You want two opposite things from the same hire, and you’re right to want both. On the work that touches a customer (the lead, the renewal, the invoice), you want zero surprises. It should happen the same way every time, no drama. And on a good day, you’d love for it to surprise you pleasantly in how it gets there: a sharper reply, a smarter save, a small touch you didn’t ask for.
Those two wishes sound like they’re in tension. They aren’t. They live in different layers, and knowing which is which is most of what makes AI safe to trust with real work.
Unmanaged AI surprises you in the wrong place
Left to its defaults, a language model is not consistent, and this catches people off guard. Ask it the same thing twice and you can get two different answers. Nothing you did changed it; the variation is baked into how it runs.
Researchers at Thinking Machines Lab made the point vivid. They sent a model the identical prompt one thousand times, with the setting that’s supposed to force the most predictable output, and counted the replies. They got 80 different answers. One prompt, a thousand tries, eighty distinct results. That’s fine when you’re brainstorming. It is not fine when the “answer” is whether a renewal went out. Surprise in the what isn’t delight. It’s a defect.
Determinism in the what, room in the how
So the job is to put the surprise where it belongs and stamp it out where it doesn’t. The what, the outcomes that must never vary, gets locked down: every lead answered within minutes, every renewal caught before it lapses, the invoice total that must be exactly right, the sentence you’re legally required to include. None of that is allowed to be creative.
The how is where an AI employee gets to be good. The phrasing of the reply, the order it handles the morning, the way it notices a repeat customer and warms the tone: that can flex, and when it flexes well, it’s a genuine pleasure. Reliability on the outcome and craft in the method aren’t a trade-off. They’re two different dials, and you set them separately.
Precise intent is what buys you both
Here’s the catch: those dials don’t set themselves. The only way to guarantee no surprises in the what while freeing the how is to say, precisely, which is which. What must be identical every time. What’s allowed to vary. Where the hard line sits between them. Vague about that, and you get the reverse of what you want: rigid where you needed warmth, improvised where you needed a lock.
That’s the discipline underneath a managed AI employee. We pin the outcomes that can’t move, watch that they don’t, and leave the method free enough to still surprise you in the good way. That’s the reliability behind why a single slip can mean nine seconds to delete, thirty hours to recover, and why someone has to steer it so that someone isn’t you. It’s also why we tell owners to start with the work you already check, where “right” is easy to define.
No surprises where it counts. Room to delight everywhere else. You don’t have to choose — you just have to be exact about the line.