What AI changed, and what it didn't

Describe a business to a model and you get a competent site in a minute: clean type, sensible spacing, copy that scans. Five years ago that was a fortnight of work. Today it is free, and everyone has it.

Which means looking professional has stopped being an advantage. It is the floor. The things that were hard in 2020 are still hard now: does the booking form know whether the room is free, does the invoice reconcile, does it still work at 2am when a driver is standing in a yard with one bar of signal. Nobody generated their way past any of that.

Reading, drafting, and the grunt. Lend Pilot runs Claude in production to pull structured data out of a borrower's uploaded documents — pay stubs, notices of assessment, statements — so nobody retypes them. That is a genuinely good job for a model: unstructured input, a human reviewing the output, and a cheap failure mode if it is wrong.

It also writes a great deal of code with me, against instruction files in each repository that spell out the money rules, the deploy runbook, and what is safe to touch in the legacy application. The assistant works to the project's standards rather than generic ones.

Arithmetic, money, and anything a regulator will ask about. In Lend Pilot the qualification, the GDS/TDS ratios and the B-20 stress test are deterministic code, with Canadian semi-annual compounding written into them, and the underwriting readout shows its working. A number a model guessed is a liability with a plausible tone of voice.

Same rule for shipping. Nothing goes out because a build went green — an agent renders the screen in a real browser or the iOS simulator, screenshots it, and reports what is actually on it. That habit catches the expensive class of bug: the change that compiles, deploys, and is wrong on screen.

When competent is free, coherent is the scarce thing. A generated site is generically good — it looks like the average of everything, because that is literally what it is. It cannot tell you why your yard runs differently from the one down the road, and that difference is the only thing worth putting on a page.

So the work moves. Less time making it look tidy, more time deciding what is true about the business and saying it plainly. The tools got dramatically better at execution. They did not get better at knowing what is worth executing.

I built conversational call routing at Brikers around 2016, on Twilio Autopilot, when making a machine hold up its end of a conversation was real work. The tooling since has improved beyond recognition. The judgement about where to point it has not changed at all, and it is still the part that decides whether the thing works.

That is the whole argument. AI raised the floor for everyone, mine included. What it did not do is make it any less important to know which parts of a business must never be guessed at.