I've been watching The Traitors, and I can't stop seeing my day job in it.

A group of smart people gather observations all day, sit down at the round table, and confidently banish… the wrong person. Almost every time.

They're not unintelligent. They're reasoning confidently from information that's incomplete, noisy, and sometimes deliberately corrupted,  and the format forces a confident answer anyway.

That's also a pretty good description of what your AI does when you point it at bad data.

When people say "the AI hallucinated," they picture the model inventing things from nowhere. But most real-world AI failures aren't spontaneous. They're the model reasoning correctly over bad inputs. Garbage in, confident garbage out.

Here's the part most teams get wrong: when a banishment goes sideways, everyone relitigates the round table-the votes, the speeches. But the mistake happened hours earlier, at breakfast, when someone misread a yawn as guilt and filed it away as "evidence."

Most AI initiatives obsess over the round table too. Tune the model, tweak the prompts, buy the bigger model. All at the moment of the decision. Meanwhile the real failure was upstream-duplicate records, stale fields, missing context, the spreadsheet where "N/A" and "0" and blank all mean different things.

You can't out-prompt bad data. You can only delay how long it takes to embarrass you.

Good news for lean teams: you don't need enterprise governance to do this well. Pick the handful of sources your AI actually depends on, make those trustworthy, know their gaps, and design for "I'm not sure." That's a weekend of focused work, not a six-month program.

The contestants who win The Traitors aren't the ones with the wildest theories. They're the ones disciplined about what they actually know.

Build your AI to play like them.

Wrestling with the data side of an AI initiative? It's what I do-let's talk. 👇

#AI #DataQuality #AIStrategy