I Let AI Run My Inbox for a Week

My inbox hit 40 unread by 8 a.m., and instead of grinding through it, I handed the whole thing to an AI assistant. Triage, sorting, first drafts, the works. Everyone says automation frees you up for the important stuff. Mostly true. But the week taught me something sharper: the tool was brilliant at the boring parts and quietly reckless at the parts that actually mattered.

I've done inbox zero the hard way for fifteen years, across Fortune 500 orgs and federal agencies where a sloppy email could end up in a FOIA request. So I went in skeptical. Here's the field report.

What it got right

The mechanical work is where it earned its keep. Within a day it was:

• Sorting by actual priority, not just sender. It learned that a two-line note from a specific client outranked a long deck from a vendor.
• Drafting replies to the routine stuff. Scheduling, "got it, thanks," "here's the file you asked for." Clean, fast, in roughly my voice.
• Killing the low-stakes back-and-forth. The four-email dance to book a call collapsed into one.

Call it the mechanical 80 percent. The volume that fills your day without moving anything. For about a week I watched it clear that layer before I'd finished my coffee, and I'll admit the relief was real. My inbox stopped feeling like a second job.

If the story ended there, this would be a product review. It doesn't.

What it got embarrassingly wrong

The other 20 percent is where it tried to bury me.

A partner emailed asking if we could "pull the timeline forward." The AI drafted a warm, agreeable yes and had it queued to send. I caught it on the way out. I had not agreed to anything, and saying yes would've committed two people to a deadline I knew we'd miss. The draft read perfectly. Confident, friendly, wrong.

Then a touchy client sent one of those emails where the words are polite and the temperature is fifteen degrees below zero. The AI read the words. It drafted something cheerful and slightly brisk, exactly the tone you do not use with someone already deciding whether to stay. A human who knew the relationship would've softened, slowed down, maybe picked up the phone. The AI saw a question and answered it.

None of these were dumb errors. That's the trap. They were helpfully wrong, plausible enough to sail through if I'd been moving fast, expensive enough to matter if they'd landed.

The failures don't announce themselves

Here's the part that kept me up. A bad auto-reply looks exactly like a good one.

There's no error code. No red flag, no bounce. A smooth wrong answer and a smooth right answer are the same shape on the screen. This is the thing people miss about handing judgment to a model: the errors are silent, and they compound. One over-eager yes becomes a commitment, becomes a missed deadline, becomes a harder conversation three weeks later. By the time anyone notices, the trail is cold.

That's not a knock on the tool. It's the nature of the thing. It generates the most probable next sentence, and the most probable sentence is usually the agreeable one. Agreeable and correct part ways more often than you'd like, especially where money or feelings are involved. Which is precisely where you can't afford a silent miss.

So I stopped trusting anything I couldn't see fail.

Where I landed

The week didn't sour me on the idea. It sharpened it. I kept the setup, with one rule bolted on: AI drafts, a human sends anything with a relationship or a dollar attached.

• Scheduling, confirmations, file sends, the connective tissue: let it fly.
• Anything touching a deal, a deadline, a negotiation, or a person who's currently annoyed: draft only. I read every word before it leaves.

The real win wasn't automating my judgment away. I don't want that, and the week made clear I can't have it. The win was that clearing the mechanical noise gave me more room for the calls that need a person. Some decisions aren't binary. The AI is very good at pretending they are, which is exactly why you keep your hand on the send button.

It ran my inbox for a week. I'm keeping it on. I'm also keeping the receipts, because the one time it almost cost me, it did it with a smile.