AI agents for email: 57 real setups, and what people actually let them do
We read 57 real setups (118 counting looser matches) from people using Grok Bot, Muse, Instinct, Poke, Claude and others. The pattern is consistent: people let an email agent act alone when a mistake is cheap to undo, and make it wait when it writes in their name. Among the 13 cleanup and 19 triage setups we read, one of each does.
That rules out a tidy “read, then draft, then act” ladder. Drafting is the rarest job in the set, at 10 of 118 rows. People jump straight from reading to acting when the stakes are low, and they slow down only when the outbound message carries their name. This guide follows that line, from the most undoable work to the least.
Cleanup: act alone, with one backstop
Cleanup is the clearest case of acting alone, because nearly everything it does can be reversed or ignored. People in the data had Muse delete 14,000 junk emails and unsubscribe from the companies they picked. Only one of the 13 cleanup rows describes an approval step.
The best setups still place a backstop on the one step that can’t be undone. One user plugged a custom email API into Grok Bot to clear a 100k-email inbox, with spam removed and important attachments backed up. That works because the backup turns a bulk purge into something recoverable. Another user’s Grok Bot cleans spam every morning and unsubscribes only after a confirmation, which saves about 10 minutes a day. That works because the gate covers only the unsubscribes, not every sorted email.
A clean inbox makes the next job possible. Sorting means little while half the mail is noise.
Triage and digests: reading, not deciding
Triage is still reading work, so it also runs alone. Grok Bot appears in 9 of the 19 triage rows we checked by hand, more than any other tool, and only one of those rows describes an approval step.
The useful versions cut the volume of pings rather than repackaging it. One user relies on Poke to flag important Gmail faster than they would check it, with normal Gmail notifications silenced. That works because the agent replaces a noisy channel with a filtered one. A Grok Bot called Dispatch scans email, Slack, LinkedIn and X DMs every night, sends calendar invites for meetings that don’t exist yet, and feeds a morning briefing. That works because the scan happens on a schedule, not when the user remembers to ask.
A digest tells you what arrived. It doesn’t tell you what you forgot to send, so that becomes the next job.
Follow-up checks: catching what you owed
Follow-up checks are where an agent starts noticing gaps instead of just sorting mail. One user found an important email sitting unsent in Drafts and had Grok Bot compare their calendar against sent mail every day to flag any follow-up they owed but never sent. That works because it cross-checks two sources that no person compares by hand. A French-speaking life-admin bot named Alfred watches critical threads and calendar deadlines alongside its triage.
Once the agent knows what is owed, it can do something about it, which is where email becomes an input rather than the work itself.
Email as the source for an action
The biggest practical win is using the inbox as evidence for a task somewhere else. Muse found a receipt in Gmail and filed a dependent care FSA claim end to end, logging into the FSA account itself. That works because the receipt is proof and the task has a clear finish line. Instinct got a $200 warranty refund for a broken backpack by finding the email receipt and handling the back-and-forth with support. A ChatGPT agent read a Lowe’s protection plan from Gmail, filed a claim and got $48.15 back.
Only 2 of the 16 email-to-action rows we read describe the user approving a step. The closest is Instinct building a Bali visa from emails and Airbnb receipts, then asking “Did I miss anything?” before submitting. These agents write to companies, and a claim form is easy to correct. Writing to people in your name is different, and that is where the gates appear.
Drafts in your voice: where the gates live
Six of the 10 drafting rows describe an approval step before anything goes out, against 4 of the 48 cleanup, triage and email-to-action rows combined. That gap is the thesis in one number.
The gates take different forms. A supplier follow-up run with Grok Bot drafted a bilingual note, waited for an OK, kept CC rules and then sent eight replies. That works because one review covers a batch of near-identical sends. A Claude user outside the US built a 36-task routine where every email waits as a draft until they hit send. Muse prospected 50 accounts and left personalized outreach in the drafts folder for one-click approval. A newsletter’s Grok Bot sales agent drafts outbound pitches, but nothing goes out on its own, and the user edits and sends from Gmail.
The ungated cases are personal and specific. One user with a severe illness had Muse draft and send three detailed medical emails they had no energy to write. It was a single, one-off task the user asked for, not a standing permission to send.
If the agent writes a lot, the next step is to stop sharing your inbox with it.
Giving the agent its own address
A separate address lets you control exactly what the agent sees and acts on. One user gave Grok Bot its own inbox and CCs it on threads, with the instruction on the last line (“Book Tuesday at 3pm and get back to them”). The agent then books the calendar and replies from its own address. That works because the user decides thread by thread what to delegate, and nothing is sent as them. The same idea shows up as a shared-inbox router built with JEV and Mastra.
How to set it up
Hand jobs over in order of reversibility, and wait about a week between steps.
- Label only: let the agent sort into fixed labels and send a daily summary.
- Add a VIP list so the people who matter always land in one bucket.
- Turn on cleanup with a backup and confirmed unsubscribes.
- Allow drafts in your tone, left in Drafts and never sent.
- Add specific auto-actions last, one at a time.
Triage my Gmail every 2 hours between 8 AM and 8 PM. Put every new email into exactly one label: Reply-Today, Waiting-On, FYI, Receipts, Newsletters, or Junk. Anything from people in my VIP list (I'll paste it below) is always Reply-Today. For Reply-Today, write a draft reply in my tone (short, friendly, no exclamation marks) and leave it in Drafts. Don't send it. Once a day at 6 PM, send me a summary: how many emails per label, the Reply-Today list with one line each, and any newsletter I haven't opened in 30 days with a suggested unsubscribe. Never delete, unsubscribe, or send without my yes.Where to start
The data gives a simple rule: grant autonomy by reversibility, not by how clever the task is. If you trust an agent with nothing yet, start with a read-only digest and a follow-up checker. Both are useful on day one and can’t embarrass you. Once those hold up, add cleanup, with a backup and confirmed unsubscribes. Next, point the agent at receipts and claims, where the job has a clear end and the recipient is a company. Leave your voice for last: drafts you approve in batches, then a separate address the agent can send from. The same reversibility logic carries into home management, where school and bill emails feed the family calendar, and into shopping, where receipts become refunds.