In The Prestige, Robert Angier hires Nikola Tesla to build him a machine. He doesn't understand what it does, and he doesn't especially care. He cares about what it would let him do — walk onto a stage, vanish, and reappear across the room a second later, to a gasping audience, every single night. He isn't chasing a mechanism. He's chasing the reveal.
Tesla builds it. It works. And what it actually does is quietly horrifying: the machine doesn't move Angier anywhere. It duplicates him. Somewhere below the stage, a version of himself drowns in a tank of water every night, so the version above can walk out and take the bow. The trick never stops costing what it always cost. It just stops being visible.
I had a smaller, much less gothic version of that conversation with a CEO last week.
The machine they wanted built
They'd hired me to scope an internal Q&A bot — employees ask a question, the bot answers it instantly from the company's own policy documents, nobody has to go hunting through folders or wait on Eric in compliance to get back to them.* Completely reasonable ask. A lot of mid-market companies are building exactly this right now.
Then I actually looked at the documents. Old versions sitting next to new ones. Contradictions between departments that had clearly never been reconciled. Decisions that existed only because someone remembered making them in a meeting two years ago and had never bothered writing them down.
Here's the part that should worry you more than a bot that doesn't work: this one would have worked. It wouldn't have failed. It would have answered every question fluently, confidently, and wrong roughly as often as the documents underneath it were wrong — which, at this company, was often. We'd have taken a problem that used to be contained — "just go ask Eric, he knows" — and handed it a company-wide microphone.
That's Tesla's machine. Impressive from the seats. And underneath it, something quietly multiplying — not a body in a tank, but a wrong answer, duplicated with total confidence, walking around the company as if it were the truth, every single day the thing stayed live.
The trick nobody claps for
There's a second magician in that film, and he's the one who actually understood the trick. Borden never builds a machine. His version of the Transported Man works because of something almost embarrassingly unglamorous underneath it — a discipline so complete it costs him half his life, and he never once lets anyone see the mechanism. No machine. No spectacle. Just the invisible work of making the underlying thing actually true, for years, before anyone ever sees the trick land.
That's what I told the CEO to go build first. Not a bot. A process. Get the right answers written down in one place. Retire the outdated versions so they can't keep contradicting the current ones. Put someone's name on owning the knowledge base, so "the truth" has an address instead of living in Eric's head and four different SharePoint folders. None of that is glamorous. None of it demos well in a board meeting. It's the twin brother nobody claps for.
Come back in six months, once the source material is actually clean, and then we build the bot — and it'll be worth building, because it'll be answering from something true instead of performing confidence on top of something broken.
They were surprised. Not offended — surprised, the way you're surprised when someone tells you the impressive part isn't the part that matters. Then they agreed.
Which machine are you building
My job is occasionally telling people the thing they're excited about isn't worth building yet. If you want a yes-man, hire one. If you want someone who'll tell you which machine is Tesla's and which one is Borden's before you spend six figures finding out, that's the work.
Automate a broken process, or feed a model bad data, and you don't get a functioning trick. You get Angier's machine — impressive from the audience, and something drowning quietly underneath it, every single night, for as long as you keep running the show.
*Names have been changed to protect the possibly annoyed.