- Time
- 12 minutes
- You work in
- Mailbox
- Connect first
Gmail
Before you start
- An employee you have hired, of any role
- Gmail connected from Integrations
- Ten minutes of your own mail to test against
The promise of inbox triage is not that a machine answers your email. It is that the twelve messages that actually need you stop being buried under the eighty that do not.
That is a lower bar than most people set for it, and it is the one worth clearing first. Sorting is reversible. Sending is not.
Connect Gmail
Open Integrations and connect Gmail. This is a first-party connector, so it is a button rather than a search.
You are granting read, label, move and draft access. Nothing in this playbook sends anything, and it is worth being deliberate about keeping it that way until you have watched the sorting work for a week.
Decide your three piles, in your own words
The default framing is "needs you", "needs an answer", "noise". It is a good starting point and a bad ending point, because the interesting judgement is where your particular boundaries sit.
Write down what each pile means for you specifically. An invoice under a threshold might be noise to a founder and the entire job to a bookkeeper. A newsletter you actually read is not noise.
Have it label, not move, on the first pass
Create the labels in Gmail first, then brief the employee. Ask for labels rather than archiving, so that nothing disappears while you are still finding out whether you agree with it.
Every hour, look at unread mail in my inbox from the last 24 hours. Sort each message into exactly one of three labels, which already exist: AI/Needs-me something only I can decide or answer personally AI/Needs-answer a real question, but anyone on the team could handle it AI/Noise newsletters, notifications, receipts, automated mail Apply the label. Do not archive, delete, move or reply to anything. If a message contains a contract, an invoice, anything legal, or anything that reads like it is about money, label it AI/Needs-me regardless of the rules above. At the end, tell me how many went into each label and list anything you were genuinely unsure about.
Read the unsure list before anything else. It is the highest-value output of the whole job. Those are the messages where your rules were ambiguous, and fixing the prompt there is worth more than correcting fifty individual labels.
Watch it for a week before you let it move anything
Run it hourly and do nothing else for a week.
You are looking for one specific failure: something that landed in Noise and should not have. Everything else is a tuning problem. That one is a trust problem, and until it stops happening you should not let the job archive anything.
When a week goes by without one, change Do not archive to allow moving Noise
out of the inbox, and keep everything else as it is.
Only then, consider drafts
Once sorting is boring, drafting is a small step: the employee reads how you have answered similar threads and leaves a Gmail draft. Drafts are safe in the same way labels are, because nothing leaves your account without you pressing send.
Sending on your behalf is a genuinely different decision and deserves its own deliberate week.
What good looks like
You open your inbox, filter to one label, and deal with twelve things instead of ninety. The other seventy-eight are still there, still findable, and no longer in the way.
If that works for a month, the case for letting it draft has made itself.
Also covers
Clearing a backlog overnight. The same job, pointed at everything instead of at the last 24 hours, is how you deal with an inbox that got away from you.
Tonight, work through my unread mail older than 24 hours, oldest first. Apply the same three labels. Do not archive, delete or reply to anything. Work in batches and stop after 500 messages, telling me where you got to, so I can check the first batch before you do the rest. List anything you could not confidently label.
Two things make this survivable. Stopping after a batch means you review a sample before the whole backlog has been sorted on a rule you might disagree with. And keeping it label-only means a wrong decision costs a filter click rather than a lost message.
Expect the unsure list to be long on old mail. Age strips context, and a thread from March is genuinely harder to place than one from this morning.