How to build AI inbox triage that drafts your replies

AI inbox triage is a scheduled prompt that reads your overnight email, ranks it, and leaves draft replies in Gmail before you open your laptop. On Zo Computer you build it in three passes: sorting first, drafts second, a real-time VIP ping third. The drafts stay drafts. You keep the send button.

Key takeaways

• AI inbox triage on Zo Computer is one plain-English prompt on a schedule, not a workflow builder or a Zapier chain.
• Build it in three passes, one week each. Pass one sorts, pass two drafts, pass three interrupts.
• Only pass three should ever interrupt you. Everything else lands quietly and waits.
• Keep the no-send rule structural, not a polite request in the prompt: Zo leaves every reply as a Gmail draft, and a focused persona with only the tools this job needs keeps a sideways prompt boxed in.
• The average office worker receives 121 emails a day and sends 40, per cloudHQ's Workplace Email Statistics 2025. That is 605 messages a workweek to sort.
• Every scheduled run spends credits whether it pings you or not. A 30-minute weekday window costs 100 runs a week. Every five minutes around the clock costs 2,016.

What is AI inbox triage, and what does it actually do?

AI inbox triage is software that reads incoming mail in full, classifies it by what you need to do about it, and surfaces a short ranked list instead of a wall of unread. Filters match on senders and keywords. AI email triage reads meaning, which is why it can tell a customer asking about a refund apart from a newsletter about refunds. cloudHQ's Workplace Email Statistics 2025 puts the average office worker at 121 emails received a day against 40 sent, and no amount of folder discipline fixes a 3-to-1 ratio by hand.

On Zo it is not a separate product you buy. An automation is Zo running a prompt on a schedule with the same files, connected apps, and judgment it uses in a live chat. Connect Gmail once through the Connect card in the chat, then describe the job in English. Zo converts your wording into a recurrence rule and confirms before saving.

Why build it in three passes instead of all at once?

Because a prompt that sorts, drafts, and pings on day one gives you no way to tell which part is wrong. The 7am list flags a thank-you note as urgent, a draft lands in the wrong tone, your phone buzzes during a call. Three bugs, tangled. Split them.

Add the end-of-day sweep and the weekly sent-mail scan below and you land at 116 runs a week. The VIP ping is 100 of them. Keep that ratio in mind when you're tempted to poll every five minutes.

Pass one: the prompt that only sorts

Nothing but read and rank. No drafting, no sending, no filing.

Every weekday at 7:00am Eastern, read my Gmail from the last 14 hours.
Sort each message into one of these buckets: needs a reply from me today,needs a reply this week, FYI only, and noise. Email me the first two buckets, ranked most urgent first, with sender, subject, and one line on why it landed in that bucket. Skip the noise bucket entirely. If both lists are empty, send one line saying the inbox is clear.

Run it five mornings before touching anything. That one-line reason is what people cut to save space, and cutting it is the mistake. It's your only window into how Zo ranks, and it tells you which words to fix.

Which categories are worth defining?

Categories that point at a next step. "Client" tells you nothing. "Client waiting on me" tells you what to open first. Four or five buckets is enough, and real AI email prioritization comes from the boundaries you draw, not the names you pick.

Define the draftable bucket by hand. Write your repeat questions into the prompt by name: pricing, availability, where's my invoice, can we move the call. Vague instructions make vague drafts.

Pass two: how do you get AI draft email replies waiting in Gmail?

You tell Zo to write the reply, save it as a Gmail draft addressed to that person, and report what it created. The draft sits in Gmail. You open it, edit a line, send it yourself.

Every weekday at 7:45am Eastern, find emails from the last 14 hours that ask one of my repeat questions: pricing, availability, where to find an invoice, or rescheduling a call. For each one, write a reply in the voice of my notes file at [path to your reply-templates file] and save it as a Gmail draft addressed to that sender. Do not send anything. Email me the list of drafts you created with sender and subject. Only draft when you're confident the question matches one of my four; flag the rest for me to write from scratch.

Two phrases carry that prompt: "do not send" and "flag the rest." One is the guardrail. The other stops Zo from confidently answering a question it half-understood. A wrong draft you catch costs ten seconds. A wrong answer you send costs an apology plus the original problem.

Pass three: when should your inbox be allowed to interrupt you?

Only for the handful of senders where a two-hour delay costs you something real. Named accounts, investors, fresh inbound asking about pricing or a demo. Everything else waits for tomorrow's 7am list.

Every 30 minutes between 8am and 6pm Eastern on weekdays, check Gmail for new mail from [name], [name], anyone at [keyclient.com], or anything that reads like fresh inbound: someone asking about pricing, a demo, or
availability, from a sender I haven't emailed before. Text me only if one arrived, with sender, company, and what they asked. Stay silent otherwise.

That last sentence is the conditional clause, and it separates an alert you trust from one you mute by Thursday. Zo still runs all 20 times a day. It just refuses to reach you unless what it found earns the interruption.

How do you make sure it never sends?

You remove the ability instead of asking for restraint. An automation runs with the same tool access as the persona behind it, so a triage job under your default all-access persona can send mail while you sleep. One ambiguous instruction is all it takes.

Build a persona for this. Give it read access to Gmail, permission to create drafts, and one outbound channel pointed at you. Nothing else. No posting, no shell, no file deletion. A missing tool is gone from the conversation entirely, so Zo cannot fumble it. Settings live in the persona scopes guide. Start read-only, add the one channel you need, stop. Worst case becomes a bad summary in your inbox instead of a bad email in a client's.

The two sweeps that catch what slipped through

Morning passes handle what arrived. These two catch what you dropped, which is the costlier miss.

End of day, unanswered. "Every weekday at 5:30pm Eastern, find email I received today that looks like it needs a reply and that I haven't answered. Email me the short list with sender and subject. Skip newsletters and automated mail. Only send if there's at least one." Fifteen seconds to read. It's the one that saves the deal you meant to answer at 11am.

Weekly, open loops in your sent mail. "Every Wednesday at 9:00am Eastern, scan my sent Gmail from the last two weeks for messages where I promised to send something or follow up, then check whether I did. Email me the ones that look unresolved. Only send if you find at least one." Read that twice before building it. Reading your sent mail is a bigger grant than reading what other people sent you. Skip this one if that bothers you. The other five work without it.

What goes wrong

"Needs a reply" is a judgment call, not a rule, and week one will prove it. Zo will flag a "thanks, got it" and miss a real ask buried in paragraph four of a long thread. That is the tradeoff for reading meaning instead of matching keywords. Fix the definition rather than abandoning the pass: name the senders who always need a reply, name the phrasings that never do, and tell it a message with no question and no request is FYI. Two rounds usually gets it close.

Credits are the thing nobody plans for. Every scheduled run is a full Zo session and bills like one, condition met or not. Three passes plus two sweeps come to 116 runs a week. That same VIP prompt on a five-minute all-day schedule fires 2,016 times a week. Twenty times the spend, and it buys you an average of roughly 12 minutes on inbound. Stretch the interval, tighten the condition. Plan details live at zo.computer/pricing.

Then there's the failure that looks like success. Gmail's token lapses after a password change, the schedule keeps firing, the Gmail call fails, and nothing looks broken from outside because a quiet automation is supposed to be quiet. Two habits cover it. Add "if you can't reach Gmail, text me" to every prompt, and when a job goes silent too long, open the run and read it. Every run is saved as a conversation with the prompt it executed and the tools it called.

One rule here has no exceptions. If you catch yourself adding "and send it" to a triage prompt, delete those words. Zo drafts. You send.

Frequently asked questions

What is AI inbox triage?

AI inbox triage is an automated pass over incoming email that reads each message in full, classifies it by urgency and required action, and hands you a ranked short list. Unlike a filter, it reads meaning rather than matching keywords or addresses.

Can I automate email triage without Zapier or n8n?

Yes. Gmail automation without Zapier works because you describe the job in plain English and it becomes a scheduled automation, with no canvas, nodes, or webhooks to wire. Zo turns your wording into a recurrence rule and runs the prompt against your connected Gmail.

Will it send emails on my behalf?

Not unless you build it to. Zo writes each reply as a Gmail draft and leaves it for you to review. Keep that gate by sending the drafts yourself and running the automation under a focused persona that has only the tools it needs, so an ambiguous prompt has no path to an outbound send.

How often should the VIP ping run?

Every 30 minutes across working hours suits most people. That's 20 runs a weekday. Shorter intervals multiply credit spend without meaningfully improving inbound response time.

Why is it flagging emails that don't need a reply?

Because "needs a reply" is a fuzzy judgment and Zo is guessing at your intent in week one. Tighten the prompt: name your real VIPs, name the categories that never need an answer, and add an explicit rule for messages with no question or request.

Does this reach inbox zero automatically?

No, and be suspicious of anything that claims it does. This is inbox zero automation in the practical sense: the sorting and the first draft are done before you sit down, which turns a 45-minute clearing session into about 20 minutes of deciding and sending.