The average knowledge worker now spends about a third of their day in their inbox. For founders and senior staff, that number is closer to half. Most of those emails don't need their attention. Most of the rest need a one-paragraph reply that takes ten seconds to think about and three minutes to write.

AI email triage takes the inbox, decides which emails matter, drafts replies for the rest, and surfaces only the messages that actually need you. Here's what it does and how to know if you need one.

What AI email triage actually does

The system sits behind your inbox - usually Gmail or Outlook - and watches new messages as they come in. For each one, it does three things:

  • Reads and labels. The email gets tagged with what kind of message it is: hot lead, cold pitch, client update, internal, newsletter, support, financial.
  • Drafts a reply where useful. If the message is something you'd typically reply to (a hot lead, a client question), a draft reply is sitting in your draft folder by the time you open the email. In your voice, with the right context pulled from your CRM and prior emails.
  • Routes the rest. Newsletters get filed, cold pitches get sent to a different label, finance emails go to a folder Ben checks once a day.

You open your inbox in the morning and see eight emails that genuinely need you, with seven drafts ready to review. The other 70 emails are already filed.

Where this earns its keep fastest

Founders, freelancers and partners. Anyone whose inbox is the bottleneck of their week. Specific signs:

  • You wake up to 100+ emails a day
  • You miss messages from real clients because they're buried in noise
  • You write the same kind of reply (intro, scoping question, follow-up) six times a week
  • You have an old promise to "get back to you next week" sitting in your inbox from three weeks ago

If you spend more than 90 minutes a day on email, the ROI on this is immediate.

How accurate are the drafts

This depends entirely on how much context the system has. A vanilla email drafter without access to your CRM, prior threads or business context will produce drafts that sound generic. A good build connects the agent to your sales pipeline, your past emails with that contact, your calendar, and any relevant docs. With that context, the drafts are usually 80% there. You read it, tweak a sentence, and send.

The 20% you change is exactly the personalisation that makes the email yours. The 80% the agent gets right is the boilerplate that drains your time anyway.

What labels make sense

Different inboxes need different label structures. The ones that work for most service businesses:

  • Action Required. Real emails that need a real reply from you specifically.
  • Waiting On. Emails where the ball is in someone else's court but you need to remember they exist.
  • Reference. FYI emails worth keeping but not actioning.
  • Pipeline. Sales-related emails grouped by stage.
  • Internal. Team emails separate from external.
  • Spam-adjacent. Cold pitches, newsletters, anything that's borderline noise.

The system gets better the longer it runs. After a month it knows your patterns - which "newsletters" you actually read, which "cold pitches" you secretly want to see, which clients always mark something as urgent that isn't.

What this won't do

It won't write your hard emails for you. The "we need to part ways" email, the "I made a mistake" email, the price-negotiation email. Those still need your full attention and your judgement.

It won't catch every nuance. Sometimes a client's casual line of "I'll have a think" actually means "I'm 90% on board, just need a nudge." A good agent flags this for human review rather than guessing.

It won't replace your inbox discipline. If your inbox is full because you treat it as your to-do list, no amount of AI will fix that. The triage works best for people who have a working inbox flow that's just over capacity.

What about privacy

This is the question we get most. The honest version:

  • The agent reads emails. That's the whole point.
  • For sensitive industries (legal, medical, finance), the build runs the AI inference locally or through a privacy-first provider with no training rights and Australian data residency.
  • The agent has the same permissions as the inbox owner. Nothing more.
  • You should still avoid sending state secrets or confidential client data through any AI, full stop, regardless of provider promises.

For most service businesses, the privacy posture is fine. For regulated industries, the build needs more care but is still doable.

What it costs

SaaS options exist (SaneBox, Superhuman, Shortwave) at $20-$50 per user per month. Useful for individuals, weak on context awareness.

A custom-built triage agent integrated with your CRM and calendar: $2,500 to $5,000 AUD, built in two to three weeks. Inference costs typically $30-$80 per user per month on top.

Most clients pay back the build in the first week of saved time.

Where to start

If your inbox is currently your biggest bottleneck, this is one of the highest-leverage automations you can deploy. Book a 30-minute call and we'll look at a typical day in your inbox to work out whether triage would actually help, and what the right first scope looks like.