The reason “AI for inbox triage” disappoints most people isn’t the AI. It’s that they ask the model to sort emails without telling it what matters. Urgent to whom? Important compared to what? Models are good pattern-matchers, terrible mind-readers. Give them your sorting rules in writing and they sort like you would. Don’t, and you get the same generic “high/medium/low priority” labels that don’t match how your day actually works.
Zapier’s 2024 survey of administrative work flagged email triage and meeting summarisation as the two most common AI uses among assistants. The ones getting real value share a habit: they wrote down their triage logic once and now paste it into every prompt.
Build a sender tier list before touching the inbox
Open a blank document and list every recurring sender or category your executive cares about, sorted into four tiers.
- Tier 1, drop everything: your executive, the CEO if your exec reports up, three to five named clients or board members, your IT team if a system is down. These get a notification, not a triage entry.
- Tier 2, same-day reply: direct reports, key vendors mid-project, anyone awaiting a deliverable from your exec this week.
- Tier 3, batch and brief: scheduling requests, vendor check-ins, internal updates from teams two or more layers away.
- Tier 4, archive or unsubscribe: newsletters, automated platform notifications, recruiting outreach you don’t pursue.
Anyone not on the list defaults to Tier 3. This list lives at the top of every triage prompt you write from now on. It’s the difference between a model that triages your inbox and a model that triages a generic inbox.
The morning briefing prompt
Open Microsoft Copilot, ChatGPT, or Gemini with full inbox access (Copilot natively in Microsoft 365, ChatGPT/Gemini via export or paste) and run this:
“You are my inbox triage assistant.
Sender tiers: Tier 1 (escalate now): [LIST] Tier 2 (same-day): [LIST] Tier 3 (batch): [LIST PATTERNS, E.G. ‘ANY VENDOR INVOICE QUERY’, ‘ANY CALENDAR INVITE FROM OUTSIDE THE ORG’] Tier 4 (archive candidate): [LIST PATTERNS]
Read the unread emails below. For each, output one row: Tier | Sender | One-line summary | Suggested action (escalate / draft reply / forward to [NAME] / file / delete) | Deadline if any
At the top, give me a three-line briefing: how many Tier 1, how many Tier 2, and the single email I should open first. Don’t include Tier 4 in the table, just give me a count and an option to bulk-archive.
[PASTE OR ATTACH UNREAD EMAILS]”
What you get back is the briefing the original post promised, except it’s calibrated to your job. The “single email I should open first” line is the load-bearing one. It forces the model to make a judgement call instead of flattening everything into one priority bucket.
The follow-up prompt for Tier 2 batches
Once Tier 1 is handled, Tier 2 is where the real time goes. Drafting six replies one at a time is slower than drafting all six at once with a shared voice. Paste the Tier 2 list back in:
“For each Tier 2 email above, draft a reply.
Constraints: under 100 words each, match the voice of an executive assistant supporting [EXEC NAME, ROLE], polite but not effusive, never use ‘I hope this email finds you well’ or ‘don’t hesitate’. Sign off [NAME].
If a reply needs information I haven’t given you, write ‘[CONFIRM WITH EXEC: question]’ in square brackets in place of that detail. Don’t guess.
Output as a numbered list keyed to the senders above.”
The “[CONFIRM WITH EXEC]” instruction is what stops the model from inventing details. Without it, you’ll get a confidently fabricated meeting time or budget number, send it without checking, and have a mess on your hands by Wednesday.
The failure mode that buries Tier 1
Here’s the trap. AI triage tools, when given a giant inbox, tend to weight by recency or by reply length, not by importance. A Tier 1 email that arrived overnight can sit below a Tier 3 thread that just had eight rapid responses. Two protections.
First, in your morning prompt, include the line: “Tier 1 emails always appear at the top of the table regardless of recency or thread length.” Make it explicit.
Second, never trust a tool to triage Tier 1 in isolation. Set a native rule in your mail client (Outlook rules, Gmail filters) that flags or stars anything from a Tier 1 sender automatically, before AI sees it. AI is doing the bulk sort. The rule is your safety net.
A counterintuitive observation
Most assistants who try AI inbox triage save less time than they expected on the triage itself, but a lot more time than they expected on the replies. Sorting 80 emails into tiers takes a model maybe 20 seconds, which is fast but isn’t life-changing. Drafting 12 Tier 2 replies in one batch is where the hours come back, because that’s the work that used to fill your morning. If you’re only running the triage prompt and skipping the batch-reply prompt, you’re stopping halfway through the workflow that actually matters.
What stays in your hands
The model is sorting and drafting. It is not sending. Every Tier 1 escalation, every flagged ambiguity, every reply involving a number or a date gets your eyes on it before it leaves your account. The promise isn’t an inbox that runs itself. It’s an inbox where the first 90% of the work is done and you’re spending your judgement on the last 10%, which is where the value was anyway.
The assistants who run the cleanest inboxes aren’t the ones with the cleverest tools. They’re the ones who wrote down their sorting logic once and now have a model that applies it consistently every morning, before they’ve finished their first coffee.