable problems. Exhibit 1 — The Management Time Stack

The 70% — the writing, the coordinating, the processing, the preparing — is where AI performs best. Agents can draft, synthesise, coordinate, and execute at a speed and volume no human matches. This is the work AI was built for, and the efficiency gains are real. But here is the structural problem: accelerating the 70% does nothing to the 30%. The decision bottleneck stays exactly where it was. You now have a faster machine producing better-prepared material for the same underused judgement capacity.

Of that 30% of management time nominally spent deciding, research shows more than half is used ineffectively — poorly structured, poorly informed, poorly executed decisions that consume deliberation time without producing useful outcomes (McKinsey, 2019). The net result: genuinely good decision-making occupies less than 13% of a manager's working week.

Add AI to this picture and the arithmetic becomes uncomfortable. Deploying agents that dramatically improve the quality of execution inputs — better research, better synthesis, better drafts — while the decision process remains structurally unchanged produces one outcome: it raises the quality and volume of material for the same 13% of genuinely good decision-making to work through. The bottleneck is now better-fed. It is still a bottleneck.

Decision-making is the new rate-limiting factor. That is the thing to measure. And the structural response to a 70/30 split is not better tools for the 70. It is an architecture that inverts the ratio.

Three Structural Failures Beneath the Throughput Metric

The decision bottleneck is the most immediate problem. It is not the only one. Two further structural failures compound beneath every throughput metric, and both are invisible to standard AI reporting.

The strategy loss problem. When a CEO or senior leader sets a direction, research suggests 85–95% of that original intent is distorted, misunderstood, or quietly reinterpreted by the time it reaches the people responsible for acting on it. Six mechanisms cause this: misunderstanding at each layer of transmission, reframing through individual priorities, filtering of inconvenient elements, resources that do not follow the stated direction, no mechanism to measure whether the intent has landed, and drift over time as operational pressures displace strategic direction. Adding AI to a structure with this problem does not fix any of the six mechanisms. It accelerates the execution of whatever misinterpretation has already formed.

The institutional knowledge drain. When an experienced person leaves an organisation, they take roughly 42% of their team's institutional knowledge with them. The decision-making capability embedded in that knowledge — the judgement about which approaches work and which do not, the context for why certain rules exist, the pattern recognition that comes from years in a specific function — cannot be transferred through documentation or replaced by a new hire at speed. Replacing a senior person typically costs 150–200% of their annual salary and takes up to two years to restore operational effectiveness. Organisations losing experienced decision-makers are not just losing people. They are losing decision-making infrastructure. AI deployment does not address this either — unless the deployment is structured so that decision-making knowledge is embedded in the human-agent relationship rather than held by the individual alone.

These three problems — the decision bottleneck, the strategy loss, and the knowledge drain — compound each other. An organisation operating with 13% genuinely good decision-making, losing 85–95% of strategic intent in transmission, and periodically haemorrhaging the institutional knowledge that makes whatever decisions do get made reliable is not a decision-making organisation. It is an execution machine pointed in an uncertain direction, now running faster.

Why Speed Makes the Wrong Thing Worse

The appeal of measuring AI performance by throughput is not irrational. Throughput is visible. It is easily quantified. It generates the numbers that populate ROI analyses and board presentations. Faster output is demonstrably better output in any process where the output's quality is not a function of the decision-making that preceded it.

The problem is that in knowledge work — which is what management is — output quality is almost entirely a function of the decision-making that preceded it. A well-executed bad decision is more damaging than a slowly-executed bad decision. It commits more resources, faster, to the wrong direction. It locks in consequences before the error is visible. It compounds.

This is what conventional AI deployment does when it accelerates the 68% without addressing the 13%. It raises execution velocity without raising decision quality. The mistakes get made more efficiently. The strategy loss reaches the front line faster. The institutional knowledge that would have caught the error is still walking out the door at the same rate.

The metric "hours saved" does not capture any of this. Neither does "output volume." Neither does "tasks automated." All three measure speed. None of them measure the variable that determines whether the speed is compounding advantage or compounding error. Compounding advantage requires a structure where decisions improve with repetition — not one where the same bottleneck processes the same ratio of good and poor judgement faster.

What to Measure Instead

Measuring decision cadence — the rate at which genuine human judgement converts into committed, executed action — is not a soft aspiration. It is a measurable structural property of an organisation. It has three components.

1. Decision Cadence measures whether the human is genuinely deciding or ratifying. It tracks deliberation time, override rates relative to error rates, and the quality of justification provided for consequential choices. A manager who approves every AI recommendation within thirty seconds, never overrides, and cannot reconstruct the rationale for their decisions post-hoc is not making decisions. They are documenting them.
   

2. Execution Fidelity measures whether the AI is doing what was actually intended — not just the letter of its instructions, but the spirit of the decision that generated them. An agent that executes precisely what it was told to do while drifting from the intent behind the instruction is producing technically compliant output for the wrong purpose.
   

3. Governance Effectiveness measures whether strategic intent and execution are staying aligned in real time — not whether last quarter's decisions were reviewed, but whether today's execution reflects this morning's direction.

These three components, multiplied together, produce what we call Cogenic Yield: the single measure of how much of an organisation's decision-making potential is actually being realised.

Critically, none of these metrics can be grafted onto a conventional structure. The ability to measure them is itself a product of structural design. An organisation that has not separated deciding from doing cannot reliably measure whether deciding is happening.

The Diagnostic

The practical question is whether any given AI deployment is addressing the decision bottleneck or simply accelerating the work that surrounds it. Four questions apply:

1. Has the deployment changed the ratio of time managers spend genuinely deciding versus executing? If the answer is no — if the 70/30 split remains roughly where it was — the deployment has addressed the 70% without touching the problem.

   
   

2. Has the deployment changed how strategic intent travels from the point it is set to the point it is acted on? If intent still moves as informal communication subject to reinterpretation at each layer, the 85–95% loss is still occurring, regardless of how efficiently the execution that follows is managed.

   
   

3. Has the deployment changed what happens to institutional knowledge when experienced people leave? If the answer is no — if decision-making knowledge is still held by individuals rather than embedded in the structure — the 42% drain is still active. A deployment that embeds knowledge in the human-agent relationship rather than in the human alone changes this calculus — not by preventing departures, but by ensuring what remains is structural rather than personal.

   
   

4. Is the deployment being measured by throughput, or by decision cadence? If the primary metrics are tasks completed, hours saved, and output volume, the deployment is measuring the right things about the wrong problem.

No deployment that fails all four questions is a structural change. It is a speed upgrade for an unchanged structure — useful, demonstrably valuable on the metrics it is measured by, and insufficient for a Superstar economy in which the differentiating variable is no longer what a firm has but how well it decides and acts.

Implications

For strategy. Every AI investment should be tested against a single question: does this deployment improve decision cadence, or does it accelerate the work surrounding decisions? The two are not the same. The organisations making the distinction now are the ones that will be structurally positioned in three years. The ones that do not will have spent considerably more on AI by then and will be no closer to the competitive variable that determines outcomes in a winner-takes-all market.

For organisational design. The unit of design in a Cogenic Organisation is not the role. It is the Cogene — the human-agent pair structured so that the agent absorbs the execution work, freeing the human to spend 70% of their time on genuine judgement rather than 13%. That is not a technology change. It is a structural change to how work is organised. (The Cogene is examined in full in The Unit of Work Just Changed for the First Time in a Century.)*

For measurement. Boards that continue to measure AI performance by throughput are reading last year's decisions in next year's report. Decision Cadence is a forward-looking indicator: it tells a board not where the organisation has been but where its decision-making capability is heading. That is the number that predicts which side of the Superstar economy divide the firm will occupy.

Conclusion

Speed is not the wrong goal. It is the wrong primary metric. An organisation that decides well and executes fast compounds its advantage with every decision it makes. An organisation that executes fast without deciding well compounds its errors at the same rate. The distinction between the two is invisible to any dashboard measuring throughput, and it is the only distinction that matters in a competitive environment where every firm has access to the same AI tools.

The firms that understand this are not asking how much faster AI can make them. They are asking how much better it can make their decisions. The answer to that question is not found in the model. It is found in the structure that governs how the model and the human work together — and whether that structure is designed to make decisions genuinely better, or merely to document them more efficiently. The gap between those two questions is the gap between a productivity upgrade and a structural advantage — and it widens with every passing quarter in which the wrong metric is being optimised. Learn more www.cogenics.ai Contact Cogenics hello@cogenics.ai Sources

• McKinsey & Company. Decision Making in the Age of Urgency. April 2019.

• McKinsey & Company. Stop Wasting Your Most Precious Resource: Middle Managers. March 2023.

• Porter, Michael E. and Nitin Nohria. How CEOs Manage Time. Harvard Business Review, July–August 2018.

• Cogenics. The Deciding Organisation. Version 1.0, June 2026.