Four-Stage AI Inbox Triage
Filter visibility, urgency, and decisions before AI drafts the response.
- Difficulty
- Moderate
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 98%
Wilkinson's Lindy workflow turns email into a sequence of small classification decisions. The first agent asks whether he needs to see a message at all and archives low-value thread replies, which he says removes about 20 percent of incoming mail. The next stage identifies messages requiring action within 24 hours and applies a dedicated label, separating emergencies from total volume. Another agent decides whether the email contains a simple choice. If so, it privately offers Wilkinson several response options; he selects a number, and the system writes and sends a thoughtful reply as him. Human judgment remains at the decision point while AI handles reading, routing, and composition. The design replaces one overloaded assistant task with several narrow agents.
Origin
Wilkinson built the workflow himself in Lindy after trying to automate every task a human could perform in his inbox. He says email work previously handled by a full-time assistant is now automated.
Core principles
- 01The first automation decision is whether a human needs to see the message.
- 02Urgency should be separated from inbox volume.
- 03Simple decisions are easier when converted into explicit options.
- 04AI can handle composition after the human supplies intent.
- 05A workflow can contain several specialized agents rather than one broad prompt.
How to run it
- 1
Filter required attention
Read each incoming message and decide whether the user needs to see it. Archive thread acknowledgments and other mail that adds no decision or useful information.
Pro tip Begin with obvious low-risk patterns such as reply-all confirmations.
Watch out A false archive can hide information, so the classification boundary must be reviewed.
- 2
Separate urgent mail
Classify whether the message must be handled within the next 24 hours and apply a dedicated label. This gives the user a bounded emergency queue.
Pro tip Define urgency by required response time, not by the sender's tone.
- 3
Detect simple decisions
Identify messages that can be resolved through a small choice, such as accepting, delaying, or declining a lunch invitation. Route complex messages back to the normal inbox.
Watch out Do not force ambiguous or multi-part decisions into an oversimplified menu.
- 4
Offer private options
Send the user a concise multiple-choice prompt describing the decision and possible intents. Let the user respond with the option number rather than compose the full message.
Pro tip Include variants that preserve timing and tone, such as yes later or different ways to decline.
- 5
Compose from the chosen intent
Turn the selected option into a polished reply and send it in the user's voice. Keep the decision human while automating the writing and delivery.
Watch out Review early drafts until the system consistently represents the user's intent.
In the wild
If Lenny emails Wilkinson asking to get lunch, the workflow privately asks whether Wilkinson wants to accept, accept in a few months, or decline in one of several ways. Wilkinson replies with an option number, and the system writes the full email back to Lenny as him.
→ The human makes the relationship decision without spending time reading and composing a routine response.
When Wilkinson has already contributed to a thread and the remaining replies only say that a plan works, the first agent concludes he no longer needs to see them and archives the messages.
→ Wilkinson says this first filter reduces his email by about 20 percent.
Common mistakes
Treating every message as required reading
An inbox stays overwhelming if automation drafts replies but never removes thread noise that needs no human attention.
Mixing urgency with total volume
Without a separate short-horizon label, the user must scan the whole inbox to discover whether an emergency is hidden inside it.
Automating the decision itself
The workflow is strongest when AI proposes and composes while the human still chooses the underlying response intent.
Is it for you?
Best for
It is best for executives with repetitive email patterns, high message volume, and many simple scheduling or yes-no decisions.
Not ideal for
It is not ideal for high-stakes correspondence that the described workflow has not yet demonstrated it can classify and draft reliably.
From the episode
I’ve run 75+ businesses. Here’s why you’re probably chasing the wrong idea.
Andrew Wilkinson (co‑founder of Tiny)