AI Does the Work. A Human Signs It Off.

Somewhere in the marketing of nearly every AI company is the phrase 'human in the loop'. It's doing a lot of work, that phrase. It reassures without specifying. Which human? Looking at what? With the power to refuse what? Usually you'll never find out, because the phrase was the product.

We use a different sentence, and we use it everywhere: AI does most of the work. Paul reviews everything before it ships. If something looks off, we fix it before you see it. It's less impressive-sounding than 'human in the loop'. It's also checkable, because it names the human, the scope and the power. This article is about why that precision matters more than the reassurance.

What AI is actually good at (and honestly isn't)

Run honestly, AI is a phenomenal production engine. It can research a sector, hold a brand voice consistently across thirty pages, apply a design system without drifting, and do the work of weeks in hours, in parallel, without fatigue. What it cannot reliably do is know when it's wrong. It produces confident output at a constant tone whether the output is excellent or subtly off. It doesn't wince. And the wince — the practised discomfort of someone looking at work that isn't right yet — is most of what quality control actually is.

Judgement doesn't automate. So we didn't automate it.

Our pipeline automates everything measurable: contrast ratios, load performance against the 95+ floor, mobile layout, banned words in copy, checks between every phase. Machines are better than people at enforcing measurable floors — they don't get tired on page twenty-nine. But 'technically correct' and 'right' are different standards. Does this feel like the business it's for? Would a first-time visitor trust it? Is anything subtly generic? Those questions get answered by a person, at the end, every time, with the authority to send work back. The delivery approval isn't ceremony. Builds get refused there.

Why we admit this at all

There's a version of our marketing that says 'fully autonomous, zero humans' — it demos better. There's another that hides the AI and implies craftspeople at benches. Both are lies in opposite directions, and both eventually get found out by the same person: you, three months in. The businesses that get burned by AI aren't burned by AI. They're burned by nobody having looked. When you know exactly who looked and what they could refuse, you can decide what our promise is worth. When a supplier won't tell you, you're being asked to trust a process nobody will describe.

Questions to put to anyone selling AI-built anything

• Name the human. Not a team, a name. Accountability that isn't a person is a shrug.
• What do they review — everything, or a sample when time allows?
• What happens when they don't like it? If the answer isn't 'it goes back and gets fixed', the review is theatre.
• What's measured automatically, and against what floor? Judgement should sit on top of measurement, not replace it.

We can answer all four in one breath, and the fourth has a number in it. That's not a boast about virtue. It's the minimum architecture for selling AI work to people who have real money on the line — and the fact that it currently counts as a differentiator tells you everything about where this industry is.