An operator, not a commentator

Nearly everything written about AI automation is written by people whose automation has never had to survive a Monday morning.

I run automation inside a print and design manufacturing business. Orders arrive, artwork gets checked, jobs move across a production floor, packages get consigned to carriers, and customers phone up when any link in that chain fails. The systems I write about here are the ones I have to answer for when they break.

That gives this writing a specific character. It is long, because the interesting detail is never in the summary. It is concrete, because vague advice about "leveraging AI" costs nothing to give and nothing to follow. And it is often unflattering — about vendors, about models, and about decisions I made that turned out badly.

What you will find here

Case studies of automations that went into production, with the numbers attached: what it cost to build, what it saved, how long before it paid back, and how often it needed a human to intervene. Honest evaluations of tooling. And a running argument that most businesses should automate their operations long before they automate their marketing.

What you will not find

Affiliate links. Roundups of "the 47 best AI tools." Predictions about AGI. Posts generated to hit a publishing quota.

Why anonymous

Because the useful parts of this writing are the parts a person with a job cannot usually publish. What a system actually cost. Which vendor's software failed, and how. The decisions that turned out badly, described with enough specificity to be useful to someone about to make them.

Attach a name and an employer to that, and the writing bends. It becomes careful about suppliers who might read it, generous about tools it needs support contracts for, and quiet about the mistakes. The result is what already fills the trade press: technically accurate, commercially inoffensive, and worth nothing to the person actually trying to decide something.

So the byline is a pen name and stays one. Everything else is verifiable. Where I cite a figure, it is a real figure from a real system. Where I cannot verify something, I say so rather than smoothing over it. You are welcome to weigh the writing on that basis, which is the only basis that ought to matter.

Contact

If you are working on something similar, or think I have got something wrong, I would like to hear it. The best correction I have received came from a reader who had run the same experiment two years earlier and watched it fail differently.

The Compositor — Operator, print & design manufacturing.