Email triage vs. Gmail filters: why one wins

I have an embarrassing number of Gmail filters. 47, last I counted. They sort newsletters into Reading, route alerts from a few systems into Ops, and tag emails from my investors with a VIP label.

They also fail constantly. Here's why — and what a triage layer does differently.

Filters match strings. That's the whole feature.

A Gmail filter is a one-line program: if the sender, subject, or body contains some literal string, do some action. That's it. There is no understanding of intent, no awareness of urgency, no model of the relationship between you and the sender.

So filters break in three predictable ways:

1. 1New sender, no rule. A new investor reaches out from a Gmail address you've never seen. Your VIP filter doesn't fire. The email lands in Promotions. You see it three days later.
2. 2Same sender, different intent. Your accountant sends two emails: one with a tax deadline (urgent) and one confirming a payment (informational). Same sender, same domain. The filter treats them identically.
3. 3Same words, different urgency. An automated newsletter with "meeting reminder" in the subject and an actual personal meeting reminder both hit your Meetings label. You miss the personal one because it's buried in a folder of 200 automated ones.

Triage understands intent

A triage layer like Triagd reads each email like a junior assistant would. It considers: who sent it, what the relationship is, what the email is actually asking, whether there's a deadline, what the sender's pattern of communication looks like.

Concretely, that means:

• First-time sender from a venture firm gets Tier 1 if the email contains investment language and the firm matches a pattern of legitimate VC outreach — even without any user rule.
• Tax-deadline email from your accountant is Tier 1; payment confirmation from the same accountant is Tier 3. Same sender, different content, different tier.
• Automated newsletter with "meeting reminder" goes to Tier 4 because the sending domain pattern is automated; the personal meeting reminder from a human goes to Tier 1.

None of that is possible with string matching. All of it is table stakes for a model that's been trained on what email actually looks like.

What filters still do better

Filters are deterministic. When you write a filter, you know exactly what it will do — every time, forever. A model is probabilistic; it will get it right 94% of the time and wrong 6% of the time, and you cannot perfectly predict which times.

For some operators in some workflows, that determinism is non-negotiable. Compliance emails that must go to a specific folder. Customer-support tickets that must hit a specific queue. We don't try to replace filters in those cases; we run alongside them. Triagd respects existing filters and complements them.

The hybrid that actually works

The right setup, in practice, is filters for the 5% of email you can deterministically describe — and AI triage for the 95% you can't. Filters for compliance routing. AI triage for everything else.

When we onboard a new Triagd user, we audit their existing filters and keep the ones that are working. Most users have 2–4 filters worth keeping (out of their accumulated 30+). The rest are scar tissue from past inbox triage attempts that no longer match the current shape of their email.