You told the team the AI pilot was optional, low pressure, nobody’s job on the line. Two weeks later, adoption sits around 20%, and the people who did try it are using it for the safest possible thing, spell-checking an email, not the actual workflow you needed off your plate. You start wondering if the team just doesn’t get it.
They get it fine. Their brains are doing exactly what brains are built to do.
Why the rollout stalls before it starts
Cognitive neuroscientist Chantel Prat laid out the mechanism on the podcast Beyond The Prompt, pointing to a model of curiosity built by researchers Matthias Gruber and Charan Ranganath. They call it PACE: prediction, appraisal, curiosity, exploration.
Prediction comes first. Your brain has to register that something is new, that your existing expectations don’t cover it, before curiosity has anything to grab onto. Next is appraisal, and this is the step that decides everything: before your brain lets you get curious, it checks whether you’re safe. Only if the answer is yes does curiosity switch on, and only then do you get exploration, actually picking the thing up and using it.
Most rollouts don’t stall on prediction. Everyone already knows AI is new. They stall on appraisal. If the tool shows up reading as a threat, faster than me, watching what I do, here to replace the slow parts of my job, the brain shuts curiosity down before it gets a vote. That’s a safety read, and it happens underneath the part of the brain that’s listening to your kickoff meeting. No amount of prompt-writing training reaches it, because the block was never technical to begin with.
The fix is naming the threat before somebody’s brain names it for them, and doing it more than once. A bigger training budget won’t touch it. Tell people directly, before they start wondering: getting good at this will not get you punished with more work for the same pay, and it will not get you quietly phased out for being efficient. Say it in the room, not just in a memo, and say it again the first time someone on the team gets visibly faster at something. That’s the appraisal step working for you instead of against you.
Never ship the first answer
Once people are actually using the tool, the next stall shows up as settling. The first draft a model hands back looks decent enough, and decent gets shipped, because reviewing feels like the slow part of the job you were trying to speed up.
There’s a story Ryan Holiday tells about Henry Kissinger on the same podcast. Kissinger asked a staffer for a report. It came back, and without reading it he sent it right back: this is wrong, do it better. The staffer tightened it, brought in better research, sent it again. Kissinger, still without reading it, sent it back a second time: no, this is still wrong, I told you to do it better. Only on the third pass did he actually sit down and read it.
Asking again costs almost nothing with a model, and that’s the actual lesson here. With a person, pushing back twice spends real goodwill and real days. With a model, it costs a few seconds and some compute you’ll never notice. If you wouldn’t send a client the first draft of a proposal, don’t send them the model’s first draft either. Ask it what’s weak, what it left out, how it would tighten this if you handed it to your sharpest editor, before you decide the answer in front of you is the answer you’re keeping.
Don’t put a checklist on the thing that isn’t repeatable
The third stall shows up later, once AI is working and you start writing it into how the company actually runs. The instinct is to turn every process into an SOP the model can execute end to end. Most of the time that instinct is right. Sometimes it kills the exact thing you were trying to protect.
Henrik Werdelin, cofounder of the dog toy company BarkBox, described this on Beyond The Prompt. His head of design, Derek, makes what the team calls Bark Magic, oddball toys nobody else on the team, and nothing AI can generate, comes close to replicating. Every time leadership tried to pin down exactly what makes Derek’s work special enough to put a KPI on it, the harder they pushed for a precise definition, the less special the work got. If you could fully define it, Werdelin said, it wouldn’t be magic anymore.
Bark Magic is rare. Most of what actually runs your company is repeatable, and repeatable is exactly what belongs in AI’s hands: first drafts, scheduling, routing, the eighty percent that follows the same shape every time. But somewhere in your business there’s a Derek, the person whose read on a hard client call or a pricing exception is the actual differentiator. Don’t build the SOP that tries to capture that judgment. Build the SOP around it, and leave the call to the person who can still make it.
Put the three together and that’s the real playbook. Roll out AI in a way that reads as safe, not threatening. Never let the first answer be the last one you accept. And know the difference between what deserves a checklist and what deserves protecting from one.
Where’s your rollout actually stuck right now, the safety read, the settled-for first draft, or a checklist creeping toward something that shouldn’t have one?
If you want a second set of eyes on which one it is, grab fifteen minutes with me. No pitch. Just a look at where it’s actually stalling.
Frequently Asked Questions
What is the PACE model and who created it?
PACE, prediction, appraisal, curiosity, exploration, is a model of curiosity built by researchers Matthias Gruber and Charan Ranganath. Cognitive neuroscientist Chantel Prat cites it on the podcast Beyond The Prompt to explain why AI rollouts stall at the appraisal step, when the brain reads a new tool as a threat instead of a safe place to explore.
Why does an AI rollout stall even when the team says they’re on board?
Saying yes and appraising something as safe are two different steps. Before curiosity ever kicks in, the brain checks for threat: job security, being watched, looking incompetent in front of a manager. A team can be verbally supportive and still freeze at that appraisal step. That freeze is a safety read, not a willingness problem.
What’s the do better technique for reviewing AI output?
It’s named for a story Ryan Holiday tells about Henry Kissinger rejecting a staffer’s report twice without reading it, only saying do it better, then finally reading it on the third pass. Applied to AI, it means never treating the first answer as final. Asking a model to try again costs seconds, not the two days it would cost a person.
How do I know what to turn into an SOP and what to leave alone?
If the work is repeatable, the same shape every time, it belongs in an SOP AI can run. If the value comes from a judgment call that gets harder to explain the more precisely you try to define it, that’s a signal to protect it, not automate it. That’s the Bark Magic problem: the more you pin down what makes someone’s work special, the less special it gets.
What’s the actual first step in a safer AI rollout?
Name the threat before someone’s brain names it for them. Tell your team directly, before they start wondering, that getting good at this will not cost them their job or get them handed double the workload for the same pay. Say it more than once, and say it out loud the first time someone visibly gets faster at something.