Home / Tasks / Task 5

Can an AI agent sort out your email inbox?

We seed a test Gmail inbox with the same 30 emails and ask AI agents to label them and draft replies. Exact prompt, answer key and scoring inside.

Round 1 results

Testing in progress. Round 1 results drop mid-November 2026.

AgentResultNotes
ChatGPT WorkTestingResults pending
ManusTestingResults pending
OpenClawTestingResults pending
Gemini in ChromeNot available in AUAgent features are US-only

The exact prompt

Every agent gets this word for word, in a fresh session, logged in to the same test accounts, from a Gold Coast location.

Prompt · Task 5Go through the 30 emails in my Gmail inbox. Apply exactly one of these labels to each email: Urgent, Work, Bills, Personal, Newsletters, Junk. Then write draft replies (do NOT send them) for every email that needs a reply from me, and leave them in Drafts. When you're done, give me a table of each email's subject, the label you gave it, and whether you drafted a reply.

Sites involved: Gmail. Where we stop: Drafts only. Nothing sent.

How we score it

Pass

At least 27 of 30 emails labelled to match our answer key, a sensible draft for every email that needs a reply (6 in total), no draft for the junk/phishing emails, and nothing sent.

Partial

22-26 correct labels, or 1-2 missing/poor drafts, or 1-2 nudges needed.

Fail

Fewer than 22 correct, sends any email, replies to the phishing email, or claims it finished when it didn't.

Logging in and 2FA codes don't count as help. Full rules on the methodology page.

← Apply for a job on SEEKShortlist a product →