An engineer who manages thousands of WordPress sites had an inbox problem most of us would find unsurvivable: every site emits alerts, and they pile into the thousands. So he wired an AI agent to his email and ran what he called a conversation with a score to beat. The result, in his own write-up: over 1,600 emails triaged and closed in a single day across two sessions — roughly 900 monitor alerts and 182 recovery-mode notices — surfacing four sites with real PHP fatal errors hiding in the noise.
It’s a genuinely impressive number. But the most instructive moment in the story isn’t the 1,600. It’s the one the human had to catch.
The mechanics scale. The judgment doesn’t.
At one point the agent confidently misdiagnosed an issue — attributing a problem to a guard that, the human knew, wasn’t actually new to that version. He pushed back, the agent re-examined, and only then found the real root cause. Left unchallenged, a tidy, plausible, wrong conclusion would have sailed straight through.
That’s the real division of labor emerging in this kind of work. The agent is extraordinary at the mechanical part — reading 1,600 messages, grouping them, closing the noise, drafting the responses. The human is still the one who knows that this explanation doesn’t smell right. Volume is the machine’s job; judgment is yours.
Triage is the killer use case — with a catch
Most people’s task lists have the same shape as that inbox: a pile where 90% is noise and 10% genuinely matters, and the work is finding the 10%. That’s exactly what an AI assistant is good at — turning a long, intimidating list into a short one worth acting on.
The catch is the same as the inbox: you want the agent to surface and sort, not to silently decide. A triage assistant that quietly closes the wrong thing is worse than no triage at all. The pattern that works is agent-proposes, human-confirms — fast, but with a person in the loop on the calls that matter.
How Lume frames it
That’s how Lume’s assistant is built to work: ask it what’s overdue, what’s slipped, what to focus on, and it does the heavy sorting across your whole list in seconds. But it surfaces and suggests — completing a task is still your move, and every change is undoable. You get the 1,600-emails-in-a-day leverage without handing over the judgment that makes triage trustworthy.
The future of getting through a list isn’t an agent that runs unsupervised. It’s one that does the reading so you can do the deciding.