How to avoid getting tricked by AI sycophancy.

"Sycophancy refers to the behavior of offering insincere, excessive flattery to someone powerful or wealthy, usually in order to gain a personal advantage, promotion, or special favor."

Simply said: AI agrees with everything you say.

AI gave you the answer in 28 seconds.

But cost you $96K to undo. What happened was...

You gave AI a messy decision.

It came back in under thirty seconds. Clean structure. Clear recommendation. You forwarded it to your team.

Six weeks later you dropped the pricing model it suggested. Two clients didn't follow you into the new structure. At roughly $4,000 a month each, that's $96,000 in ARR you spent the next quarter trying to replace.

You gave it a bad input and the AI just returned a polished version of that bad input.

You asked: "Should we raise prices?" when the real question was: "Why are clients churning before month three?"

The model answered what you asked. The answer was coherent, supported, and built on a frame that was already broken.

This is the failure that never shows up in the post-mortem.

When a decision goes wrong, founders blame the market, the timing, the execution. Almost never the question they handed AI. Because the output looked credible. Because confident prose signals rigorous thinking.

The failure is structural. You were stressed. You opened the chat. You typed the question already forming in your head, assumptions included. The model took your frame and built on it. It does not push back on a loaded question. It runs.

Here is what to do before you hand a real decision to AI.

Write two things before you open the chat:

1. What you know for certain: Facts you can point to, numbers you have, patterns that have repeated

2. What you're assuming: Things you're treating as true that you haven't verified

The second list is where most decisions break.

For the pricing example, the fact list had one item: "margins were tighter than last year."

The assumption list had five:

- "clients would follow the new pricing",

- "the churn was price-driven",

- "the market could bear the increase",

- "retention would improve at higher commitment",

- "existing clients valued the service enough to re-sign."

None of those five had been verified. All five were inside the question handed to AI.

When the assumption list is longer than the fact list, the decision needs more information. Hand AI the question of how to get that information, then come back.

When the lists are close to even, hand both to the model. The output quality jumps because the input quality jumped.

$96,000 is a steep price for a fast answer to the wrong question.

13 comments