The most common request I get isn't "build me an AI system." It's "add AI to this" — where "this" is a process that's already unreliable, poorly defined, or held together by one person's memory of how it's supposed to work. That request, taken literally, usually makes things worse, not better.
Automation doesn't fix ambiguity, it multiplies it
A manual process with an unclear step still gets caught, eventually, because a human doing it notices when something looks wrong and asks a question. An automated version of the same process doesn't ask questions. It executes the ambiguous step exactly the same way, every time, at whatever speed and scale you gave it — including at 3 AM, including a thousand times before anyone notices.
I've seen this play out directly: a reporting process that occasionally miscounted because of an edge case nobody had documented. As a manual weekly task, it was wrong maybe once a quarter, and someone caught it. Automated without fixing the underlying logic first, it would have been wrong every single day, silently, because nothing was watching for the mistake anymore.
What "fixing the process first" actually means
It's not a call for months of process documentation before you're allowed to touch automation. It's smaller and more practical than that:
- Know exactly what "done correctly" looks like for the task, in specific, checkable terms
- Know what the failure modes are — not hypothetically, but the ones that have actually happened
- Have a way to notice when the automated version fails, because it will eventually hit a case you didn't plan for
If you can't describe those three things about a process, that's the actual work to do before adding AI to it — not a blocker to automation, but the design spec for it.
Where AI genuinely earns its place
The systems I've built that actually stuck weren't the ones that replaced a person's judgment. They were the ones that removed the boring, well-defined, high-volume parts of a job — running the same deliverability test every day, generating the same report from five data sources, watching a hundred servers for the same failure pattern — and left the judgment calls to the person who could actually make them.
That's a less exciting pitch than "AI automation," but it's the version that survives contact with a real operation.