They showed up on LinkedIn two years ago and they’re still here: “prompt expert”, “prompt engineer”, “instruction engineer”. €2,000 courses to learn how to talk to ChatGPT. Companies opening a role to hire someone who “masters the art of the prompt”. And most of them are spending money to solve the wrong problem. Not because writing good prompts doesn’t matter —it does— but because it’s, by far, the easiest and least scarce part of building AI that actually works in production.
Why hiring a prompt expert sounds good and turns out badly
The idea is seductive because the prompt is the only visible part. When you watch someone pull a brilliant answer out of a model, it looks like magic, and magic seems worth paying for. But you’re confusing the switch with the wiring. The prompt is the switch: the part you touch. The system that makes the light turn on reliably, every day, with the right data and without burning the house down, is something else entirely —and that’s where the real work is.
The result of hiring a “prompt expert” and stopping there is predictable: you get a gorgeous demo that impresses in a meeting and falls over the first week it meets reality. Because the perfect prompt over data that doesn’t exist, with no connection to your systems and nobody watching what happens when it fails, isn’t a solution. It’s a parlor trick.
The prompt is the part that’s least scarce
Writing a good prompt is a real skill, but you learn it in weeks, you document it in a page, and —increasingly— the models do it themselves. Today’s models are far more tolerant of imperfect instructions than those of two years ago: they get what you want even if you don’t phrase it as a ritual. Betting your competitive edge on writing better sentences than everyone else is betting on a moat that dries up on its own with every new model version.
What’s scarce isn’t the sentence. What’s scarce is everything around it: knowing which process in your company is worth automating, where the real data the answer needs comes from, how it connects to your ERP or your CRM, what happens when the model gets it wrong and who finds out. That isn’t learned in a weekend course, and it’s exactly what separates a demo from a system.
What’s actually scarce: the system around it
When you break down an AI system that’s been running in a company for months, the prompt is a tiny fraction of the whole. What holds up the result is this:
- The data. The most brilliant prompt over information the model can’t consult produces answers that are confident and false. The work is in connecting the real source —the order, the invoice, the status— so the system looks it up instead of inventing.
- The integration. An answer in a chat window does no work. Value appears when the system reads your email, writes to your CRM, fires your workflow. That’s plumbing, not prompting.
- The evaluation. How do you know it works 95% of the time and not 60%? Without a system to measure hits, misses and edge cases, you’re operating blind and calling it confidence.
- The operation. What happens when it fails, who reviews it, how it escalates to a human, how it updates when the model changes. A system with no owner isn’t a system: it’s a time bomb with good packaging.
None of this is solved with a better prompt. It’s solved by building. If you want to see how the piece that actually does the work gets built, the guide on how to create an AI agent breaks down the parts that matter, and the one on training the agent with your information explains why your own knowledge —not the instruction— is the defensible asset.
What to hire instead
Don’t hire the person who knows how to talk to a model. Hire the one who knows how to build the system that does the work and leaves it running in production —measured, with a safety net and with an owner. The right question in an interview isn’t “show me your best prompt”; it’s “show me something you built that’s been running for six months without you watching it”. The first filters for theatre; the second, for judgment.
In practice, that’s almost never a person writing sentences: it’s someone wiring up your data sources, your integrations and your escalation rules until a specific process stops eating your team’s day. When what’s stealing your hours is the same emails over and over, the fix isn’t a prompt: it’s actually automating the repetitive emails. And when the pain is operational and cross-cutting, operations automation builds the whole system —prompt included, as the small piece it is.
The “prompt expert” doesn’t disappear: it dissolves into a bigger, more useful role. Knowing how to ask the model well is a skill every good implementer has, just as a good plumber knows how to open a tap. But you don’t hire a plumber for knowing how to open the tap. You hire them for everything behind the wall.