This year I designed, built, and shipped my own product — Avklare — using AI as a genuine development partner. Long design conversations, real back-and-forth, thousands of messages. It went well. The product exists, it works, and people use it.

It also shipped with a decision that quietly cost four production bugs. We found them weeks later — not because a user complained, but because we went looking. Nobody had reported a thing. That's the part that should worry you more than the bugs themselves: at small scale, a system can be meaningfully broken and still feel completely fine.

Here's the uncomfortable part. I saw this one coming. I said so, three separate times, over eleven days, in writing. It shipped anyway.

Early in the build, the AI and I made a call: two different kinds of thing in the product — full tasks and simple list items — would share one database table, told apart by a flag. Every individual argument for it was correct. Fewer moving parts, simpler screens, less to maintain. What it didn't account for is something you only learn by living downstream of a decision like it for years: that kind of shortcut turns into a condition every future piece of the system has to remember, unaided, forever. It's free today. It costs someone a couple of weeks of confusion, later.

I felt that coming, and I said so — more than once. "I am concerned that this is one of those things you would regret later and would be hard to unwind." "Maybe we should have built this as two separate things." Real notes, with real dates.

Then I went back through the record to see where I'd lost that argument. There wasn't a moment where I made my case and got outreasoned. What I found instead was simpler, and more useful to know: every one of those messages was a question or a hedge, not a statement. And a hedge, aimed at something that answers everything fluently and instantly, just gets answered. Clearly, reasonably, confidently — and the conversation moves on. Nobody overruled me. The concern just closed on its own.

That's the real finding: what decided the outcome wasn't who knew more. It was who sounded more certain. I wasn't hedging because I doubted myself — I was hedging because that's what a hunch sounds like before you can fully justify it. Thirty years of pattern-matching, running quietly in the background, coming out as "maybe" instead of "no."

Why this matters if you're bringing AI into your business: AI is, genuinely, the best development partner I've worked with in thirty years — fast, broadly correct, tireless, and able to read an entire system in seconds in a way no person can. But it has a real blind spot, and it's structural, not carelessness: it has never had to live with a decision. It can't price a cost it's never paid. So it will help you build something that works beautifully today and is expensive in a way that won't show up for a year — with complete sincerity and perfectly reasonable-sounding logic the whole way.

That's not an argument against using AI to build software. It's an argument for making sure someone in the room has felt the cost before, and says so in a sentence that can't be politely answered and moved past.

What I'd actually check for, if I were you:

  • Ask whoever's building your systems how they catch the mistakes that don't show up in a demo — the ones that only cost you a year from now.
  • Ask what happens when someone on the team has a hunch and can't yet justify it. Does it get written down, or does it just quietly lose to the next confident answer?
  • Before anything gets built on an assumption, check whether the assumption is actually true. The load-bearing argument for our design — that a list item would occasionally get promoted into a full task — turned out to have happened zero times, ever, in six weeks of live use. One query would have told us that on day one.

AI has closed the knowledge gap almost completely — it will answer any question you ask, well, at any hour. What it can't do is tell you which questions you forgot to ask. That's still the job of the person who's lived through it before.

— Alan