The one inbox job I'd hand an AI this week, and the one I'd keep to myself

Google is quietly building a triage inbox for Gemini. Here's the email job worth handing over now, and the button I never let it press.

TLDR: Every lab is racing to point AI at your inbox, and the loud promise is that it'll answer your emails for you. The higher-leverage job to hand over this week is the quieter one upstream: triage. Let the AI sort your mail into what needs you, what can be drafted, and what's noise, so your attention goes to the few threads that move money. Draft replies too, right up to what I call The Send Line, the point where a human presses send. Below: the two prompts I run and the threads I never hand over.

Google got caught building something telling this month. Testers spotted a dedicated inbox inside the Gemini app for Workspace, with three filters, follow up, done, and one labeled "Needs review," that pulls messages out of Gmail into a to-do list a person walks through (TestingCatalog, July 4). What jumped out at me was that "Needs review" filter, because it's an admission from the company with the most email data on earth that the AI's job is to sort and surface, and the human's job is still to decide.

By the end of this you'll have a morning routine that clears your inbox to a handful of real decisions in about ten minutes, plus a hard rule for where the AI stops and you start.

That's the same instinct I run my own inbox on. The first thing I do with any new AI tool that can touch my email is work out what it's allowed to do without me, and the answer is: read and sort, yes, press send, no. I laid out a version of this test yesterday for spreadsheets, the Bounded-Job Test, and the inbox runs on the same three questions, is the job bounded, does it run on data you already have, and can you eyeball the result and know it's right. Triage clears all three, while sending a reply on your behalf clears none of them, because you can't eyeball an email after it's already gone.

Here's what's inside:

• The triage pass: one prompt that sorts a day of unread mail into four buckets so you only read what needs a human.

• The Send Line: the boundary the AI drafts up to and never crosses, and why it's the whole game on the outbound side.

• The draft-to-the-line prompt: copy it, and every reply lands in your drafts folder in your voice, with the commitments left blank for you to fill.

• The threads I keep fully human: negotiations, complaints, anything with feelings or a contract in it.

Who this is for: operators running a real business off their own inbox, ChatGPT or Claude or Gemini already in the mix, drowning a little in email and tempted to just let the robot answer it.

Why triage is the job worth handing over

Email still eats around a quarter of the average workweek, and inbox zero stays the thing most people give up on by week two (Missive, 2026). Most of that time goes to deciding more than writing: opening things, scanning them, working out what deserves a reply and what doesn't, and carrying the low hum of a hundred unread items around all day. That deciding layer is the expensive part, and it's the part AI is genuinely good at right now, because sorting is bounded, it runs on the mail already sitting in front of it, and you can glance at the buckets and know instantly if it got them right.

Answering is the part everyone wants to automate and the part that keeps biting people. A draft that sounds right isn't the same as a draft that's right, and once it sends, there's no glancing at it first. So the move is to split the inbox job in two, hand the AI the sorting and the drafting, and keep the send button on the human side of a hard boundary. I call that boundary The Send Line. The AI can read, sort, summarize, and write a full draft, and it stops the instant a message would leave your account, because crossing that line is your call and not the model's.

If that sounds familiar, it's the outbound twin of something I named back in June. When AI browsers like Comet showed up promising to read your whole inbox, I wrote about Comet Walls, the permission boundaries you set so an agent can read a sensitive stream without acting on it. The Send Line is the same wall pointed the other direction: inbound, let it read; outbound, it drafts but you send. It's the same principle running on both sides of the mailbox.

The tools to do this are already sitting in plans you pay for. If you're on Google Workspace Business or a personal Gemini subscription, the Gemini side panel in Gmail already summarizes threads and drafts replies with Help me write and Help me reply (Google support). If you're a ChatGPT or Claude person, connect your Gmail and the same routine runs there. And if you'd rather not connect anything yet, you can paste a screenful of subject lines into any chat window and get most of the value on day one.

The morning triage pass

Here's how it works. Once a day, before you start opening things one by one, you hand the AI the last 24 hours of unread mail and ask it to sort, not answer. The whole point is to turn a scroll of forty things into four short lists so your eyes only land on the one list that needs a person.

The prompt I run, which works in the Gemini side panel, in ChatGPT or Claude with Gmail connected, or on pasted subject lines:

You are my inbox triage assistant. Here are my unread emails from the
last 24 hours [paste them, or read from my connected Gmail].

Sort every message into one of these four buckets, and nothing else:

1. NEEDS ME: a real person waiting on a decision, a reply, or money.
   One line on what they want.
2. CAN BE DRAFTED: a reply I'll approve before it sends.
3. READ LATER: useful, no action today.
4. NOISE: newsletters, receipts, automated notifications.

Rules:
- Do not send, archive, or delete anything. Sorting only.
- If a message could be money coming in or money going out, it goes in
  NEEDS ME, even if you're unsure.
- Flag anything that reads angry, legal, or like a cancellation, and put
  it in NEEDS ME with the word REVIEW next to it.
- NEEDS ME is capped at what a person could clear in 20 minutes.
  If it's longer than that, move the softest items down to CAN BE DRAFTED.

Give me the four buckets as short lists. Nothing else.

What a good result looks like: NEEDS ME is short, a handful of threads, and the money and the angry ones are sitting right at the top with REVIEW next to them. The bulk of the volume falls into NOISE and READ LATER, which is the whole win, because that's the pile you were burning attention on without meaning to. You read one list, you make a few decisions, you move on.

The failure mode to watch, and it's the common one: NEEDS ME comes back with twenty-five items in it. That means the model hedged and pushed everything up to the safe bucket, which gives you back the exact overwhelm you were trying to kill. When that happens, tighten the cap ("NEEDS ME holds at most eight items, everything else drops down") and it recalibrates. The rarer failure is an empty NEEDS ME on a day you know had a real ask buried in it, which is the model under-escalating. If you see that even once, keep reading manually alongside the AI for a few more mornings before you trust the sort, because a triage pass that misses the one email that mattered is worse than no triage at all.

Walk away with: a once-a-day routine that collapses a full inbox into four lists and puts the threads that move money at the top, in about the time it takes to make coffee.

Drafting up to The Send Line

Sorting gets you to the handful of threads that need a reply. The AI can write those too, and this is where holding the line earns its keep. The rule is simple and it never bends: the AI writes the draft, the draft lands in your drafts folder, and you are the one who reads it and presses send. Nothing leaves your account on the model's decision.

The prompt:

Draft a reply to this email in my voice. Do NOT send it. Leave it as a
draft for me to read and send myself.

The email:
[paste the thread]

My voice: [paste 2-3 of your own real sent replies here, or describe it:
direct, warm, short sentences, no corporate filler].

Draft rules:
- Match the length of the incoming message. A two-line email gets a
  two-line reply.
- If a date, price, or commitment is involved, leave a blank for me to
  fill: [DATE], [PRICE], [SCOPE]. Never invent a number.
- If the thread is a negotiation, a complaint, or anything with money or
  feelings on the line, do not draft a full reply. Write one line telling
  me what's at stake and let me handle it.
- End every draft with: READY FOR YOUR REVIEW.

A good draft comes back short, in your register, with blanks sitting right where the commitments are, and that READY FOR YOUR REVIEW line at the bottom so you know it stopped where it was supposed to. You read it in five seconds, fill the blanks, send. The response-time drop is real, and the teams running this at scale almost all keep send on the human side anyway, because auto-send stays rare even where it's technically supported, the cost of one wrong autonomous reply running much higher than the minutes it saves (Missive, 2026). You get the speed of a drafted reply and keep the judgment of a human read.

The failure mode here is the dangerous one, so calibrate on it hard. If the draft fills in a price or a date you never gave it, the reply is contaminated, the model guessed a commitment, and if that had been on auto-send it would've gone out as your word. That's the tell to keep those threads human and to make the "never invent a number" rule louder. The softer failure is a three-paragraph draft in response to a one-line email, which means your voice sample was too formal or you skipped it, so feed it two or three of your own real short replies and it snaps into your cadence.

Walk away with: replies written in your voice sitting in your drafts folder, commitments left blank on purpose, and a hard rule that the send button is yours.

What I keep fully human, and the honest tradeoffs

The reason The Send Line matters is that some threads shouldn't get an AI draft at all, and knowing which ones is the real skill. Charles Hudson, who runs Precursor VC, built agents that watch his inbox, classify it, and draft his intros and declines, and he still keeps the send button on his own side. He doesn't trust the agent to send on its own, so as he told Missive, "I have a draft only flag on" (Missive, 2026). That posture, an always-on agent doing the work with a hard draft-only stop on the send step, is the one I run too, and the threads I'd never even let it draft are the ones where sounding right is a trap: active negotiations with real terms on the line, complaints and conflict, anything legal or contractual, and any message that needs genuine empathy. The model produces something that reads fine in every one of those, and reading fine is the danger when the stakes are a client relationship or a signed contract.

A few things worth flagging before you wire this up. The triage sort is only as good as the model's read of your business, so the first week it'll misfile things, and the fix is feedback, correct the buckets a few times and it sharpens fast. The native Gemini side panel is bundled into Business and Enterprise Workspace plans and personal Gemini subscriptions, so if you're on a bare Workspace tier you might not see it yet, in which case the ChatGPT-or-Claude-plus-Gmail route does the same thing. And the deeper tradeoff is data, connecting any AI to your inbox hands it read access to everything in there, which is precisely why the read-but-don't-act boundary from Comet Walls matters, and why I'd start with pasted subject lines before I'd hand a brand-new tool a full mailbox connection.

There's also a real cost to over-trusting the sort. A triage pass that quietly misroutes the one email that mattered doesn't feel like a failure, it feels like a clean inbox, right up until you find the buried thread a week late. So the discipline is to keep glancing at the NOISE and READ LATER piles for the first couple of weeks, the way you'd double-check a new hire's work, because the AI can't learn what "important to me" means until you've corrected it a few times.

Recap

Point the AI at the deciding layer, not the sending layer. Run a once-a-day triage pass that sorts your unread mail into four buckets so you only read what needs a person, let it draft the replies that can be drafted, and hold The Send Line so nothing leaves your account without you pressing the button. Keep negotiations, complaints, legal, and anything with feelings in it fully in your hands. The industry consensus landed in the same place for a reason: these tools are co-pilots on the inbox, and the operators getting hours back are the ones who used them to decide faster, not to stop deciding.

If you run a booking calendar or a lead inbox, this compounds fast, because the same triage-and-draft pass is what keeps a fresh lead from sitting unanswered for six hours while you're heads-down. That gap between a lead landing and a human touching it is the kind of leak my Showtime skill pack is built to close. Worth a look if your inbox is where your money either moves or stalls.

Tomorrow: the calendar version of this, the one scheduling job I'd let AI run end to end and the one I'd never automate.

Andrew

P.S. A few to pull the thread further:

• Yesterday's Bounded-Job Test for spreadsheets, the same three questions applied to Gemini in Google Sheets.

• Comet Walls, the inbound version of The Send Line, for when an AI browser wants your whole inbox.

• The Standing Brief, the one-page business brief that makes every draft sound like you.

• The Showtime skill pack if your lead inbox is where the money leaks.

• New here? The newsletter lives at muddventures.substack.com, a short tactical read most weekday mornings.