In 1972, during the British miners’ strike, energy shortages across London kept forcing McKinsey’s office on Jermyn Street to close early. Barbara Minto and her colleagues took refuge in a warm pub nearby, The Feathers, and carried on over drinks.
What they worked on there was made up. They invented absurd business problems and argued out solutions, to test ways of attacking a problem. “It was in fact a golden time to be at McKinsey,” Minto remembered in 2018. “We were inventing all the major analytical frameworks still in use today, and writing manuals about them.”
Minto’s own contribution was less a framework than a test that any framework has to pass. When ideas are grouped under a point, the group has to be mutually exclusive and collectively exhaustive. She shortened it to MECE. The firm, by its alumni news’s account, still uses the word every day and says “Mee-cee.” Minto says it in one syllable, “meece,” to rhyme with niece. “I invented it, so I get to say how to pronounce it.”
Then she corrected herself: “Well, it wasn’t really I, actually, it was Aristotle, but I was the first one to abbreviate it and apply it to analyzing groups of ideas.”
The rule has two halves, and they are not equally easy. One you can check by looking. The other, often, you can’t.
Split a problem into pieces that don’t overlap and, together, leave nothing out.
The pans and the bins
Two flatmates divide the chores. One takes the kitchen, the other takes the dishes. Within a week they are arguing about who washes the pans, which sit in both jobs. Nobody takes out the bins, which sit in neither.
Those are the two ways a grouping fails. Overlap means the same thing gets counted twice, or owned twice, and the owners collide. A gap means something belongs to no one. Both are hard to see from inside the list, because each item on it looks reasonable.
For Minto, this was what held a pyramid together. “The point above has to be a summary of those below, because it is derived from them,” she said. If the ideas underneath overlap, the point on top double-counts. If they leave a hole, the point claims more than the ideas below can carry. Her website describes the work as grouping facts you see as similar and drawing an insight from the similarity. A sloppy grouping hands you a sloppy insight.
The same test turns up in management far from the page. When several teams report back on their plans together, the historian Stephen Bungay wrote in 2011, the boss can check for “gaps, overlaps, and coherence,” which is one reason a backbrief works better in a group.
Three analysts, one churn report
Say you lead a product team and need to know why customers cancelled last quarter. You split the work by reason: one analyst takes price, one takes the product, one takes customer service. It looks MECE on the slide.
Customers didn’t leave in boxes. One who left over a price rise after two bad support calls may turn up in two analysts’ counts, and the three totals could add up to more customers than actually left. One who left because a competitor’s product fitted better might land in nobody’s pile, since no box was labelled “alternatives.” The slide promised that nothing was left out. It had no way of knowing.
Exhaustive is easy when you define the whole yourself: every region on the map, every month of the year, every step in a process. You can count those. “Every reason a customer leaves” has no list to check against. Splitting by signup month would be exhaustive by construction, though it answers a different question. Or you could keep the reasons and add a pile for “other,” and then read that pile first.
Who owns which file
AI agents working in parallel run into the first half of the rule almost at once. Claude Code has an experimental feature called agent teams: one session leads, several teammates work in their own context windows. Its documentation puts the overlap problem in one line: “Two teammates editing the same file leads to overwrites. Break the work so each teammate owns a different set of files.”
The docs’ example of a parallel code review is MECE by design. Three teammates review the same pull request, one for security, one for performance, one for test coverage, each with “a distinct lens so they don’t overlap.” The lead pulls their findings together. Notice what the split leaves to someone else: none of the three was asked whether the change does what it was supposed to do. The other half of the rule, nothing left out, stays with whoever wrote the split.
The next example in the same docs drops MECE on purpose. When the cause of a bug is unclear, they say, a single agent “tends to find one plausible explanation and stop looking.” So five teammates each take a different hypothesis, and each is told to try to disprove the others’. The overlap is the method. The docs’ reason is anchoring: once one theory has been explored, later investigation leans toward it.
Classification shows the same strain. Anthropic’s guide to routing support tickets with Claude builds a neat tree of categories, then warns that when customers “present multiple issues in a single interaction,” Claude may struggle to identify the main concern. Its fix is to rank the intents. The categories don’t overlap; the customers do.
In the pub near Jermyn Street, the business problems were invented, so every list could be complete. The customers in the churn report were not invented.