The most qualified team is often the one that learns least

A technology consultancy, a mid-size Brazilian company (figures in Brazilian reais), 140 people. A six-month project is delivered nine weeks late and with a negative margin. The partner in charge calls a learning meeting, with good intentions and a well-built agenda. Two hours later, the conclusion recorded in the minutes is: scope poorly defined by the client, client team unavailable, and a regulatory change midway through.

All of that is true. And none of it is learning — because none of the three causes is under the control of anyone who was in that room. Six months later, another project runs late, with three other equally true external causes.

01The idea: two types of learning #

Chris Argyris worked with Donald Schön on the distinction they presented in Organizational Learning (1978), which he brought to a managerial audience in Teaching Smart People How to Learn, published in Harvard Business Review in 1991.

Single-loop learning corrects the action while keeping the premises intact: the project ran late, so next time we add more people and review the schedule weekly. It's useful and insufficient — it's the thermostat that turns on the heater without ever asking whether the set temperature makes sense.

Double-loop learning questions the premise that generated the action: why did we accept a project with undefined scope? What commercial incentive leads us to sign before understanding the problem? Why did no one flag it in month two, when the deviation was already visible?

The article's uncomfortable finding is this: the more successful someone's track record, the higher the chance their reaction to failure will be defensive. Argyris observed this among high-performing consultants — people capable of precisely diagnosing a client's problem and unable to apply the same rigor to their own behavior when the outcome was bad.

He called defensive routines the organizational behaviors that protect people from embarrassment while, in the same move, blocking examination. The most common routine isn't aggression: it's politeness. Argyris addressed this in Good Communication That Blocks Learning (1994) — the cordial conversation that avoids the hard question and leaves everyone comfortable and nothing resolved.

02Why it still holds #

Because the post-mortem became ritual. Many companies hold the meeting, fill out the document, and file it away. The format exists; the function doesn't. And the sign that it turned into theater is easy to spot: if every cause named is outside the reach of anyone in the room, the meeting produced no learning.

One caveat matters here, because the easy reading of this article is a moral one — concluding that people are dishonest or cowardly. That's not it. The defense is automatic and serves a real function: in an organization where error costs reputation, examining your own mistake in public is irrational from an individual standpoint. It's the same mechanism Amy Edmondson measured in 1999 and that we describe in teams that report more errors make fewer.

There's also a practical, less-discussed obstacle: memory. Six months later, no one remembers which week the deviation appeared, who flagged it, what the response was. Without that record, the meeting works from reconstruction — and reconstructed memory favors, without malice, the version that spares whoever remembers.

03The cost of learning that never happened #

At the scale of the 140-person consultancy, with 18 projects a year. Explicit assumptions:

Want to see how this looks inside a real operation? Explore the platform.

The math: an average project of 1,800 budgeted hours, a 22% overrun means 396 hours per project; × 5 projects, 1,980 hours; at an average billing rate of R$ 315 an hour, roughly R$ 620 thousand in revenue that ceased to exist. What matters isn't the absolute figure but the last line: three reviews held, zero premises revised. Until double-loop learning happens, the same overrun gets budgeted again the following year.

04Where this gets stuck in practice #

1. The agenda asks what went wrong, not what we assumed. The first question produces a list of external causes; the second produces a rule revision.
2. The boss speaks first. From there, the meeting becomes about agreeing with him — politely and completely uselessly.
3. There's no record of the timeline. No one can point to the week the deviation appeared, and the conversation runs on impressions.
4. The conclusion never becomes a verifiable change. It stays as a generic recommendation to improve communication, which obligates no one to anything.
5. Only the disaster gets analyzed. The project that came in on time out of luck is never examined — and it has as much to teach as the one that failed.
6. Whoever raised the flag early came out worse. If flagging risk costs reputation, the next deviation will be flagged late.

05What data-driven management answers #

The defense loses force when the fact is available. If the record shows the hours deviation appeared in week seven and the decision not to renegotiate scope was made in week nine, the conversation can't end at scope poorly defined by the client. Not because someone was brave, but because the timeline is on the screen.

This also makes the review cheap. Much of today's two-hour meeting is spent reconstructing what happened; with the history ready, there's time left for the only part that produces learning. On the pattern that repeats beyond the isolated case, see organizations that learn; on distinguishing real deviation from normal variation, variation isn't error.

At Relevanti, projects, processes, and metrics keep that trail by default — the modules are on the platform, the breakdown by area is in solutions, and a conversation helps look at a real case from your operation.

If every conclusion from your last failure review sat outside your reach, the meeting was comfortable. Learning rarely is.