Everyone’s posting AI reliability horror stories lately — stats, surveys, all pointing at how unreliable AI-assisted code supposedly is. Here’s a contrarian data point, from someone who ran the previous model too.
What outsourcing actually cost
Before AI, when I outsourced small dev projects, requirements had to be airtight or the code came back unusable. Keeping scope small and focused helped a lot — but string enough small ad-hoc pieces together and you’re the one left making sure they still add up to the bigger picture. Even then, “working” often meant weeks of my own review, refactor, and testing to get it to a standard I’d actually ship: code review, unit/integration/end-to-end tests, BDD, documentation, audits — all on me, all after the fact, all the same regardless of how the code was written or by whom.
AI hasn’t removed that tax — it’s moved it
My rule hasn’t changed: trust nothing until I’ve verified it myself. Human-written or AI-written doesn’t matter, and I’ve got horror stories from both sides to prove it. What’s changed is where the effort goes. Less time reviewing finished, often poor, work after a slow contractor round-trip. More time up front on requirements, architecture, and tests — the parts that actually decide whether something’s correct before a line of implementation exists.
Early on, AI got things confidently, completely wrong often enough that I leaned harder on sanity checks — and building verification into the pipeline turned out cheap enough to just become part of the cycle, not a bolt-on. That’s the part the reliability surveys tend to miss: the number that matters isn’t “how often is it wrong,” it’s “how expensive is it to find out.”
Where the effort goes now
The result isn’t “faster because we skipped steps.” It’s shorter, tighter iteration loops, because verification is immediate instead of a contractor round-trip measured in days. For a bootstrapped build, that’s the whole game: quality held constant, cost and time both come down.
The biggest shift for me personally: I spend far more time in the research and requirements phase, exploring how one architecture holds up against another before committing to it. Plenty of solutions get the job done, even done well — finding the right one is a different question, and it’s the question worth spending the saved time on.
Betting on evidence, not assertion
When the increments are small enough, quality doesn’t get sacrificed for speed — each iteration beats the last, because now you can back the claim with evidence, not just assert it. “This way is faster.” “This way is better.” We’ve all heard those lines before. How often did anyone actually verify them, when verification used to cost more than it was worth? That’s what’s different now: the cost of checking dropped enough that skipping the check stopped making sense.
Curious what this has looked like for other technical founders — what’s actually worked with AI-assisted building, what hasn’t, and what you changed because of it.
Kenneth Springer is the founder of snapWONDERS, a digital forensic analysis platform for images and video. The verification discipline described here — trust nothing until it’s checked, regardless of who or what produced it — is the same standard every snapWONDERS release is held to. snapWONDERS forensic analysis — no account required.