Where does value come from in the age of cheap execution?
Writing Pierre-Alexandre Balland
When anyone can build anything, the scarce resources are vision and verification – and most organizations are overlooking both, creating a feeling of productivity at the individual level while limiting it at the organizational level.
For most of industrial history, the binding constraint was execution. The scarce thing was the capacity to actually produce the report, the model, the code, the campaign. Leadership was largely the art of marshaling that scarce capacity. Today, AI has taken that constraint and driven its cost toward zero. Yet here is the part we can easily miss: when a scarce resource becomes abundant, the constraint does not disappear. It migrates to whatever sits next to it. The new scarce resource is the human capacity to decide what is worth building – and to confirm that what got built is any good.
These are the two Vs that now determine business value: vision and verification. A massive amount of value is moving to those two layers, yet nearly all the investment is still pouring into the one that has been commoditized.
The verification bottleneck
My colleagues Christian Catalini, Xiang Hui and Jane Wu have given this a sharp formulation. They model the AI transition as a collision between two cost curves – an exponentially falling cost to automate execution, and a cost to verify what is inherently bottlenecked, bounded by human time and attention. The binding constraint on growth is no longer intelligence but human verification bandwidth: the capacity to validate outcomes, audit behavior and underwrite responsibility once execution is abundant.
The asymmetry is the whole point. You can parallelize execution without limit – spin up 10 agents, then 100, and the work multiplies. You cannot parallelize the checking. Attention does not fork. The reviewer reading the 100th plausible draft is the same finite human who read the first, with the same finite judgment, only now stretched thinner. So a gap opens between what your systems can produce and what your people can realistically verify – and that gap widens precisely as execution gets cheaper.
This reframes what your output actually is. Only the verified share of it is productive. The rest is plausible, but left unchecked, is indistinguishable from the verified part until something breaks. The word that keeps coming back when senior people describe this to me is ‘noise.’
“With AI I got much more work from my team,” one executive told me, “which looked great at first. Then I quickly realized I also got a lot more noise in my pipelines – and that is as bad as news can get when your biggest asset is your reputation.”
An organization that triples its output while holding verification flat has not tripled its production. It has tripled its noise and kept its real output roughly constant. This is a large part of why productivity statistics can remain unimpressive despite spectacular AI adoption numbers. At the individual level, everyone feels faster: drafts appear in seconds, analyses in minutes. At the organizational level, that acceleration piles up in front of verification gates that are not evolving at the same speed.
Cheap execution multiplies output that looks like signal, and it shows up in small but meaningful ways. Take AI meeting notes. They are often praised as a key tool of office life – and just as often dismissed as useless and a source of clutter. The summaries are usually impeccable: every attendee recorded, every action item captured, every decision neatly organized. Yet in one high-stakes negotiation I attended, the entire outcome hinged on just two seemingly minor parameters. The AI dutifully documented them, buried halfway through a perfectly structured list. Meanwhile, someone scribbling notes on a folded sheet of paper had circled those two game-changing elements in red. The machine produced a complete record – a useful one, even. But the human’s precise pattern recognition produced the most value – the critical judgment that would form the basis of subsequent decision-making.
This is the paradox of cheap execution. AI compresses the 95% of information that is routine, making everything appear equally polished. But in doing so, it often flattens the crucial 5% – the anomaly, the tension, the unanswered question, the subtle signal that deserves disproportionate attention. Intelligence is not the ability to capture everything, but the ability to recognize what should interrupt your thinking.
Consider the manager in the consulting or research world who has started to receive an endless stream of AI-generated dashboards: visually stunning, technically flawless and increasingly devoid of insight. Information production has exploded while information density has declined. The bottleneck is asking the right question before building anything. Organizations are recreating the PowerPoint problem all over again – the dashboard is becoming the deliverable, rather than the insight.
One pattern worries me more than almost anything else: consultants, analysts and junior staff increasingly using LLMs to find data points or perform analyses as if they were search engines or statistical software. They are not. An LLM is a language model: it predicts plausible text, as opposed to reporting verified facts. AI is an extraordinary tool for thinking, writing and exploring ideas – but when it comes to evidence, statistics, or analysis, it should never replace the underlying data or the human judgment needed to verify it.
The trap of verifying with more AI
The obvious objection is that this is what evals, test harnesses and AI-checking-AI are for. Let the machines verify the machines and the bottleneck dissolves, runs the argument. It is the right instinct, yet it has a hard limit.
Automated verification works beautifully where the criterion for correctness is cheap to specify: does the test pass, does the ledger reconcile, does the output match a known answer? It breaks down where judgment was the entire point: is this the right strategy, is this argument sound, is this the tasteful call, is this claim actually true rather than merely well-formed? Pointing a second model at the first does not escape this. It relocates the judgment to a new question – do I trust this verifier? – which is itself unverified, and now one layer harder to inspect.
The automated layer extends your reach across the easy ground and leaves the hard ground exactly as scarce as before. An augmented organization is not one that produces more. It is one where verification scales alongside capability, rather than being outrun by it.
There is a second-order version of this problem that the leaders I work with feel before they can name. When a senior delegates verification to a junior who delegates it to a model, the chain of people who understand the work thins out – and this is the very loop that used to produce the next generation of people qualified to verify. The cheaper execution gets, the faster you erode the human judgment you will need to check it, unless you rebuild on purpose the learning loops that cheap execution dissolves by accident.
This is the crisis hiding behind the disappearance of entry-level jobs. Execution was never just production: it was the apprenticeship through which juniors earned the pattern recognition that verification requires. Nobody becomes a good reviewer of financial models without having built a few hundred of them, badly, under supervision. If AI absorbs the building, organizations must deliberately reconstruct that ladder – structured review, verification rotations, deliberate exposure to failure cases.
Investing in vision and verification
The implication is almost the opposite of what most AI budgets reflect. You spent years and capital building the capacity to execute. That capacity is now cheap to rent. The discipline now is to invest in vision and verification with the same seriousness – and to understand that the two are not separate line items but two ends of one mechanism.
Start with vision, because it is the least understood of the two. Most people hear ‘vision’ as inspiration – a stirring sense of direction. Its real function in this economy is mechanical and far more valuable: vision is upstream verification. When you can build anything, prioritization stops being a planning exercise and becomes the hardest thing leaders do – because every project you greenlight is a verification bill you will personally have to pay later. If the acceptance criteria cannot be articulated, the project is not ready: not because it cannot be built, but because it cannot be checked.
Sharp vision does not just point the way. It minimizes the surface area that ever needs checking: fewer things built, narrower acceptance criteria, less plausible output flooding the queue. The leader who lets a thousand experiments bloom without this discipline has not empowered their team. They have manufactured a verification crisis.
Then treat verification as your last line of defense – and resist the urge to make it a bloated one. The goal is not an army of checkers; that simply moves the bottleneck without relieving it. The goal is to verify the things that matter deeply, and to design the work so that less of it needs deep verification in the first place.
In practice, in a world of intense competition and limited resources, this will most likely mean tiering. Reserve deep human review for what is irreversible, reputation-critical or client-facing, and accept lighter-touch sampling elsewhere. Build auditable loops: know which claims trace back to verified data and which are model-generated plausibility. And make verification visible in the organization: a named responsibility, a budget line, a skill that is hired for and promoted on, rather than an invisible tax absorbed by whomever happens to care.
Orchestration is the connective tissue here, and the reason good execution still matters enormously even as its cost collapses. The orchestrator’s job is to use cheap execution in service of the vision without flooding the verifier – structuring work so that what reaches a human for judgment is already filtered, already narrowed, already worth the attention it will cost. Execution becomes the thing you organize and structure.
The organizations that win the next decade will be those that understand, earlier than their competitors, that abundance on one side of the equation makes the other side priceless. Everyone can build anything. The advantage goes to whoever knows what is worth building – and can tell whether it’s any good. AI is the defining technology of our time, yet the bottleneck is not technological. It is organizational – and unlike the models, it cannot be rented. It has to be cultivated within.
Pierre-Alexandre Balland is chief data scientist at CEPS, visiting professor at Harvard Growth Lab, and a Duke CE educator
