How I work
Rules where rules work.AI where they don't.
Most automation disappoints for one reason: somebody used AI for a job a rule should have done. Here is how I decide which is which — and why it matters to you.
The difference
A rule
“When an order over $500 comes in, tell the account manager.”
- —Does the same thing every single time
- —Can be tested before it goes near a customer
- —Costs effectively nothing to run
- —Cannot invent anything
A job for AI
“Read this invoice, when all forty suppliers lay theirs out differently.”
- —No rule could ever cover every case
- —Needs checking, because it can be wrong
- —Costs money on every single run
- —Worth it anyway — the alternative is a person
Most of what a business wants automated is the first kind. It is boring, repetitive, and completely predictable — which is exactly what makes it automatable. When that work gets handed to AI it becomes slower, more expensive, and occasionally wrong in ways nobody can explain. That is why so many companies have an automation project they no longer trust.
So I build the predictable majority as rules, reach for AI only where the input is genuinely unpredictable, and tell you which parts are which — so you know exactly where the system can surprise you, and where it never will.
I take my own advice
Same person, same month, opposite tools.
There is a free tool on this site that compares what six thousand AI models cost across every company selling them. There is no AI in it at all. Comparing numbers is a job for arithmetic, and arithmetic does not have opinions.
Building it, I found sixteen prices in the industry's own public data that were wrong by a factor of a million — a typo that would have made one model look like it cost eight million dollars a month. A rule caught that, because a rule can be told “no real price is ever above this line.” I also cross-check two independent price sources and flag the sixty-five places they disagree, rather than quietly averaging them into a number that is wrong in a new way.
Meanwhile the deployment platform I built does use AI — for exactly one thing. When a release fails, it reads the error log. Error logs are unbounded: every failure looks different, and no rule could ever cover them all. That is a genuine AI job, and it sits inside a system where everything around it is rules.
That is the judgement you are hiring.
What to expect
- Plain English, always
- If I cannot explain what I am building and why, in words you use, that is my failure — not yours to decode.
- It runs on things you own
- Your servers, your accounts, your data. You can hand it to someone else, or to nobody, and it keeps working.
- Fixed price, agreed before I start
- You know what it costs and how long it takes. If it is more work than I thought, that is my problem.
- I will tell you when the answer is no
- If a week of my time will not pay for itself, I would rather say so than take the money and have you feel it later.
- You get shown, not told
- Backups get restored in front of you. New processes run alongside the old one until you trust them.