Every team "uses AI" now. Cool. That tells me nothing.
Put it on a slide, put it in the pitch, put it in the job posting. Every team on earth will now tell you, unprompted, that they "leverage AI throughout the development lifecycle." Fine. Doesn't mean anything. The question was never whether you use it. The question is where you let it near the wheel, and whether you actually know the difference between those two things.
Most teams don't. They point a model at a hard problem, take whatever comes back because it compiles and the demo works, and ship it. That's outsourcing your judgment to a glorified autocomplete engine and hoping nobody checks, not AI adoption.
What AI is actually good at
Boilerplate. Scaffolding. The CRUD endpoint you've written four hundred times. The regex you'd otherwise spend twenty minutes fighting. The first draft of a test suite. Summarizing a log dump at 2am so you don't have to read all six thousand lines yourself. Turning a rough idea into a rough first pass, fast, so a human can start editing instead of starting from a blank page.
That's real leverage. We use it for exactly that, constantly, and we're not going to pretend otherwise to sound more artisanal than we are. A senior engineer who refuses free velocity on the boring parts is just working slower on purpose.
What it is not good at
Deciding what actually matters. Knowing which shortcut becomes a six-figure incident eighteen months from now. Reading a room full of operators and knowing which requirement they're too polite to push back on. Knowing when "technically correct" and "actually right for this business" have quietly stopped being the same thing.
That's taste. It's built from having watched things fail, in production, with your name on the postmortem. A model has read about failure. It has not sat in the room when the failure happened and felt what it cost. Those are not the same kind of knowledge, and no amount of context window fixes that gap.
Top tip
If you can't explain why a piece of AI-generated code is correct without re-reading it line by line, you haven't reviewed it. Ship that enough times and eventually one of those lines is load-bearing.
The Wix problem
Here's the analogy we actually use internally.
If you're selling muffins out of your kitchen, get a Wix site. Templated, cheap, up in an afternoon, and genuinely the right call. You have better things to do than hand-roll a homepage for a fifteen-order-a-week bakery. Nobody is wrong for doing that.
Now scale that same instinct up to a business doing real revenue, with real operational complexity, real edge cases, real money on the line if the system gets it wrong at 2am with nobody watching. Run that on templated, best-guess, nobody-really-owns-it output and you don't get a cute homepage. You get an outage with your name on it, a customer who churned because the exception case wasn't handled, a number on a board that quietly stopped moving and nobody can say why.
AI, used without judgment, is a Wix site for problems that were never Wix-sized to begin with. It looks finished. It demos great. It has not been asked a single hard question yet.
How we actually run it
We let AI touch anything that's reversible, boring, or bounded. The stuff where a wrong answer costs you five minutes, not five figures. We do not let it anywhere near architecture decisions, anything customer-money-touching, anything where "mostly right" is functionally the same thing as wrong. Those get a senior human, full stop, every time, no exceptions for being busy that week.
That split is a rule we actually hold our people to. It's the difference between a team that ships fast because they're good and a team that ships fast because nobody cares enough to check that the job was done correctly.
The standard
AI didn't lower the bar for good engineering. It raised it, by removing every excuse for the boring parts being slow. What's left after you take the boilerplate excuse away is taste, and taste is still a human thing. We'd rather be the team that knows the difference than the team that finds out the hard way which parts of the system nobody actually reviewed.