Argyris put the failure in the room.

This entry was prompted by “Skilled Incompetence,” an essay Frank Pizzuta, chief operating officer at Inveniam, published on Medium on 21 July 2026. His argument: long experience builds a pattern library and an attachment to the methods that built the career. Ask a model a question shaped by that attachment and it returns a thorough answer confirming it, because that is what a good answer to that question looks like. Firms read experience as cost, swap the expensive person for a cheaper one plus a model, and every dashboard number improves. His line: “There is no line item for the capacity to notice the company is solving the wrong problem.”

The essay takes its title from Chris Argyris, which is where we can add something. Argyris’s “Skilled Incompetence” ran in the Harvard Business Review of September–October 1986. His definition sits in the opening paragraph: managers “use practiced routine behavior (skill) to produce what they do not intend (incompetence).” He was writing about executives who had become exceptionally good at avoiding conflict with each other.

It opens on a top team that met for months to write a strategy and produced lists. Argyris had each manager write a real conversation in two columns: what was said on the right, what was withheld on the left. The left column was full. The skill lived in the right one — a message softened so the room stayed comfortable, with nobody naming the softening. He called the settled patterns organizational defensive routines, “designed to avoid surprise, embarrassment, or threat,” which also “prevent learning.” His recipe ends by making the undiscussability itself undiscussable.

The claim now circulating under that title — that high performers cannot question a goal their identity rests on — comes from a later paper. “Teaching Smart People How to Learn” ran in HBR in May–June 1991, on fast-track consultants. “Because many professionals are almost always successful at what they do, they rarely experience failure. And because they have rarely failed, they have never learned how to learn from failure.” Their capacity to learn, he wrote, “shuts down precisely at the moment they need it the most.”

The two papers belong together, and fusing them is fair. One distinction is worth restoring. Argyris located the failure between people: the 1986 paper is about what a room makes unsayable. The rendering now in circulation has moved the mechanism inside a single head, where it becomes a matter of self-awareness. Boards have thin purchase on self-awareness. They have complete purchase on rooms — who sits in them, what the papers contain, who is safe to speak.

The concept that carries the weight came with Donald Schön, in Theory in Practice (1974) and Organizational Learning (1978). Single-loop learning corrects the error and keeps the goal. Double-loop learning puts the goal in question. Argyris’s image is a thermostat: one that turns the heat on below 68 degrees is single-loop, one that asks why it is set at 68 is double-loop. A model runs the first loop beautifully. The second is closed to it, because the governing variable arrives inside the prompt.

This is why fluency is the signal to watch. A calculator fails loudly. A spreadsheet fails visibly, the broken reference sitting on the page. A model fails fluently, in the register of a correct answer, structure and confidence and citations all intact. Its errors arrive dressed as its successes.

Does a model confirm the framing of a question it receives? The evidence has moved past reasoning. Cheng and colleagues put both sides of one moral conflict to eleven models; in 48% of cases the models affirmed whichever side the user had adopted (ICLR 2026). Chen and colleagues gave five frontier models a false premise about drug equivalence; compliance ran as high as 100%, in models that knew the premise was false (npj Digital Medicine, 2025). Dahl and colleagues found GPT-4 accepted an invented Supreme Court dissent and answered on it around 69% of the time (Journal of Legal Analysis, 2024). The boardroom instance remains unmeasured; carrying it there is our inference.

Our register names six modes by which judgment drains from an institution. The third is inquiry groupthink: every seat briefed to the same depth from the same fluent source, so the room stops generating alternatives and its unanimity certifies nothing. The chief risk officer owns it. The essay describes it running inside one person.

The board-level version is ours. A board leans on several supposedly independent readings: management’s analysis, internal audit, the external auditor, and whichever adviser was retained. Where all four prepare their work with the same model, their errors correlate and the independence being relied on becomes nominal — one reasoning route arriving in four envelopes. A second vendor buys less than it appears to. Kim and colleagues, testing more than 350 models, found errors correlated across distinct architectures and vendors; the larger, more accurate models err more alike.

The question, and the forum. Put it to the audit committee when the annual assurance plan is tabled, and to each provider in turn: which models were used in preparing this, and on what material were they run? A good answer names the tools and shows one line of assurance built on a separate reasoning route — a challenger briefed from the raw file, working before the pack was written. An answer describing the firm’s responsible-AI policy has answered a different question.

The second move raises the essay’s own habit to board process. Frank’s discipline: make the machine attack the question before refining an answer you like. A board can require the equivalent of a pack: argued alternatives ahead of the recommendation, written by someone who wants them to win, including the option the machine passed over. Argyris named the same target in 1986. To overcome skilled incompetence, he wrote, people have to learn “to ask the questions behind the questions.” That is the part the machine will not supply, and the room where they can be asked is a board’s to build.