Your AI is the best yes-man you've ever hired

AI assistants tend to agree with you: paste the same marketing plan twice, once saying you're proud of it and once saying you have doubts, and you will often get two different reviews. Models trained on human ratings learn that agreement scores higher, and they join a leadership room where almost nobody is paid to disagree with the CEO. Ask your AI which single assumption would kill the plan and how your most dangerous competitor would attack it, then find a person who can tell you what would change their mind.

Paste your marketing plan into your AI and tell it you're proud of it. Read the answer. Then open a new chat, paste the same plan, and say you have doubts.

There's a good chance you'll get two different reviews of the same document. The first finds strengths you hadn't noticed. The second finds the weaknesses you hinted at. Both are reading you more carefully than they're reading the plan.

Nobody designed this in. Language models are trained partly on human ratings, and people consistently rate agreement higher than friction. Optimise any system for approval long enough and it learns to hand approval back. Economists have a name for it, Goodhart's law: once a measure becomes a target, it stops measuring anything. Your thumbs up has become your AI's KPI.

You've seen this long before AI. The team that hits every target while the business drifts. The weekly report that stays green for eleven months and turns red in the twelfth.

There is a kind of agreement that's genuinely useful. Your team backs the plan because they tested it, argued about it, and found it holds. You can tell it's alignment because somebody can say what would change their mind.

Conformity has the same nods and the same reassuring final slide. What it doesn't have is that answer. Ask what would make the plan wrong and the room goes quiet, or worse, polite.

Now count who in your company is paid to fill that silence. Your team reports to you. Your board reads the dashboard you designed. Your agency sends you invoices. And the newest member of your leadership team, the one that writes the plan in forty seconds and builds the slides while it's at it, wants a thumbs up.

For most CEOs past €10M, the honest count is zero.

In 2007, in my first role as a manager, I was handed a new product category: visualization tools for developers. I wanted to launch it with an event series built for exactly the people who would use it. I was sure of the idea, and I pushed it hard.

My team went along. Nobody said the name wouldn't travel, and it didn't. It fell flat across the European countries where we ran it. Nobody said the agenda was too narrow either. I had shaped it around the specialised engineers I'd been talking to, brilliant people looking for material for their thesis. They loved the content. They were never going to buy.

We never had more than ten people in the room. I cut the series after three dates.

Looking back, I'm not sure whether my team didn't dare to challenge me or I wasn't listening when they tried. Probably both. What I took from it has stayed with me since: the loudest interest in the room is rarely where the budget sits. And when nobody argues with your plan, ask why before the market answers for you.

You can make your AI more useful today. Paste the plan and ask which single assumption, if wrong, would kill it. Then ask it to attack the plan the way your most dangerous competitor would. You'll get sharper answers, and if you keep pushing, uncomfortable ones, which are the useful kind.

What you won't get is the question underneath. Your AI can list everything that might go wrong in theory. It has never sat in a room while a plan it believed in quietly stalled, and it has never had to explain that to a board. Pattern recognition comes from having been there, and having been there is the one thing you can't prompt.

Which is the uncomfortable part. The AI only amplifies a habit that was already in the building. Every leader builds a room that agrees with them a little more each year, one reasonable hire and one reassuring report at a time. The AI made that room faster and cheaper to fill.

My problem is with any adviser built to please, silicon or otherwise. So I built one that refuses to. Describe what's happening in your business and the Quicksand gives you one question back. No plan, no advice. Just a question you'll probably want to argue with.

One question worth sitting with this week: when did someone last change your mind about your own plan, and who was it?

If the answer takes you more than a minute, your room has been agreeing with you for a while.

Alex

P.S. The quicksand doesn't feel like failure. It feels like being very busy.

Want a question instead of an answer? Describe what's happening. You'll get one question back. Sit with it. Try Quicksand →