Photo by Anne Nygård on Unsplash

I build the instruments organizations use to watch themselves. Dashboards, databases, pipelines, semantic models, reports, and the row-level security rules that quietly decide who sees which version of the truth.

The first thing that work teaches you is how easily a number stays green while the thing beneath it dies.

Modern leadership is fluent in alignment. Aligned to strategy, and to targets. Aligned to shareholder value, and aligned to the quarter. It gets spoken almost as a moral quality; the disciplined enterprise, every part of it pulling in a single direction. But alignment is a vector, not a destination. It tells you a system is pulling hard, but it says nothing about where. A machine can be perfectly aligned to the wrong objective and still be called a success by the people who built the dashboard.

That gap between being aligned and being well is the subject here. It is not a story about villains, but one about instruments, and about what happens when we hand the most powerful instrument we have ever built to institutions that have already confused the two.

The sovereign metric

Every metric begins as a stand-in for something much harder to measure.

Revenue stands in for value created, while margin stands in for efficiency in its creation, and retention stands in for whether people actually want the thing you make. The numbers become a map, and the territory they describe is the health of the institution: its capacity to keep producing something worth paying for.

The trouble is that the map is always cheaper to improve than the territory. You can raise a margin by building a better operation, or you can raise it by deferring the maintenance that next year’s operation depended on. Both arrive on the chart looking identical.

And so, gradually, the organization learns to optimize the map. The number that was meant to describe health becomes the thing pursued in place of it.

Call it the sovereign metric: the moment the scoreboard stops reporting on the kingdom and starts governing it. Other fields have a name for the underlying failure – a measure that becomes a target stops being a good measure – but the version that matters to a leader is plainer than that.

You stop being asked whether the institution is healthier. Instead, you start being asked whether the number moved.

Alignment is not effectiveness

The danger is not that these numbers are fake. They are real, and they matter. Earnings, margin, and the credibility of a forecast are all things a serious institution has to answer for. The quarter does honest work, too: it disciplines complacency, punishes the lazy management that would otherwise coast on a good story, and forces a recurring, uncomfortable answer to whether anyone still wants what you make. A company with no scoreboard does not become humane; it becomes unaccountable.

The danger is subtler. An organization can answer every one of those questions correctly and get measurably worse at the thing it exists to do.

Alignment to shareholder value reliably produces motion: expense discipline, headcount reduction, asset-light operating models, acquisition-driven growth, tax optimization, and pricing leverage. None of these is a crime, and each, in the right context, is exactly the right call. But none of them, on its own, means the business has become better at creating value. They mean the scoreboard improved. The two are easy to confuse precisely because the scoreboard was built to make them look the same.

A goal can be clear, measurable, and financially rewarded while being structurally corrosive. The clarity is part of the trap. A legible goal will always crowd out an illegible one, because someone is accountable for the number and no one is accountable for the thing that has no number.

Trust does not have a cell on the dashboard. Neither does institutional memory, nor the operational slack that lets a system absorb a shock it didn’t forecast. So those are the first things spent, because they cost something now and pay back later, to someone else, in a currency that doesn’t impact the score.

I have watched this from inside the measurement layer. Build a clean dashboard and a capable team will manage to it with real skill and real dedication. They will hit the number. And the work that never made it onto the board – be it documentation, cross-training, or any of the other unglamorous requirements for resilience – quietly goes unfunded, because the instrument that decides what gets attention was never given a way to see it.

I have done it myself. With one capacity overloaded past the point of usefulness, I moved a set of reporting workspaces onto another to relieve it. It was the right call, and the load recovered almost at once. What the load number did not show, however, was that the two capacities had been stood up in different regions, and that the semantic models, by virtue of their storage format, were bound to the region they were first published in.

Relieving the metric took an afternoon. Paying for it – republishing and relinking every region-locked model by hand – took the weekend, and it was only that expensive because of an earlier convenience that nothing on any dashboard had ever been asked to price.

Extraction is easier than development

There are two ways to make a number go up.

You can develop: increase the system’s actual capacity to create value. Build a better product, deepen a skill, or make a process more resilient than it was.

Or you can extract: convert capacity that already exists into financial return. Acquire the competitor instead of out-building them. Limit supply to protect pricing. Push the cost of dysfunction onto the people least able to refuse it, be they employees, suppliers, customers, or the community that hosts the plant.

Either will move the quarter; the difference is what happens after.

Development is slow, capital-intensive, and inconveniently honest: its payoff often arrives long after the people who authorized it have rotated out. Meanwhile, extraction is fast, defensible, and cleanly attributable to whoever happens to be standing there when the results land. So the scoreboard, quarter after quarter, makes a consistent recommendation. Often, it recommends extraction; less because anyone decided to hollow the place out, than because extraction is simply the cheaper way to move the only number the institution agreed to watch.

The absence of sacrifice is the tell

Here is the part the dashboard cannot show you. A good quarter looks the same whether it was built or extracted. You cannot see stewardship in the number, because the number records the result and discards the cost.

You see stewardship only in the quarter where someone chose the slower, more expensive right thing and absorbed the difference. Accepting a lower margin to hold quality when costs increased. Investing in people before the return is visible, to ensure they are ready for the next challenge. Telling shareholders an uncomfortable truth before the market has decided to reward honesty. These are acts of sacrifice in the precise sense: doing the costly thing because it is right, rather than because the scoreboard asked for it.

A leader who never asks capital to wait is not balancing stakeholders. They are merely deciding which of them are cheapest to spend.

The captured scoreboard will reward them for it, because it has no way to distinguish a leader who grew the organization into a good number from one who simply found something unmeasured to remove. The absence of sacrifice is the tell, but it is invisible to any institution that has let the metric become sovereign.

The same logic, about to be amplified

The first thing I learned about Excel macros is that they repeat exactly what you tell them to do. That can spare you a thousand small manual changes. It can also leave you choosing between starting over and correcting the thousand mistakes that happened in the screen-flickers before you managed to hit Esc.

It took me longer to accept that the macro was not malicious, only obedient. It had no opinion about my intent. It amplified my instruction, and my instruction was wrong.

That is the whole of what follows, which is exactly why it should worry a careful reader. Automation does not import new values into a system. It executes the ones already there, faster and at a scale that removes the friction that used to let you catch the mistake.

The usual frame for AI risk is the machine’s inner values, focused on whether the model will come to want the wrong things; and that is a real question. But it is not the first one, and for most institutions it is not the operative one.

The machine does not have to want anything. It enters the organization the way every tool enters: through procurement, productivity mandates, vendor roadmaps, budget cycles, and the renewal conversation twelve months later. It gets funded, integrated, and kept on the strength of a single question – conveniently, the same one every other tool answers to. Did it move the number?

An institution that has defined usefulness as cost removed, labor compressed, throughput raised, and risk transferred will reliably select for the AI that does those things. Less because the system believes in shareholder primacy – the model itself believes in nothing – than because it is being chosen, again and again, by something that does.

The selection pressure is entirely external to the model. You do not need a misaligned intelligence to produce a misaligned outcome. You need only an ordinary one, evaluated by an institution that already mistook its scoreboard for its purpose.

So the machine does not corrupt the institution; that frame reverses the direction. It helps the institution become more fully itself: faster, with different cost centers, and more thoroughly committed to whatever it was already optimizing.

The concern should not be that something alien arrives carrying strange values. It is that something tireless arrives and accelerates the values already in the building.

The scoreboard is the alignment layer

If demand is where the alignment problem actually lives, that is not cause for despair. It is a location, and a location is something an architect can work with.

The thing that decides which AI survives inside an institution is the institution’s own definition of useful: its metrics, its procurement criteria, and the standards it holds a deployment to six months after the pilot.

These are usually treated as administrative furniture, but they are not. They are the alignment layer, in the same way a permission boundary in an agentic system is not compliance overhead but load-bearing architecture. What you choose to count is what you will get more of, and you are about to hand that choice to something that optimizes far harder than any team you have ever managed.

Which means the leaders who handle this well will not be the ones with the most ambitious AI strategy. They will be the ones who repaired what they reward before they bought the thing that optimizes for it.

That repair is concrete work, not a posture. An organization that wants intelligence to serve a decade has to put at least one decade-scale number on the board and give it real weight: regretted attrition, the share of departures the organization actively wanted to keep, rather than raw headcount cost. Time-to-recover from an unforecast shock, measured and tracked like any other reliability metric. The ratio of maintenance and remediation to new build, so that deferred debt has a cell of its own instead of hiding inside next year’s margin.

None of these is perfect, and that is the point: each one prices something the quarterly scoreboard was built to ignore. Failing even that, the minimum viable discipline is to refuse the quarter the right to be the only number with a vote.

Stewardship has always been the willingness to leave some value unmeasured and protect it anyway. That discipline was survivable as a weakness when the optimizer was human and slow, but it stops being survivable when the optimizer is neither.

The intelligence we reward

Intelligence will inherit the incentives of whoever buys it long before it sustains any values of its own. That is the alignment problem underneath the one the industry usually argues about: not whether the machine can be made to obey, but whether obedience to our current institutions would be a good thing at scale.

The kingdom can survive a bad quarter; most already have. The open question is whether it can survive being governed by the conviction that the quarter is the kingdom, now that the governing is about to be systematically automated.

If I have learned anything in the years since LLMs took over the news cycle, it is that we will get the intelligence we reward. The work, before any of it is bought, is to be honest about what we have been rewarding.

And then to build a scoreboard we would not be ashamed to watch something optimize.