AI inbox triage: leave only what actually needs a human

Most of your mail doesn't need a decision from you — it just needs to be filed. Notifications, CCs, newsletters, order confirmations. You still scroll through it anyway, because it sits in the same pile as the three emails that genuinely need your head. And that pile costs more than it looks like: it isn't the reading, it's the dozens of micro-decisions of "is this for me?" that leave you tired without having accomplished anything.

The goal of triage isn't inbox zero — it's an inbox that contains only work for a human. Everything else should be tidy, searchable, and visible whenever you choose to look. This guide walks you from your first filter to a routine that sorts your mailbox on its own every day — and, above all, to the rules that make it trustworthy.

The process has six phases: categories by action (not by topic), dividing the work between hard filters and AI, an escalation rule for mail that must never go into hiding, the daily routine's instructions, a weekly audit of false positives, and boundaries that never get crossed. One rule governs all of it: triage may move and label, never delete. A misfiled message can be found in a folder; a deleted one can't.

A typical scenario

Lenka, an accountant, gets around ninety emails a day. Realistically, about eight of them are actually hers to deal with. The rest are automated messages from the billing system, CCs from projects where she's only there for information, and informational mailing lists. She still skims every single one, at least with her eyes, afraid something important might be hiding among them. That's half an hour a day spent scrolling through things that require no decision — and she still misses something once a month anyway, because an important message got buried among fifty unimportant ones.

After setting up triage, her mailbox looks different. The main folder holds nine messages, all from people, all waiting for a reply. Automated messages sit in a System folder, informational CCs in a To Read folder, mailing lists in Newsletters, and anything the routine wasn't sure about lands in a Review Sorting folder — usually one to three items a day.

Lenka checks the main folder in the morning and the others once a day after lunch. Or she skips it on a packed day, and nothing breaks. After a month, what surprised her most was that she stopped missing things: messages from her boss and her three biggest clients are on the escalation list, so the routine may never move or label them no matter what they say. Wherever the automation could cause damage, it simply doesn't touch anything.

Phase 1: categories by action, not by topic

This is the decision the rest of the system stands on. Most people set up labels by topic — projects, clients, areas — and then discover it didn't help. The reason is simple: when you open your inbox, you're not asking what a message is about. You're asking what you need to do with it. A "Project Alpha" label doesn't tell you whether something's waiting on you today or whether it's just a CC.

Four categories are enough

• Reply today — someone is asking a question, wants something from you, is waiting for a response. This is the only category that stays in the main inbox.
• Delegate — belongs to someone else on the team; you're just forwarding or assigning it. Its own folder, because it's a different kind of work than replying.
• Read — CCs, meeting notes, reports, industry newsletters you actually read. It's about content, not action; you go through it in one block.
• Archive — order confirmations, automated notifications, invoices from systems. Not for reading, just needs to be searchable.

Four is a ceiling, not a target. Every extra category means more deciding for every message — the exact opposite of the point. When a fifth one tempts you, ask what other action it triggers; if none, it belongs in one of the existing four.

One refinement on naming: name categories with a verb, not a noun. "Reply today" tells you what to do. "Important" tells you nothing, and a month later that folder holds thirty messages nobody ever comes back to.

Categories built on your actual mail

The generic scheme is a starting point, not the finished product. The precise boundaries come from what actually lands in your mailbox.

I'm attaching a list of 60 emails from the last two weeks —
for each one, the sender, subject, and first two sentences.
Names and companies are anonymized.

Help me set up triage by ACTION, not by topic.
Return:

1. A sort of all 60 messages into these categories: reply
   today, delegate, read, archive. For each message, the
   reason in one word.
2. Messages that didn't fit any category, and a suggestion
   for what to do with them — either where they belong or
   why I need a fifth category.
3. For each category, write a three-to-five-sentence
   definition that automatic triage will use. The definition
   has to be decision-ready: what signal identifies the
   message, what sets it apart from the neighboring
   category, and one borderline example from my list.
4. Three category pairs that will most often get confused,
   and a rule that resolves the tie.
5. How many messages from my sample would stay in the
   main inbox.

For point 3, don't write generic definitions like "important
messages" — write them so that someone who doesn't know my
job could decide correctly from them.

You'll get back four usable definitions and — more valuable — a list of borderline cases you'd have gotten stuck on yourself. Check point 5: if thirty out of sixty messages would stay in the main inbox, the definitions are too soft and need tightening, or you'll end up with triage that doesn't actually sort anything. Save the finished definitions to a file or a Project — you'll be coming back to them every week.

What to do with what's already sitting in your inbox

New rules handle new mail. Cleaning up two thousand accumulated messages retroactively is a separate operation, and the safest way to do it is as a one-time, manually directed pass — not through the routine.

I have [2000] messages older than [14 days] sitting in my
main inbox. I don't want to sort them one by one and I don't
want to delete anything.

Suggest a one-time cleanup in steps:
1. Which groups of messages can be moved in bulk without
   reading them (by age, sender, domain, bulk-mail header)
   and where to. For each group, estimate how many messages
   that is.
2. Which ones, on the other hand, I need to look at with my
   own eyes because they might contain unfinished business —
   and how do I filter for them to keep that number as low
   as possible.
3. What search queries to use for this in [Gmail /
   Outlook], write them exactly as I could paste them in.
4. What order to do the steps in so they don't overlap.
5. What to do at the end so the old mail stays searchable
   but doesn't get mixed up with the new triage.

No step may involve deletion. Where you're not sure about
a count, say so instead of a number.

You'll get back an afternoon's worth of steps. Always run the search queries from point 3 first and look at the result before you act on it in bulk — a query that's one domain too broad moves a thousand messages that shouldn't have gone anywhere.

Phase 2: where filters end and AI begins

The second key decision. AI is a tempting answer for everything here, but for a large share of your mail it's needlessly expensive, slower, and less reliable than a rule that's existed for thirty years.

Hard rules first

A rule is free, instant, and never gets it wrong. When a message always comes from the same address or always has the same string in the subject line, it belongs on a filter, not sent to a model. It typically catches sixty to eighty percent of the volume:

• automated messages from the systems you use (billing, timekeeping, monitoring, CRM);
• notifications from tools (task trackers, calendars, shared documents);
• newsletters and mailing lists — identifiable by an unsubscribe footer or a bulk-mail header;
• order and shipping confirmations;
• CCs where you're in the CC or BCC field and not in the To field.

A dedicated tip covers filters for mailing lists in newsletter filters in Gmail; for Outlook the same applies via rules. Build these before you reach for AI — a model that gets a clean input makes noticeably fewer mistakes than one buried under notifications.

What a rule can never tell

What's left is mail from people, and there a filter fails, because the decision depends on content. A rule can't tell "please send a quote" (reply today) apart from "just letting you know the deadline changed" (read), even when both come from the same person with the same subject line, "Order 2024/118." It can't spot a question buried in the fifth sentence of a long email. It can't tell that "thanks, sounds good" closes out a conversation and needs nothing further.

This is work for AI, and it's the only part where it pays off. Leave the rest to filters.

Here's an export of my current mail filters and rules:
[insert list]

And here are 40 messages from the last week that got past
them into the main inbox:
[insert sender, subject, first two sentences]

Do an analysis:
1. Which of those 40 messages would a plain rule have caught?
   For each, write the specific condition (sender, domain,
   text in subject, bulk-mail header) and the target folder.
2. Which ones, on the other hand, require judging the
   content and can't be sorted by a rule? For each, say
   exactly what would have to decide it.
3. Which of my existing rules overlap or contradict
   each other.
4. Which rules are too broad and risk catching something
   they shouldn't.

Don't propose new rules for messages from specific people
whose content varies — a rule doesn't belong there.

You'll get back a division of labor between filters and AI, plus a review of what you already have. Points 3 and 4 are most useful for people who've been adding filters for years: overlapping rules produce behavior nobody can explain ("why did this email end up in archive?"). Review the proposed conditions before turning them on — the model sometimes proposes a rule based on a word that commonly shows up elsewhere too.

Phase 3: the escalation rule

This is the most important section of the whole guide, and also the one people typically set up only after a first unpleasant experience. Set it up now.

Automatic triage has one asymmetry: a mistake in one direction costs a second, a mistake in the other costs trust. When a newsletter stays in your main inbox, you glance at it and move on. When an email from your boss, with a subject that looked like an ad, ends up in the Read folder and you see it three days later, that's a problem serious enough to make you turn the whole system off.

The fix isn't a better model — it's a hard list of exceptions that isn't up for discussion.

The "never filter" list

The list should include: your direct manager and company leadership, key clients (by name, not category), addresses that contracts and legal matters come from, messages where you're in the To field and the sender is inside your own company, and anything from a circle where a delay has real consequences — for an accountant, the tax office; for operations, the supplier of a critical component; for engineering, an outage report.

The rule should be phrased as a prohibition, not a recommendation. Not "be careful with these messages," but "you may not move, label, or hide these messages, under any circumstances, even if they look like a mailing list."

Help me build an escalation list for the automatic triage
of my mail — recipients and situations that triage must never
touch a message for.

Context: [role], [industry], [company size].
Who I report to: [manager's role].
My key counterparts: [3–5 types, not names].
What can't tolerate delay in my work: [description].

Return:
1. A list of sender categories that belong on the escalation
   list, and one sentence for each explaining why.
2. Situations recognizable by content, not by sender
   (escalation, complaint, deadline today, approval
   request) — for each, say what signal identifies it.
3. Rule wording I can paste into the routine's instructions.
   Write it as a prohibition in the imperative, not as advice.
4. Three situations where the list could end up too broad
   and triage would lose its point.
5. How often I should review the list and what should
   trigger a review (a role change, a new client, an
   ended engagement).

Don't write specific people's names into the output, I'll
fill those in myself.

Read point 4 carefully — an escalation list tends to grow until half your mail is back in the main inbox. And put point 5 on your calendar: the list goes stale quietly, typically right when a client leaves or your manager changes. Add names to the list by hand; you don't need to put them in the prompt.

Phase 4: the daily routine run

Now you turn the rules into a task that runs on its own every day. It assumes mail is connected via a connector — how to set that up and what a connector can see is covered in the reply-draft routine; the process is the same. For triage, the right to read messages and add labels or move them between folders is enough. Never the right to delete.

The complete routine instructions

Every workday at [12:30], sort my mail.
Follow this exactly and don't make anything up:

1. SCOPE
Take messages delivered to the main inbox since the last run
(at most 24 hours back). Don't touch the [HR, Legal]
folders, or messages older than 24 hours.

2. ESCALATION RULE — TAKES PRIORITY OVER EVERYTHING ELSE
Messages from [list of addresses and domains], and messages
that look like an escalation, a complaint, an approval
request, or a deadline within 24 hours, must be LEFT IN THE
MAIN INBOX unlabeled. Don't move them, even if their content
looks like another category. Just list them in the summary.

3. SORTING
Sort the remaining messages into one category per the
attached definitions: reply today / delegate / read / archive.
- leave "reply today" in the main inbox
- move everything else to the folder with the matching name
- decide the category by what ACTION the message requires,
  not by its topic
- for threads, decide based on the last message, not the first
- never put a message that contains a question addressed
  to me, anywhere in the text, into "archive"

4. UNCERTAINTY
When you're not sure how to classify something, DON'T GUESS.
Move the message to the [Review Sorting] folder and, in the
summary, note which two categories you were torn between and
why. More uncertainty flagged is better than a silent mistake.

5. PROHIBITIONS
Don't delete anything, don't archive anything permanently,
don't mark anything as read, don't reply to anything, don't
send anything. Only move messages and apply labels.

6. SUMMARY
Write an overview, up to [5] lines per category:
- how many messages in each category; for "reply today"
  list the sender and topic
- what fell under the escalation rule
- what ended up in [Review Sorting] and why
- which definitions weren't enough to decide with

You'll get a sorted mailbox and a five-line summary. Verify three things right after the first run: that the escalation rule really does take priority (send yourself a test from a trial address on the list), that the routine didn't touch older messages, and that the uncertainty folder isn't empty — an empty one means the model is guessing instead of admitting it doesn't know.

When to run it

Once a day, in the middle of the day. A morning run doesn't make sense, because most of your mail arrives during the morning; 12:30 or 1:00 is a good compromise. If you get a lot of mail, you can add a second run at the end of the workday.

Don't run it every five minutes. Continuous triage means messages move out from under you, which is unpleasant and, more importantly, makes it impossible to check anything — you never know whether an email was there and disappeared, or was never there at all. One run a day creates a clear boundary: whatever's in the main inbox after the run belongs there.

Two weeks running dry

Before you run the routine for real, let it run for a week proposal-only: nothing gets moved, it just writes down what it would have filed where. A correction at this stage is free, while a correction after two weeks of running live means digging through folders.

Trial triage run — today, don't move anything, don't label
anything, don't change anything. Just propose a classification
and show me your reasoning.

Go through the messages from the last 24 hours in the main
inbox and return a table:
sender | subject | proposed category | confidence
(high/medium/low) | what you decided it on
(a specific word or sentence from the message, not a
general reason) | falls under the escalation rule yes/no

Below the table, write:
1. which messages had low confidence and what you were missing
2. two messages you're still not sure about even now,
   and how you'd classify them on a second look
3. which of my definitions turned out unclear in practice
   — quote the specific wording that didn't help
4. how many messages would be left in the main inbox
   per your proposal

Don't inflate your confidence. "Medium" is a legitimate answer.

The "what you decided it on" column is the whole point of the trial run: you'll see whether the model is deciding based on content or based on sender and length. If a reason like "looks like a notification" shows up five times, that's a signal it's deciding based on form and will get the first atypical email wrong. Five such tables in a row are usually enough to get your definitions right.

Phase 5: the weekly false-positive check

Triage that nobody checks quietly degrades: new senders show up, your workload shifts, definitions go stale. Ten minutes a week takes care of it.

Two kinds of mistakes and their cost

A false positive is a message that left the main inbox when it should have stayed. That's the expensive mistake — you don't see it until someone follows up a second time.

A false negative is a message that stayed in the main inbox unnecessarily. A cheap mistake: you glance at it and move it.

That tells you how to tune it. When you're unsure, set it up so the message stays. Triage that's cautious is usable; triage that's ninety percent accurate and quietly hides the other ten isn't.

Audit: what you filed wrong last week

The single most valuable prompt in this whole guide. Run it once a week, ideally on a Friday.

Audit your own triage for this week. You have access to my
mail, so compare what you filed with what I actually did
about it.

Return:

1. FALSE POSITIVES — messages you moved out of the main
   inbox that I then pulled back out, replied to, or opened
   within [24 hours]. For each, write: sender, topic, where
   you put it, why you thought it belonged there, and what
   you should have recognized instead.
2. LEFT UNNECESSARILY — messages that stayed in the main
   inbox and I just moved elsewhere without replying. For
   each, what rule is missing.
3. UNCERTAINTY — the contents of the [Review Sorting]
   folder: where they ended up belonging and what you were
   missing to decide.
4. PATTERNS — are the mistakes random, or do they repeat
   for a certain type of message, sender, or wording? List
   at most three patterns, with two examples each.
5. ESCALATION — did you touch anything on the escalation
   list? If so, list it first, even if it's a single message.

Be critical of your own work and don't tell me it went well.
Where you don't have enough data to conclude something,
say so instead of guessing.

Read point 5 first — touching the escalation list is the one mistake you can't let slide. Points 1 and 4 together usually show that the mistakes aren't random: they typically cluster around one message type (automated messages that contain a question, say, or long threads). The method in point 1 is an inference from your behavior, not a certainty, so spot-check the findings against your actual mailbox.

From audit to rules

Findings alone fix nothing. Once a week, turn them into a definitions update.

Here are the findings from this week's triage audit:
[insert audit output]

And here are my current category definitions and escalation list:
[insert]

Propose specific changes:
1. For each recurring mistake, say whether the problem is
   in the category definition, in the escalation list, or
   whether a plain filter is missing — and so where the
   fix belongs.
2. Rewrite the affected definitions. Show the old and new
   version side by side and say what changed in one sentence.
3. Check that the new definitions don't contradict the rest
   or overlap with each other.
4. Flag rules that never came into play over the past
   month and suggest cutting them.

Don't add rules for the sake of adding them — when the same
thing can be solved by editing an existing definition, do that.
Keep the whole set of definitions to one page.

Point 4 fights bloat: rules pile up easily and are hard to remove, until you end up with a system nobody understands that behaves unpredictably. One page of sharp definitions works better than five pages of comprehensive ones.

When it's done: when the audit finds zero false positives three weeks in a row and the uncertainty folder holds only a handful of items a week. Then move to a monthly check.

Phase 6: boundaries that never get crossed

• Never delete. Triage may move and label. Not delete, not permanently archive, not mark as read — a message marked read and filed away is, from a human's perspective, the same as deleted. Don't grant the automation delete permission even at the technical level, if that's configurable.
• Never reply. Triage is triage. If you also want reply drafts, that's a separate routine with its own rules and the same principle: AI proposes, the human approves — only a human ever sends.
• Sensitive folders stay out. HR, legal, and health matters are excluded from the routine explicitly in the instructions, not by a "be careful" rule.
• Company account. Triage sends message content to a model. It belongs on a work account with contractual data protection, not a personal chat.
• Minimum content out. In the instructions, ask only for a classification and a short reason, not a retelling of the message. A summary that contains the content of ninety emails needlessly widens what leaves your mailbox.
• Quarterly access review. Check which services the routine has access to and what it actually uses. Permissions quietly pile up.

The last point can be scheduled directly as a quarterly task:

Prepare the material for my quarterly triage routine review.
Don't change anything, just report the current state:

1. Which services and data you have access to, and what of
   that you actually use for triage.
2. Which folders and labels the routine created or changed
   over the past quarter.
3. Which folders it excludes, and whether that list matches
   my rules — list the differences, not the matches.
4. Everything you process from message content during
   triage, and which of that shows up in summaries.
5. Which categories were barely used over the quarter and
   could be dropped.
6. Three questions I should ask myself during this review
   that you're not able to answer for me.

Where you're not sure, write "I don't know" instead of guessing.

Point 6 is deliberate: the model can't see your company policy or your data processing agreements. Use the output to disconnect what you're not using — the fewer access rights, the less that can go wrong.

Common mistakes

• Sorting by topic instead of by action. Labels named after projects look tidy and solve nothing — you still don't know what to do first in the morning. What should decide is the action a message requires.
• Letting AI handle what a filter can do. Newsletters and system notifications belong on a rule: it's free, instant, and never gets it wrong. The model should only handle cases where content is what decides.
• Not having an escalation list. One hidden email from your manager is enough to make you shut the whole system off. A "never filter" list is cheaper than that experience.
• Allowing automatic deletion. A misfiled message can be found in a folder. A deleted one can't. This boundary has no exception, not even for obvious spam.
• Forcing the model into a decision. Without a permitted "I don't know" answer and an uncertainty folder, you get silent mistakes instead of visible ones. Uncertainty belongs out in the open.
• Setting it up and never checking again. Without a weekly audit, accuracy quietly degrades as your work changes and new senders show up.

The best tools

• Built-in filters and labels in Gmail, rules and Focused Inbox in Outlook — the cheapest, fastest step, and your data never leaves your account. Start here so AI gets a clean input.
• Claude with a mail connector — reads and files messages directly in your mailbox per the categories you described, runs as a scheduled task once a day, and can even audit its own work weekly.
• Gemini in Workspace and Copilot in Microsoft 365 — triage and summaries inside an environment your company has already approved; the least friction with IT.
• SaneBox — a ready-made service for separating low-priority mail into its own folders; it works on top of your existing mailbox without switching clients or building your own logic.
• Zapier, Make, or n8n — for when you want custom categories and connections to other systems; n8n for when mail content shouldn't leave your own infrastructure.

What you get out of it

• Time: 20–40 minutes a day depending on your mail volume. It's not just the reading — it's mainly the dozens of micro-decisions of "is this for me?" that cost attention.
• Money: half an hour a day is roughly ten hours a month. For billable work, that's a day and a half; for a manager, it's room for things they'd otherwise never get to.
• Peace of mind: your inbox stops being an endless list. Nine items are manageable; ninety aren't — and the difference mostly comes down to whether you sit down to your mail with dread.
• Fewer things missed: counterintuitively, the risk of overlooking something also drops. Important messages aren't hidden in the noise anymore, and the escalation list watches exactly what it needs to.

Pro tip

An advanced trick that costs five minutes: have the routine list, once a week, senders whose messages you never opened over the past month. It's not about sorting, it's about the source — every such sender is a candidate either for unsubscribing or a hard filter, and eliminating mail entirely always beats sorting it cleverly. After two rounds of this, daily volume usually drops by tens of percent, and what's left is easier to sort.

And the final rule: triage should be cautious, not precise. A system that leaves a message in the main inbox whenever it's in doubt is one you'll keep using for years. A system that decides confidently and quietly hides something once a month gets switched off after the first mess — and with it, all the time it saved you up to that point. Add fixed email-reading blocks on top of it, and your inbox stops being a place you keep checking in between work.