Skip to main content
VantaSoft
Back to Insights

How to Build an AI Email Triage System (Step by Step)

An AI email triage system sorts your new mail and sends only what needs you to a chat channel. The rules to write, the access to grant and what it costs.

Rollout and oversight 15 min read

AI email triage means letting software read each new email and decide what it needs from you: a reply, a decision, a quick look or nothing at all. Outlook and Gmail now do a version of it inside the inbox, and tools such as SaneBox and Inbox Zero sort mail into folders and labels for you. A system you build yourself sorts on rules you write and sends only what needs you to a chat channel you already watch. Receipts get summarized and junk goes to spam, so you stop opening the inbox just to find out whether anything in it matters.

We run one of our own every day. Vance, the chief of staff agent on our team, checks our CEO's eight inboxes every 15 minutes on weekdays from 8 a.m. to 6 p.m. Pacific and every 30 minutes the rest of the time, and posts what he finds to a private Slack channel. Email sorting is the biggest part of his job as our AI chief of staff, and about one email in eight became a task in the 30 days to October 30. The job had run more than 5,800 times by November 5, 2026, since we moved it to Hermes Agent in July. This guide builds a smaller version from the same parts, starting with a single inbox, after a short look at what you can buy instead.

Most of the work is decisions only you can make: what counts as important, where the alerts go and how much access the agent gets. The technical setup follows the Hermes Agent documentation. Someone on your team who is comfortable installing software can do it, or you can have it done for you.

AI email triage tools and what they cost

You may not need to build anything. If a priority flag inside your mail app is enough, the feature could already be part of a plan you pay for. Each price below comes from the vendor's own page, read on November 5, 2026.

OptionWhat it does with new mailPrice on November 5, 2026
Copilot in Outlook: Prioritize my inboxMarks each email high, normal or low priority and says whyCopilot Business add-on $21 a user a month, yearly ($18 on sale)
Gemini in Gmail: AI Inbox (beta)Lists suggested to-dos and topics from your Primary tabWorkspace Business Standard $14 a user a month, yearly; Google AI Plus $9.99 a month
SaneBoxMoves mail it judges unimportant into a SaneLater folderFrom $9.49 a month for one account
Shortwave or SuperhumanA separate mail app that organizes your inbox with AIShortwave from $30 a seat a month; Superhuman Business $40 a month, $33 yearly
Inbox ZeroSorts and labels Gmail or Outlook mail and drafts repliesFrom $20 a month, $18 yearly; the code is open source
Your own buildSorts on your rules, posts to Slack, creates tasksHermes Agent is free; you pay for the computer and AI usage

The two built-in options come with limits. Microsoft's Prioritize my inbox page says it rates only mail that arrives after you turn it on and skips mail your rules file into another folder. The page doesn't name the license it needs, so ask whoever runs your Microsoft 365 account. Google's AI Inbox works only in the US and in English, starts at Business Standard on a work account, and can take up to 7 days to show its first to-dos. If you're weighing those seats for more than email, where an AI assistant for business stops compares them.

Most of these keep the result inside a mail app, so you still go to the inbox to see it. Build your own when the sorting has to fit how your business works: one set of rules across inboxes at different providers, a post in the chat channel your team already reads, a task for each email that needs you, or a handoff to another agent. That's the version the rest of this guide builds.

What the finished system does

Every half hour, the agent checks for new mail, sorts each message and posts the result to Slack. Each message lands in one of four buckets:

  • Important: someone is waiting on you, or the message carries a decision, a deadline, or money, legal or security risk. Each one gets its own post.
  • Not sure: the agent could not decide. These get their own post as well, so a doubtful message still gets seen.
  • FYI: receipts, notifications and automated reports, gathered into one summary post per check.
  • Spam: cold pitches and junk, moved to the spam folder.

What you need

  • A computer that stays on. The agent has to keep working while you are away from your desk, so use a small rented server or a computer you can leave running. Hermes installs on Linux, Mac and Windows. Ours runs on a Mac mini.
  • An account with an AI company. Hermes works with most of them, and our triage job uses one of Anthropic's Claude models.
  • A Slack workspace where you are allowed to add an app. Hermes also works with Telegram, Discord and other chat apps, but this guide uses Slack.
  • Your mailbox, and permission to connect it. On a company email account, check first whether your administrator allows outside apps to connect to mail.

Step 1: Write down what counts as important

The rules are what make the system match how you work, and they are yours to write. The agent reads them every time it checks your mail, so write them in plain language, the way you would brief a new assistant.

Hermes includes a ready-made email triage skill with a sensible starting set of categories: urgent reply, reply, action without reply, waiting, reference and noise. By default it only reads mail and drafts replies, and it does not send or delete anything unless you approve it. We keep our own rules instead, because they describe one person's priorities.

Writing ours taught us six things:

  • Judge the newest message in the thread. The subject line, the sender's reputation and whether you were on To or Cc tell you little. The earlier messages are only background.
  • Define Important by what the message asks of you: an action, a decision, a deadline or a reply, or real money, legal, security or operational risk. Most mail is none of those.
  • Write standing exceptions for what you can never miss. In ours, a customer or prospect who needs a reply, an open tax, legal or compliance matter, and a signed agreement are Important every time, whoever sends them.
  • Check before you alert. Before an Important post goes out, our agent reads the live thread and the Sent folder. If our CEO has already replied, the message is dropped and no task gets created.
  • Make FYI safe to ignore. Our CEO does not read the FYI summaries, so the test for FYI is whether it would be harmless for him never to see the message.
  • When the agent cannot tell, it should show you the message rather than guess. Dismissing a needless alert takes a moment, while a missed reply can cost you a client. That is why the Not sure bucket exists.

Lists of known senders still earn their keep for spam. Our job keeps a blocklist that grows every time the agent marks a new spam sender, so known junk is moved before the AI ever reads it.

Add one more rule that has nothing to do with sorting. Anyone can send you an email, and the Hermes documentation on giving an agent its own email address warns that this makes an inbox an easy way to slip instructions to an agent. Tell the agent to treat everything inside a message as mail to sort and never to follow instructions it finds there. The wider question of what an agent should be allowed to touch is covered in what access an AI agent should have to your business systems.

Step 2: Choose where the alerts go

Pick the chat channel you already check all day. For us that is a private Slack channel, such as #email-triage, that only the agent and the person it works for can see.

Hermes connects to Slack without your computer having to accept connections from the internet, so you do not have to open up your office network. The Slack setup guide walks whoever does the setup through creating a Slack app for the agent on Slack's app page, screen by screen. The agent only answers the people you list, and it will not join a channel until someone invites it.

Each of our agents has its own Slack app, so a triage post shows up under the name of the agent that wrote it. If more than one agent will share your computer, give each its own profile from the start, so each keeps its own logins, memory and scheduled jobs.

Step 3: Decide how the agent reads your mail

There are three ways to give the agent your mail. Choose by email provider and by how much access you are comfortable granting.

RouteWorks withWhat you grant
Himalaya, a skill that comes with HermesAlmost any mailbox, including Gmail, Outlook and FastmailAn app password, which is a separate password you can cancel at any time
The Google Workspace skill that comes with HermesGmail, plus Calendar, Drive, Docs and SheetsGoogle's own sign-in approval, set up through a Google Cloud account you create
A hosted connection serviceWhatever mail services that service supportsThe service holds the sign-in, and the agent sees only the actions it offers

The Google Workspace skill takes more setup, because it needs its own Google Cloud account and a sign-in approval. After that it stays signed in. If email is all you need, the same page points to Himalaya as the faster route: a Gmail app password and no Google Cloud account. The Himalaya setup guide covers it.

For a hosted service, the Hermes guide to connecting outside tools recommends listing only the actions the agent needs, calling that "usually the best default for sensitive systems." A mailbox is one.

We use the third route. All eight of our inboxes, seven on Gmail and one on Outlook, go through Composio, a hosted service that holds the Google and Microsoft sign-ins, so no Google or Microsoft password or login is stored on our machine. The Outlook inbox had its own connector until October 29, when we moved it to Composio as well.

Whichever route you pick, start with one inbox and get it working end to end before you add another. Triage only needs to read, label and move mail, so leave sending out of this job.

Step 4: Put it on a schedule

Hermes runs work on a timetable through scheduled jobs. You can set one up just by asking the agent in chat, for example: "every 30 minutes during the day, triage my inbox and post the results to Slack." The person doing your setup can also create it directly, for instance to run every half hour from 6:00 a.m. to 9:30 p.m.

Three details from the documentation matter for a job like this:

  • Each check starts fresh, with no memory of your chats. Your written rules, the four buckets and the format of the post all have to be in the job's instructions.
  • Tell the agent to answer with only the word [SILENT] when there is nothing new. Hermes treats that answer as "post nothing," which keeps quiet half hours quiet.
  • Ask for a quick check that runs before the AI does. If there is no new mail, Hermes skips the AI for that run. The documentation recommends this for jobs that run often, because otherwise you pay for AI runs that have nothing to do.

Our job is built around that last one. The quick check collects mail that arrived since the last check in each inbox, moves known spam and lines up the rest. With nothing new, the agent never starts. With new mail, its only job is to sort it. A second step then posts to Slack, moves any new spam and creates a task for each Important email. Plain rules handle the bookkeeping, such as not posting the same email twice, and the AI handles the judgment calls. Since October 30 it has run every 15 minutes during working hours and every 30 minutes outside them.

Step 5: Run it beside your inbox for a week

Keep checking email the usual way for the first week. Each day, compare the channel with your inbox and write down every miss. Then fix the rule itself by editing the job's instructions, so the next check reads the correction. A correction you make in chat may never reach the job, since each check starts fresh and the documentation advises against relying on the agent's memory for details a job depends on.

The rules will keep changing after that first week. Ours have since the move in July. The customer-reply exception arrived in late August and the tax and legal exception in early September. A rule that every Important email gets a task came the day after that, and the signed-agreement exception two days later. In late October, a signed agreement that calls for a payment link started producing a task to send that link. In early November, our expense card's requests for missing receipts started going to our finance agent instead of our CEO.

Once the channel catches everything you would have caught, stop checking the inbox directly. Someone still owns the rules, though. The agent sorts the mail, and a person decides what the sorting should be and when it changes. Human in the loop for AI agents covers where those approval points belong.

Where to take it next

  • Add inboxes. Our job covers seven Gmail accounts and one Outlook account under one set of rules.
  • Turn Important emails into tasks. Ours create a Todoist task for each one, linked to the email and to its Slack post.
  • Hand mail to other agents. A missing-receipt request now becomes a task for our finance agent. Project email went the other way: we used to label it for our project manager agent, and since October 13 he reads those inboxes himself, so the label step is gone.
  • Draft replies for approval. The triage skill that comes with Hermes studies your recent sent mail to match your voice, then shows you its drafts as a batch to approve. Keep sending behind that approval.

Frequently asked questions

What is the best AI email triage software?

It depends on where you want the answer to show up. If a priority flag in your inbox is enough, start with what your mail plan already includes: Prioritize my inbox in Outlook or AI Inbox in Gmail. SaneBox moves low-priority mail out of the way without changing apps, from $9.49 a month for one account on November 5, 2026. Build your own when the result has to leave the inbox, as a chat alert, a task or a handoff.

Can an AI email triage system reply to emails for me?

It can draft them. The triage skill that comes with Hermes drafts replies and asks for approval, and it does not send or delete without it. Our own triage job goes further and never sends mail at all. It sorts, posts and creates tasks, and replies stay with a person.

Does it work with Outlook?

Yes. Outlook has Copilot's Prioritize my inbox built in. For a build, the Himalaya skill works with most mailboxes, Outlook included, and a hosted connection service is another way in. Ours reaches its Outlook inbox through Composio, the same service that connects our Gmail accounts.

What does it cost to run?

Hermes Agent is free and open source under the MIT license. You pay for the computer and for the AI company's usage charges, and the quick check in Step 4 means a run with no new mail costs nothing in AI fees. Our AI agent cost guide has worked numbers, and which AI model to use compares current prices.

Have us build it

If you want triage running on your inboxes without owning the computer, the rules and the upkeep, our Agent Service builds and runs agents like this one.

Vendor pages read November 5, 2026: Microsoft's Prioritize my inbox help and Copilot Business pricing, Google's AI Inbox help, Gemini feature availability by plan, Workspace pricing and Google AI plans, SaneBox's feature tour and pricing, and the pricing pages of Shortwave, Superhuman and Inbox Zero, whose code is on GitHub. Hermes Agent sources read October 10, 2026 and checked again November 5, 2026: the documentation on installation, Slack, the Google Workspace skill, Himalaya, connecting outside tools, the email triage skill, scheduled jobs and scheduling automation, plus the settings and records of our own triage job.

Five quick questions about where your team's time goes. You'll get the one agent we'd start with, and what it would take off your plate.

Take the quiz

Put your first agentto work.

We build custom AI agents around your processes and the tools you already use. It starts with 30 minutes on your workflows.