The question has changed shape over the last couple of years. It used to be “why should I hire you instead of a cheaper freelancer,” and now it’s some version of “can’t AI just build this.” It’s a fair question, and the honest answer is: for a meaningful chunk of the actual typing, yes. Where that answer stops being “yes” is the more interesting part, and it’s the part that actually determines whether a project ships correctly.
What AI genuinely handles well now
I’m not going to pretend otherwise, because clients can tell when you are and it costs you credibility on the parts that matter. Scaffolding a new service, drafting a first-pass component, writing boilerplate CRUD endpoints, converting a design into markup — an AI coding agent does a large fraction of that competently, fast, and it’s changed how quickly I can get a client from “let’s start” to something clickable. Refusing to use that because it feels like it undercuts the value of hiring a human would just make me slower for no reason a client should care about.
What it doesn’t do, and why that’s the actual job
Scoping an ask that isn’t fully formed yet. A client saying “we need a booking system” isn’t a spec, it’s a starting point. The airport transfer platform I built didn’t start as a clear requirements doc — it started as a business running bookings by phone and WhatsApp, and the actual first job was figuring out what “self-serve” needed to mean for their specific customers: a real-time quote, live flight tracking so a delayed arrival doesn’t strand someone, TWINT as a payment method because that’s what Swiss customers actually use. None of that comes from an AI agent parsing a prompt — it comes from asking the client the right questions before writing anything, several rounds of them, and knowing which questions to ask because you’ve shipped booking flows before and know where they usually break.
Making the tradeoff calls that don’t have a clean right answer. Self-host or stay managed. Build a custom CRM or push back and recommend an off-the-shelf one. Ship a phased milestone now with known gaps, or wait two more weeks for something more complete. These aren’t technical questions with a correct answer sitting in a training set — they’re judgment calls that depend on a specific client’s risk tolerance, budget, and timeline, and getting them wrong doesn’t look like a bug, it looks like a project that technically works and still fails the client.
Catching the mistake before it ships, not after. I’ve written elsewhere about an AI-generated Terraform change that would have destroyed a production database if I’d applied the plan without reading it line by line. The agent wasn’t wrong to produce that diff — it did what I asked. Catching it was the job. That’s true more broadly: AI-assisted output still needs someone who understands the system well enough to know when “looks right” and “is right” have quietly diverged, and who reads for that on purpose instead of skimming for vibes.
Owning the outcome after it ships. When the booking platform’s payment flow breaks at 11pm on a weekend, “an AI wrote that code” isn’t an answer a client wants or should accept. Accountability doesn’t delegate to a tool. That’s not a technical capability gap, it’s a structural one — someone has to be the party who answers for the outcome, and that’s what a client is actually contracting for when they hire a person instead of assembling a stack of tools themselves.
What this actually means for how I work now
It means I’ve stopped billing for typing speed, because typing speed stopped being the scarce resource a while ago. What I’m actually pricing is the two judgment-heavy stages on either end of that diagram — turning an ambiguous ask into a scoped, sequenced plan, and reviewing what comes out the other end closely enough to be the one who’s accountable if it’s wrong. The middle stretches faster than it used to. The parts on either side of it didn’t get any faster, because they were never a typing-speed problem to begin with.
The clients who’ve pushed back hardest on hiring a person instead of “just using AI” tend to be the ones who haven’t yet had the experience of a plausible-looking output being subtly, expensively wrong in a way that only shows up under real use. The ones who have had that experience don’t ask the question anymore. They ask who’s going to be accountable for the answer.