Working notes, with real numbers
Most AI content is written to sell something. These are engineering notes from building and running production systems — including the parts that argue against buying from us.
What AI automation actually costs in Canadian dollars
The gap between "AI is cheap" and the invoice. Real per-token arithmetic on a 5,000-document-a-month workload, what the cloud line adds, and why the model API is almost never the expensive part.
ResidencyWhere your data actually goes when you use a US AI platform
A practical walk through the request path — your app, the vendor's servers, the model provider's region — and what Canadian privacy law, U.S. sectoral rules and a procurement reviewer will each want to know about it.
EvaluationWhy a self-graded accuracy score isn't accuracy
Several agent platforms will generate their own evaluation criteria from the prompt, then score themselves against it. Here's how to build a held-out labelled set instead, how big it needs to be, and how to freeze it.
JudgementWhen not to automate a process
Four situations where automation loses money — low volume, unstable upstream process, exceptions that outnumber the happy path, and a task where being wrong is expensive. Written by someone who gets paid to build it.
New pieces land roughly every two weeks. No newsletter, no gate, no lead magnet — if you want to know when something's up, the sitemap is public and so is everything here.
Have a process worth measuring?
The readiness assessment applies this thinking to your operation — volume, cost per task, payback, and an honest read on whether automating it makes money.
See the assessment