AI's returns are earned downstream, in governance

A set of figures has been circulating in AI strategy decks all year, and they lean one comforting way. Companies that rebuilt their cost structures around AI are pulling far ahead: 3x greater cost reductions, 1.6x higher operating margins, and 2.7x the return on invested capital of their competitors. The Boston Consulting Group put those numbers out in a 2026 report on building an AI-first cost advantage. The lesson most people draw from them is clean. Building software is getting cheap, so the human edge is now judgment. Taste. Knowing what to build.

The plain version, if you tuned out the news this quarter: AI can now write a lot of the software and do a lot of the work a company used to pay people for. Enterprises are deploying that cheap output at volume. The BCG study says the ones who restructured around it are winning on cost and profit by wide margins. And a quieter strand of research says the difficulty AI removed from building doesn’t actually leave the building.

That second part is the whole story, and it changes what the numbers mean.

The complexity AI strips out of building software doesn’t vanish. It moves downstream and lands as a governance problem of roughly the same weight, so the enterprise returns everyone reads as proof of better taste are mostly the return on something duller: the capacity to trust AI output at scale without it quietly going wrong.

Read the 2.7x again and ask what mechanism produces it. It isn’t taste. Nobody earns 2.7 times their competitor’s return on capital because their product managers have sharper opinions about what to build. Taste improves your hit rate, and hit rate is a slow, noisy lever. It shows up over quarters of shipped bets, a few more winners than the other team managed. A step change in return on capital across an enterprise comes from somewhere else: running cheap, AI-generated volume through the business without the error rate eating the savings. That means checking the output, monitoring it in production, and knowing who answers for it when it’s wrong. That is governance capacity, and it’s what the number is actually paying for.

So the headline evidence points at a different edge than the headline claim. There are two moats here, and they keep getting mistaken for one.

One is a discovery edge. It’s individual, it sits upstream, and you exercise it by knowing what to build. It’s scarce because domain expertise is scarce, and it’s real. Marty Cagan of Silicon Valley Product Group has argued all year that collapsing delivery costs finally expose product discovery as the bottleneck that was always there. Anthropic’s own analysis of 400,000 work sessions found that domain knowledge, not coding background, predicts who gets leverage out of AI.

The other is a governance edge. It’s organizational, it sits downstream, and you exercise it by being able to trust AI output at volume. It’s scarce because verification capacity has to be built, and building it lags badly. Research on AI cost structures (arxiv paper 2506.22440) describes the difficulty re-emerging past implementation as an “accuracy ceiling”: the point where you can’t push more AI-generated work through without more of it being wrong than you can catch. New bottlenecks, new operational risks, all downstream of the cheap building everyone’s celebrating.

These are not one moat wearing two hats. They ask for different investments, a great product manager at one end and an instrumentation layer plus a new reliability role at the other. They sit at opposite ends of the workflow, and they fail for different reasons. Conflate them and a team can believe it’s winning on taste while it’s quietly losing on governance, pushing out more at scale and mistaking the volume for value. That is the exact failure the accuracy-ceiling research warns about.

This is more than one study’s quirk, because the same downstream signal shows up from unrelated directions.

Harvard Business Review researchers, writing in early 2026, found that what blocks enterprises from turning AI pilots into real transformation is organizational: governance, incentives, and trust, well ahead of model quality. They call the pattern “pilot-rich, transformation-poor,” and read it as a coordination failure rather than a technology gap. Separately, the analytics firm Amplitude has been naming a new category it calls agent analytics, a measurement layer for whether an AI agent actually gave a user a useful answer. The context for it: Gartner figures show task-specific agents jumped from under 5% of enterprise apps in 2025 to 40% in 2026, an eightfold expansion in a single year, while the ability to measure whether any of them work well lagged far behind. And in day-to-day product work, research on delegating admin tasks to agents finds the same shape. Handing work to an agent frees up a product manager only when accountability is drawn explicitly. Leave it fuzzy and ownership blurs, until nobody’s sure whether the human or the agent owns the quality of what shipped.

Four separate strands, all pointing at the same downstream layer as the place the constraint actually went.

The honest counterargument is that the taste camp isn’t wrong, and it isn’t. Domain expertise really is a durable edge, and knowing what to build really did get more decisive as building got cheap. Cagan and the Anthropic data make that case well. But it explains which of your bets land. It does not explain a near-tripling of return on capital across a whole company. The taste story borrowed a number that measures something else and held it up as proof.

Which leaves a question sharper than “is taste the moat,” and a testable one. Do the BCG-scale advantages track with a company’s discovery quality, its product judgment and hit rate, or with its governance maturity, its instrumentation depth and how cleanly it assigns accountability? Find a firm with mediocre product taste and heavy governance investment. If it still captures the cost and return advantage, the moat is downstream, and every “hire for taste” plan built on these numbers is aimed at the wrong end of the pipe.

That’s worth settling before you write next year’s headcount plan around the more flattering half of the story. Taste is a hire you can make this quarter. Governance is a rebuild, and the numbers being quoted to justify the hire were paying for the rebuild all along.