The AI Email Workflow That Saves Me 5 Hours Every Week

I tracked my email time for two weeks in March — every minute, logged with a timer — because I suspected I was lying to myself about it. I was. The real number was 11.5 hours a week. Reading threads I was cc’d on, drafting replies I’d already mentally composed, deleting newsletters, negotiating meeting times. Knowledge-work classics.

Today that number sits between 6 and 7 hours. The difference — roughly five hours a week — comes from an AI workflow that runs mostly on my Mac, plus one non-AI discipline change that I’ll get to, because it matters more than any model. Headline first, math later. And the math will be itemized, because “saves me 5 hours” articles that never show their accounting are exactly the genre I’m trying not to write.

Triage: AI summaries of threads you were dragged into

The biggest single win is thread summarization. Not summarizing every email — summarizing the 23-message threads where you were cc’d on message 9 and someone “just wants your sign-off” on message 23.

My setup: an Apple Shortcut called Summarize Thread. In Apple Mail, I select a thread, hit a keyboard shortcut, and the Shortcut grabs the selected messages’ text and pipes it to a local model via Ollama’s API. The core of it is one “Run Shell Script” action:

curl -s http://localhost:11434/api/generate -d "$(jq -n \
  --arg prompt "Summarize this email thread in 5 bullets max:
  who wants what, what was decided, what is still open,
  and what — if anything — is being asked of Jakub.
  Thread: $THREAD_TEXT" \
  '{model: "qwen3:14b", prompt: $prompt, stream: false}')" \
  | jq -r '.response'

The result pops up as a notification-sized digest. Decision quality on “do I need to actually read this” is near-perfect, because the prompt forces the model to answer the only question that matters: what is being asked of me?

Why local? Because my inbox contains contracts, salary discussions, and other people’s confidential problems. None of that should transit a third-party API just so I can skim faster. A 14B model on Apple Silicon summarizes a long thread in under ten seconds, which is faster than I can read message 9 of 23.

Drafting: the personal style prompt

Generic AI email drafts are recognizable from orbit — “I hope this email finds you well” — and sending one is mildly insulting to the recipient. The fix is a style prompt: a standing instruction that teaches the model to write like you. Mine took twenty minutes to build and it’s the highest-leverage prompt I own. Here’s the reusable template:

You draft emails as me. My style rules:
- Greeting: first name only ("Hi Petra,"). Never "Dear" or "Hello".
- First sentence answers or addresses the main point. No warm-up,
  no "hope you're well".
- Short paragraphs, 1-3 sentences. One idea per paragraph.
- Direct but warm. I say "I can't do Thursday" not "Thursday may
  be challenging".
- I write numbers as digits. I use em dashes. I never use
  exclamation marks in work email.
- Sign-off: "Thanks," or "Best," then "Jakub". Nothing flowery.
- Length: shorter than the email I'm replying to, always.
Given an incoming email and my one-line intent, produce a ready-to-send
reply. Output only the email body.

The “one-line intent” part is the workflow: I read the email, type a fragment like “yes but only after the 15th, and ask for the figures first”, and get back a complete, send-able draft in my voice. Composing that fragment takes five seconds. Composing the polite three-paragraph version myself takes four minutes.

For privacy-sensitive mail this runs through the same Apple Mail + Shortcuts + Ollama pipe as triage. For low-stakes external mail I sometimes use Claude in the browser instead — cloud models are still noticeably better at delicate phrasing, and a public conference reply contains nothing worth protecting. That’s my routing rule in one sentence: sensitive content gets the local model, hard diplomacy gets the cloud, and most email is neither and gets whichever is closer.

The four reply templates that cover 80% of my outbox

Logging my sent mail revealed something embarrassing: four reply types cover the overwhelming majority of it. Each now has a saved intent template I fill in:

1. Yes-with-conditions. “Agree to X. Conditions: A, B. Confirm timeline.” The model wraps it in the social padding.
2. The graceful no. “Decline X. Reason: bandwidth. Leave door open for [quarter].” Hardest one to write by hand without sounding cold; the style prompt nails it.
3. The status nudge. “Chasing the thing from [date]. Friendly, assume they forgot, restate exactly what I need and the deadline.”
4. Meeting-time negotiation. “Can’t make proposed time. Offer Tue 10:00 or Wed 14:00 CET, 30 min, their call link or mine.”

Meeting requests deserve their own footnote: a Shortcut reads my next five days of calendar availability and injects free slots into template 4 automatically, so I never alt-tab to the calendar to compose an “I’m free at…” email again.

The honest 5-hour accounting

Two weeks measured before, four weeks measured after. Per week:

• Thread summarization: ~1.5 hours saved. Roughly 25 long threads a week at 4–6 minutes of reading each, replaced by 10-second digests plus selective deep-reads.
• Reply drafting: ~2 hours saved. About 40 substantive replies a week. Average hand-written reply: 4 minutes. Average intent-fragment-plus-review: 1 minute. The arithmetic is boring and that’s why I trust it.
• Newsletter and notification triage: ~1 hour saved. A daily digest Shortcut summarizes everything matching newsletter senders into one morning briefing; I unsubscribed from a third of them after seeing how rarely the digest contained anything I acted on.
• Meeting-request handling: ~0.5 hours saved. Maybe 8 scheduling exchanges a week, each compressed from 3–4 minutes to under one.

Total: ~5 hours. Not 20, not “10x my productivity.” Five, measured, with a timer. Your number will differ with your inbox shape — if you get 30 emails a week, this whole apparatus is overkill and you should close this tab.

The discipline part nobody mentions

Two unglamorous truths that determine whether any of this works.

First: every AI draft gets a 10-second human pass. Every one. I read the draft top to bottom before sending. About one in ten gets a tweak; about one in forty gets rewritten because the model misjudged the politics of the situation. That pass is non-negotiable — the day you start blind-sending is the day you sign off on a commitment you never read. The 10 seconds are already counted in the math above.

Second: batching beats models. I process email twice a day — 11:30 and 16:30, calendar-blocked — and the Mail dock badge is off. This single non-AI change is worth more than the entire AI stack, because the true cost of email is not reading time, it’s the context-switch tax of checking it 30 times a day. The AI workflow makes each batch session fast; the batching makes the sessions finite. Without the second part, you’ll just use your reclaimed five hours to check email more often, and the machine wins after all.

Steal the style prompt, build the summarize Shortcut on a rainy afternoon, and log your own two weeks of before-and-after. If your number comes out at three hours instead of five, you still bought back a Friday afternoon every two weeks — for software that runs quietly on the Mac you already own.