Case Study: Solo Founder Cuts Email Time 70% with an AI Assistant

Background

The founder of ShipKit — a 2-person bootstrapped analytics tool — spent the first 90 minutes of every day in Gmail. Investor updates, customer bugs, partnership intros, newsletters disguised as urgents. Real work started at 10:30.

He'd tried Superhuman, SaneBox, and inbox-zero systems. The bottleneck wasn't reading speed — it was deciding what needed a reply and writing the first sentence.

The problem

Hiring wasn't an option at $8k MRR. He needed automation that made him faster, not a bot talking to customers.

Architecture

Built from the AI Email Assistant workflow:

Cron (every 5 min, VPS)
  → Pull new unread from Gmail
  → Classify: urgent | reply | fyi | archive
  → FYI/archive → label + skip
  → Urgent → flag in daily digest
  → Reply → draft in founder's voice (Gmail draft API)
  → 8am digest email: urgents + stats

Hard rule: no gmail.send scope. Every outbound message required a human click.

Implementation

Week 1 — Classification only

Structured JSON output via Claude Haiku:

{
  "bucket": "urgent | reply | fyi | archive",
  "reason": "one line",
  "confidence": 0.0–1.0
}

He labeled 80 historical emails manually. Holdout accuracy: 91%. FYI/archive auto-labeling alone reclaimed ~20 minutes/day.

Week 2 — Voice-matched drafts

He fed 15 sent emails into a style guide prompt: short sentences, no exclamation marks, signs off with first name only. Draft acceptance started at 34% and climbed to 58% over three weeks of edits.

Rejected drafts still helped — they surfaced the right thread context even when the wording was wrong.

Week 3 — Cost routing

Classification moved to Haiku ($0.002/email). Drafting stayed on Sonnet ($0.03/draft). LiteLLM proxy handled both with one config file. Monthly cost: ~$12.

Results

The founder's quote: "I didn't get an AI assistant. I got a morning back."

What he'd do differently

1. Label 80 emails before writing code — his first classifier used too many buckets and confused "reply" with "urgent."
2. Build the digest first — even without drafts, the 8am urgent summary was valuable on day three.
3. Keep send manual forever — one bad auto-reply isn't worth any time saved.

Takeaway

Email automation for individuals isn't about AI sending mail — it's about AI doing the sorting and first draft so you can batch the human part into 25 minutes. The workflow guide is the blueprint; voice examples are the secret ingredient.