"Automation" gets thrown around until it stops meaning anything — a chatbot, a dashboard, a cron job, an AI model that summarizes a document. We use the word a lot too, so it's worth being specific about what we mean when we say it.
For us, automation is narrow by design. It's a manual, repeatable step in someone's workflow — the kind of task a person does the same way, over and over, because nobody has gotten around to removing it — rebuilt as a system that does it without a human in the loop. Not "AI that helps you work faster." Not a general assistant. A specific, boring task, gone.
That narrowness is deliberate. Broad automation promises — "we'll automate your whole business" — are usually a sign that nobody has scoped the actual work yet. The projects that hold up are the ones where we can point at the exact step being removed, describe what triggers it, and say plainly what happens when it fails. If we can't do that before we start building, we're not ready to build.
Large language models are part of the toolkit now, and sometimes they're the right part — reading unstructured text, drafting a first pass, flagging something for a human to check. But a model is a component, not a strategy. We treat it the way we'd treat any other dependency: useful where it's reliable, wrapped in checks where it isn't, and never the only thing standing between a client and a wrong answer they'll act on.
The honest version of "we build automation" is less exciting than the marketing version. It's fewer demos and more questions up front: what exactly happens today, who does it, how often, and what breaks if it's wrong. That's the work. The system is just what's left after you've answered those questions properly.