Monday, 7:40 a.m., Tampa Pasta House. Margaret Ellison has twenty minutes before the kitchen needs her, and twelve new e-mails.
The first one shouts in capitals: "URGENT!!! LAST CHANCE," an advertising promo that "ends TONIGHT." The fourth one barely says anything: "Tuesday delivery — one small change." Inside, the tomatoes for the whole week won't arrive Tuesday. The truck is now Thursday.
If you hand that inbox to the AI assistant on your phone, which one does it read first?
Tampa Pasta House is our sample restaurant. The twelve e-mails are made up, written the way a real Monday inbox looks.
📬 What we actually tested
On October 15, 2026, we gave Gemini 3.8 Flash the twelve e-mails one at a time, with the same instruction every time: say what the sender wants in one line, put the e-mail in a queue (Today, This week, Can wait, Ignore), name the deadline if the e-mail has one, quote the one sentence the owner must read, and write the one question the owner has to answer. No replies to senders. No facts that aren't in the e-mail. Then one more request: a summary of the whole morning. Thirteen answers came back.
The loud e-mails didn't win. "URGENT!!! LAST CHANCE" went to Ignore. So did a fake "FINAL NOTICE" to renew the restaurant's web domain for $249 from a company that isn't the registrar, and a "featured listing" with a one-time fee of $890. We didn't tell the model the domain notice was fake. It worked that out on its own. The tomato e-mail went to Today. In 9 of 12 e-mails the queue matched ours. In 12 of 12 the quote was a real sentence from the e-mail, and nothing was made up: no amounts, dates or facts that weren't there.
Two mistakes worth knowing about:
- The water bill went into the wrong pile. "Your autopay for 09/2026 ($312.40) was declined by the bank. Please update the payment method within 5 business days to avoid a late fee." The model read the deadline correctly, "within 5 business days," and still put it in This week, not Today. It saw the clock and didn't raise the priority. The morning summary repeated the same mistake.
- Deadlines it had to work out, it didn't. When the date was written in the e-mail, the model named it every time: 5 of 5. When the deadline had to be worked out (the tomatoes must be settled before Tuesday, the birthday dinner for 14 needs an answer before Saturday, the shift swap before Friday), it wrote "None," or mentioned the day only as the date of the event. Zero of three in full.
It also leaned the other way twice: a 2-star Google review and a catering request for 40 people went into Today, where we had them in This week. That's the safer kind of mistake, but it still crowds the Today pile.
The loud e-mails lost. The quiet one with a fee attached slipped. Sorting by volume is solved; sorting by consequence isn't yet.
🗂 One card instead of a pretty reply
What made this useful wasn't a drafted reply. It was a five-line card for each e-mail. Here is the card Gemini 3.8 Flash actually returned for the tomato e-mail:
Ask: The sender asks to confirm if delivering the 12 cases of San Marzano tomatoes on Thursday instead of Tuesday is acceptable.
Queue: Today
Deadline: None
Quote: "Hi Margaret, the San Marzano tomatoes (12 cases) won't make Tuesday; the truck is now Thursday."
Question for the owner: Does Thursday work for the delivery of the tomatoes?
And the card we wanted: the same, with Deadline: before Tuesday's service, and a question that also asks whether Tuesday needs a backup supplier. The model got the queue, the quote and the question right. The date that matters most, the one nobody wrote down, is still the owner's job.
🛠 How to repeat without code
- Pick five to ten real e-mails from one morning. Remove names and amounts you don't want to paste anywhere.
- Write your queue names first: Today, This week, Can wait, Ignore. Use the same instruction for every e-mail; ours is above, in "What we actually tested."
- Paste one e-mail per request, not the whole inbox. Put today's date at the top of the request, so "Tuesday" means a real day.
- If your tool lets you set them (Google AI Studio does), use a low temperature such as 0.2, a short answer limit and no extended reasoning. We ran our test through a command-line tool that doesn't expose those settings, so ours ran on its defaults.
- Compare each card with your own judgment and mark only two things: the wrong pile and the missing deadline. Those are the mistakes that cost money.
🧾 The bill and the limits
Thirteen requests in about two and a half minutes. The tool doesn't print token counts, so this is an estimate from text length: about 3,600 tokens in and 1,200 out. At the paid API list price of Gemini 3.8 Flash, that's less than one cent for the whole morning.
The limits are real. Twelve made-up e-mails, one morning, one model, one run. We didn't test replies to senders, attachments or long threads. And a first attempt failed for a reason that had nothing to do with the model: the tool tried to run commands on our computer instead of answering, and 8 of 12 requests came back empty until we told it, in plain words, to answer in text only.
✅ Before you trust the assistant with an e-mail
- ☐ Give the assistant your queue names first; "urgent" in a subject line is the sender's word, not yours.
- ☐ Make it quote one sentence from the e-mail, so you can check every card against the original in a second.
- ☐ Ask for the deadline as a date, and work out the unwritten ones yourself: deliveries, bookings, shifts.
- ☐ Anything about money, inspections or staff ends with a question for you, never with an action.
- ☐ Show it two scam shapes from your own inbox: a fake domain renewal and a paid "featured listing."
- ☐ Read the cards once a day at the same hour; the assistant sorts, you decide.
Next in "AI on staff": Mateo from Casa Lista and ten Instagram messages that arrive at 10:40 p.m.
What was the last "urgent" e-mail that turned out to be noise?
