1981. A ZX81, a cassette tape for storage, and the first proper lesson anyone ever taught me about computers: GIGO — garbage in, garbage out. You don't hear the term much any more. It's still the truest thing anyone's ever told me about technology, AI included.
The idea is almost insultingly simple: feed a system rubbish, and no amount of clever programming rescues you from rubbish coming back out. Doesn't matter how elegant the code is. Doesn't matter how powerful the machine. Garbage in, garbage out. Full stop.
Here's what most AI marketing to local businesses gets backwards: it sells you the model. Look how clever it is, look what it can write, look how fast it answers. All true, and almost beside the point. A modern AI model is dramatically more capable than anything I touched on that ZX81 — and GIGO applies to it exactly as hard as it did in 1981, just with higher stakes, because now the "system" is your actual business, not a homework program.
Feed an agent a stale contact database, and it will confidently, efficiently, and very quickly help you waste money contacting people who left years ago. Feed it a vague, unexamined process, and it will automate the vague, unexamined process — just faster. The model isn't the bottleneck. It never was. What goes in is.
It's why the flagship example I use elsewhere on this site — auditing a contact database before doing anything with it — starts with checking what's actually true, not with sending anything. Audit first. Know what's garbage and what isn't. Only then act. That order isn't a nice-to-have; it's the entire discipline, and skipping it is the single most common reason "AI automation" projects for small businesses quietly fail to deliver anything real.
It's also why the Working Diagnosis is built the way it is — a consultation based on your actual working day, not a generic questionnaire. A generic questionnaire is itself garbage in. You'd get a generic answer out, dressed up as a diagnosis. That's not worth £199 of anyone's money, mine or yours.
Not from reading about AI — from four decades of watching real inputs go right and wrong, across some very different businesses.
Door-to-door sales and telesales, early on, where you learn fast that a bad lead list is worse than no list at all. Corporate spells at Microsoft, Applix (business intelligence software, later sold to IBM), and Salesforce, where I saw what "good data" actually looks like inside serious enterprise software, not just claimed.
Joining Selling People as its first employee and building the team up around me until I was running it as General Manager — a UK agency that gave overseas tech companies a real UK market entry, sales and marketing teams built from nothing, which only works if you know exactly what you're feeding into the pipeline.
Running offshore customer service operations at Ellery Communications (2000–2004) at real scale — up to 8,000 agents worldwide at one point — where garbage input across thousands of people and several time zones stops being an abstract problem and becomes a very expensive one, fast.
Website and lead-generation consultancy, email marketing, a CRM I built from scratch myself, and a stretch heading up venture capital investment in agritech worldwide. Every one of those jobs was, underneath, the same question: what's actually going into this, and is it any good?
I'm not the person to ask if you want the deepest possible explanation of how a model works under the hood — plenty of people can out-technical me, and I wouldn't pretend otherwise. I'm the person to ask if you want someone who can look at your actual business and tell you, honestly, what's garbage and what isn't, before anything gets automated. That's not a smaller skill than knowing the technology. Forty-odd years of evidence says it's the one that actually decides whether any of this works for you.
This is what the Working Diagnosis is actually for — finding the real input before anything gets automated, not after.
See how it works