Brief
|
|
At a Glance
On a Monday afternoon, the capital projects team at a large mining company spent an hour setting up the review of its most critical investment projects: which milestones to test, which data to trust, what variances mattered. Then, the agents went to work. By Tuesday morning, they had reconciled tens of thousands of documents against cost and schedule data, and the first pass was on the table. The work was no longer waiting to be done. The question for the team was which of the problems the system flagged were real. When results can be checked quickly and reliably, AI can do the work that people used to do. Call that fast and trustworthy read on whether the work was done correctly a signal. Where a signal exists, AI has something to learn from, and the company has a way to judge how well AI is performing. Everyone has access to the same AI tools. Competitive advantage accrues to the companies that move the frontier by building signals no one else has. Those signals settle many decisions competitors still have to judge, freeing these companies to concentrate human judgment on the most valuable decisions that remain uncertain. The judgment premium goes to the companies that make those calls better than their competitors do. And it compounds: the data, workflows, and lessons learned from those calls become proprietary intelligence that pushes the frontier further and makes the advantage harder to copy. What can AI do? It depends on the signalApplying AI to work is a business transformation, and the right place to start is the work that matters most to growth, cost, or customer value. But value tells you where to look, not where AI can actually advance. The intuitive test is difficulty, but tasks that look equally difficult to a trained professional can land on opposite sides of what AI can do. A signal is the better test: Can someone tell quickly and reliably whether the result is correct? Across any business, work falls into three groups. Some work already has a fast signal, and that is where AI is advancing most quickly: Code runs, or it does not; tests pass or fail. Engineers turned that feedback into automated evaluations that they could run again and again. Other work is structurally slow. A market entry succeeds or fails, but the result is tangled with luck, timing, and hundreds of other choices, with no way to repeat the experiment. Most work, though, falls into the dangerous middle. Nobody has measured it yet (which is not the same as unmeasurable). Judgment feels irreducibly human until someone finds a way to score it. For example, underwriters' sense of risk looked irreducible until someone built the signal.
The frontier between work that can be checked and work that cannot is not something companies just encounter; they move it. They build a proxy, a fast measure that stands in for a slow outcome; they encode an expert's unwritten standard into a rubric against which a model can be scored; and they use simulation within which the environment can be reproduced and success graded. Each codifies a judgment that was never written down, turning it into feedback from which a machine can learn. The honest version of "AI can't do that" is often "no one has built the signal yet." But building a signal does not make it true. Goodhart's Law warns that once a proxy becomes the target, it can drift. Recognizing that drift before the damage is obvious is exactly the kind of judgment AI cannot supply. Where you move the frontier is the most consequential choice. Building a signal is slow and expensive, so it belongs where the work is what makes you distinctive. An insurer that builds a signal to check data entry on a claims form has built a tool any competitor can buy. But a signal built from years of proprietary underwriting history matched to what those policies eventually cost can grade AI’s risk selections against outcomes no competitor can see, and that signal gets sharper with every policy year. The move is not new. Consider how lenders with no credit files on their customers were able to build a proxy out of mobile operator data, reading airtime top-up patterns as an indicator of who could repay. A mobile credit-scoring industry formed around the score. What AI changes is the speed and range of work for which companies can build such signals. Building the signal is one thing; keeping it is another. Climate Corporation was a crop insurance start-up that had to price weather risk field by field. It built soil and climate measurement data across US farmland where almost none existed. Monsanto acquired Climate Corporation for about $930 million in 2013, and Bayer inherited it when it bought Monsanto in 2018. More than a decade later, Bayer traces its current AI capability to that deal, and its CIO argues that the data, not the model, is the part competitors cannot copy.
That is the test to apply before investing: Will you own what the signal produces? Durable competitive advantage depends on owning the data, workflow, and feedback loop (i.e., the proprietary intelligence) that together make the system improve with use. Where companies should concentrate human judgmentIn our March 2026 cross-industry survey, 91% of respondents said that AI helps them be more productive, and 53% reported higher-quality work. Only 27% said it helps them do new types of work. But as companies move work across the frontier, something predictable happens. Production stops being the constraint, and judgment becomes the bottleneck. Anthropic reports that as agent-written code moved through its own organization, human review became the new limit. A signal can grade the work, but it cannot grade itself. The judgment premium goes to companies and the people inside them who work faster but can still tell whether the work is right. Agentic AI does not remove accountable judgment. In our work with clients, agentic AI consistently concentrates judgment into four human responsibilities surrounding production: intent, constraints, validation, and stewardship. AI can help with each, but it cannot decide what to pursue, what to permit, what to accept, and what to keep. These responsibilities are not new, but they used to be undefined, or thought of as the natural output of human experience. Now, they have to be assigned separately. Most failures we’ve seen were not AI model failures but instead trace back to one of these four responsibilities going unassigned: The system solved the wrong problem, crossed a line nobody drew, produced something plausible that no one checked, or did work no one owned. Intent is defining the problem and what a good answer looks like. Get it wrong, and the machine solves the wrong problem faster and more persuasively. Constraints are the limits: what the system may use, what it may decide alone, where it must stop. The hardest constraints were once tacit. Stripe had to write down a rule that had never appeared in its access controls: Data from two unrelated customer contexts must not appear in the same analysis, even in cases in which one person is cleared to see both. AI can follow rules you write down; it cannot inherit values you never taught it. Validation is deciding whether the output is right. Not whether it is polished or internally consistent. Right. In an experiment with 791 professionals at Procter & Gamble, AI raised the quality of the ideas generated; the value of human judgment concentrated on selecting among them. And because agents generate more output than humans can review, validation is also an attention problem: Someone has to decide not just if each output is right but also whether the standard for what qualifies as right still holds. Stewardship is deciding what the system keeps, both within a workflow and across the institution: corrections, standards, and the judgments worth preserving. In our agentic work, we find that stewardship takes roughly a third of the human effort. OpenAI helped an accounting firm build a tax preparation agent and recorded what it proposed, what the accountants changed, and what was finally filed. The share of returns reaching a baseline quality threshold went from 25% to 86% in six weeks. The agent didn’t get smarter on its own. Stewardship turned human corrections into a compounding asset rather than a transcript nobody reads. A company that runs a learning loop like this for a year starts its next project ahead of a competitor starting from zero, and the gap widens with every cycle. This is how proprietary intelligence is built. Human teams tacitly understand these four responsibilities, but different teams have different shared norms, context, and assumptions. AI agents cannot navigate those differences without clear rules. In a recent Anthropic survey of more than 500 technical leaders in the US, 57% were running agents on multistage workflows, but only 16% were running agents across processes spanning teams. The implicit must become explicit for the company to operate at agent speed. An Anthropic survey of more than 500 technical leaders found57
running agents on multi-stage workflows 16
running agents across processes spanning teams Two questions remain. The first is how much an agent can decide on its own. An ability to score sets the automation frontier. Reversibility sets the autonomy frontier: whether errors can be caught and undone, and what they cost when they cannot. Where errors can be caught and reversed, let the system run, and audit the exceptions. Where they cannot, use the signal to accelerate the review rather than remove it. The second question is how to develop human judgment. Before AI, judgment was simply a byproduct of work. Junior work created output and expertise at the same time. AI separates the two. Firms book productivity gains immediately; the expertise losses arrive later and without warning because the people who would notice are the same ones whose judgment is still masking it. Protecting the human judgment on which competitive advantage relies requires companies to deliberately redesign both work and learning. Where to start building competitive advantage with AIStart where the value is, then test the signal. Make the bet explicit: Which work matters most to your economics, and why? Value tells you which work to transform. Whether results can be checked tells you how far AI can go. Take one of your most important workflows, and break it into tasks because the frontier runs inside a workflow rather than around it. Deploy where results are already checkable; do not pour agents into work no one can grade. Build the signal where it is worth owning, and check that it is yours. Pick one high-value workflow in the dangerous middle and make a deliberate bet with a defined horizon. Before funding it, ask whether the resulting loop would be something a competitor could buy. If it is, you are building efficiency, not advantage. Assign the four responsibilities to people by name, not to a committee. Intent, constraints, and validation belong to the one person who answers for the output. Stewardship works at two levels: Someone owns what a workflow keeps between cycles, and someone owns what the institution keeps across workflows. If any of the four responsibilities has no clear owner or is owned by committee, you have found the failure mode before it finds you. Before scaling agents, build the environment they need. This requires a shared state so that:
Then, measure three things: speed, which tells you that the tools work; quality, which tells you that the team design works; and compounding, which tells you whether the system is actually getting better or just accumulating records. Which judgments should stay human as AI scales?The work AI can score will get easier to commoditize; what will not is deciding which work matters, which signals are worth building, and which judgments must stay with people. Find where your advantage really sits, build the signal there, and put your best people on the calls that the signal cannot make. Everyone will have AI. The advantage comes from which signals you build and own, and where you put the judgment that remains. |