Writing · Behind the build
Why we give the AI as little to do as possible
We’re an AI-native agency. Our systems hand the AI a deliberately narrow job. The parts of a business that must not wobble don’t sit on a model.
Ricky Bell · 29 September 2026 · 4 min read
For an AI company, we hand our AI surprisingly little to do. That’s not a confession. It’s the design.
Take the audit. The AI doesn’t fetch your website — code does. It doesn’t measure your speed — Google’s PageSpeed Insights does, and we take that number as it comes. It doesn’t choose your competitors — code discovers and verifies them, and the model is expressly forbidden from inventing one. It doesn’t compute the overall score — that’s arithmetic, with fixed weights. It doesn’t store anything or send anything. What it does is the one thing only it can do: read your homepage the way an experienced reviewer would, judge the copy and the AI readiness against a fixed rubric, and write the findings in English a human wants to read.
Judgement is the job. Plumbing is not.
Language models are genuinely remarkable at judgement and language — and unreliable as calculators, random as plumbing. So the split writes itself. You hire a brilliant analyst for their judgement; you don’t also ask them to wire the building and remember every invoice. You give them systems for that. In ours, the model is the analyst. The code is the office around them.
The payoff is that the numbers can be trusted. Arithmetic can be checked. Fixed weights can be published — we did, in how we score your website — and there’s a written rule for what happens when a measurement fails: the report says “unavailable”. It never guesses.
A lot of AI products are a prompt in a trench coat
I’m sceptical of a good deal of the AI market, and I build with this technology every day. Too many products are an expanded search box: a prompt, an interface, a subscription. The tell is what happens at the edges — when the input is odd, when the model is wrong, when something upstream fails. If the answer is “the model handles it”, there is no system. There’s a hope with a login page.
The same split, in your business
When we build a tool for a client, the division of labour travels with it. The AI drafts the review response, qualifies the enquiry at nine in the evening, writes the first cut of the proposal. Code decides when it runs, where the data lives, what gets logged, and which actions wait for a human sign-off. The result behaves the same way on a Tuesday as it did in the demo — which, for a business tool, is the entire point.
Overall: the measure of an AI product isn’t how much the model does. It’s how little it’s allowed to do unsupervised. If you’d like to see this philosophy running, the free audit is it — every number on the report comes from the part we deliberately built to be boring.
Frequently asked
Is the Fordi audit just ChatGPT with a form on top?
No. Google’s PageSpeed Insights measures your speed, deterministic code discovers and verifies competitors, arithmetic blends the final score, and code stores and sends everything. The model’s job is judgement against a fixed rubric — scoring your copy and AI readiness the way an experienced reviewer would — and writing the findings in plain English.
Why not let AI do everything?
Because models are superb at judgement and language and unreliable as calculators and plumbing. A system is trustworthy when the parts that must never wobble — measurement, arithmetic, storage, sending — are boring code you can inspect, and the model is spent where judgement actually lives.
Does Fordi use AI in client work the same way?
Yes, the same split. The AI drafts the review response, qualifies the enquiry, writes the proposal draft. Code decides when it runs, where the data lives, what gets logged and what needs a human sign-off — so the system’s behaviour is inspectable, not vibes.