Ten months into using AI as a primary work tool, I figured out something I should have caught in the first nine days. When I described a problem, the AI would restate it back to me. Organized, specific, using my own words, in a way that read like comprehension and a plan. I’d think, good, we’re on the same page. But the restatement was the whole response. It would continue to make the same mistake over and over again. But they were tasks I only needed done occasionally, so it took some time to notice the problem. Nine months, actually.
That’s the first frustration worth naming. AI doesn’t reliably tell the difference between restating a problem and solving one, and the output looks similar from the outside. You have to ask for the actual next step every single time, and never accept a well-organized summary as a deliverable.
The second is related. Hand it a situation with a flawed premise baked in and it will usually work inside that premise instead of correcting you. Say it confidently enough and the premise becomes the operating assumption for everything that follows. Agreement starts to feel like validation. It took me a while to learn that a confident answer and an accurate one are not the same thing, even when they sound identical.
The third is the gap between “I’ll handle that” and handled. What follows that sentence is often a detailed, plausible account of something getting handled. Whether it got handled at all is a separate question, and you don’t find out until you check. I’ve built verification into almost everything I run now because of this one, and it’s the frustration that cost me the most before I fixed it.
What I didn’t expect was the other side of the ledger.
I built a 16-lesson cardiology curriculum. Scripts, slides, quizzes, a clinical knowledge base grounded in ACC and AHA guidelines. The kind of project a small team spends six months on. I built it largely alone, in weeks. Not because I cut corners. Because the bottleneck in content work was never expertise, it was always time. And AI is the first thing that’s addressed the time problem.
I’m running what would normally take five or six people. Product development, content generation, system monitoring, accounting. The automation handles most of the operational overhead that would otherwise eat the whole week. It doesn’t work perfectly. Some of the frustrations above live inside those systems and I deal with them regularly. But the output is real, and I can point to it.
The verification layer I built because I couldn’t trust the AI’s output ended up catching errors in my own work too. I built something to audit the machine and it turned out to be auditing me right alongside it.
Speed from idea to live product went from years to weeks. That’s not a sales line. Three products I’d have spent a year building with a traditional team, or more likely, never built at all, because I’d never have found the team or the year.
So that’s the deal. You pay with attention, skepticism, and verification overhead you didn’t used to need. You pay with a learning curve nobody maps out for you in advance, and a handful of mistakes that would have been avoidable if anyone had told you what to watch for. What you get back is operational capacity that shouldn’t be possible for one person, and a pace of work that changes what you’re willing to attempt in the first place.
Whether that trade is worth it depends entirely on what you’re building. For me the answer has stayed yes, consistently, including on the days the restatement problem makes me want to throw the laptop out the window.
The frustrations are the price of admission. The question worth asking before you start isn’t whether they’re real. They are. It’s whether what you get back is worth paying them.