Let’s lock in an inbox that sorts itself — quick wins today, smarter automation by the weekend. We’ll blend simple rules (fast) with a focused AI classifier (smart) and add guardrails so nothing critical goes missing.
Where AI fits: rules catch the obvious; AI handles the fuzzy messages that waste your attention. Start in “suggestion” mode for 2–3 days, measure, then let high-confidence items move automatically. Expect rapid improvement as you correct edge cases.
What you need
- Email admin access (Gmail or Outlook).
- 6–8 labels with clear intent (Action-High, Action-Low, Finance, Clients, Newsletters, Internal, Optional: Uncertain).
- 20–50 example emails per label (kept in folders for testing).
- Optional: Zapier, Make, or Power Automate; an AI API key if using a third-party model.
- A simple privacy plan (see notes below).
Build the foundation (15–30 minutes)
- Define labels + response rulesKeep labels purpose-driven. Add a rule per label (e.g., Action-High: reply within 4 hours; Finance: review daily).
- Create native filters (fast wins)Gmail: Settings > See all settings > Filters and Blocked Addresses > Create a new filter. Use From (domains), Subject (keywords), and Has the words (invoice, unsubscribe, receipt). Apply label + auto-archive if appropriate (e.g., Newsletters).Outlook: Rules > Manage Rules & Alerts > New Rule. Use sender/domain, subject phrases, and add “assign to category” and “move to folder.”
- Collect examplesCreate temporary folders like TRAIN-Finance, TRAIN-Clients. Drag 20–50 real emails into each. These power your AI prompt testing.
Smart layer: AI classifier flow (Zapier/Make/Power Automate)
- Trigger: New email (exclude anything your native rules already filed).
- Sanitize (privacy optional but wise): replace email addresses, names, and numbers with tokens before sending to AI (e.g., [EMAIL], [NAME], [AMOUNT]).
- AI step: send the email text + the prompt below. Get back JSON with label and confidence.
- Decision: if confidence >= 80 and label ≠ Uncertain, apply label and move; else add label Uncertain or prefix subject with [SUGGESTED: Finance].
- Log: save predictions and your corrections to a simple spreadsheet for weekly tuning.
Insider template: Routing policy + robust AI prompt (copy-paste and customize)
“You are my email routing assistant. Use this policy and return ONLY JSON.
Labels and purpose:• Action-High: time-sensitive requests, commitments, clients with deadlines.• Action-Low: non-urgent tasks, FYIs that require action later.• Finance: invoices, receipts, payments, contracts with amounts/dates.• Clients: client comms that are not primarily finance or urgent action.• Newsletters: marketing/news updates, bulk mail, promotions.• Internal: team, HR, ops from our domains.• Uncertain: use only when confidence < 75 or content conflicts.
Guardrails:• VIP senders always Action-High: [ADD VIP EMAILS OR DOMAINS].• If email contains invoice/receipt/payment with amounts or PO numbers → Finance unless it’s clearly a newsletter.• If multiple labels apply, choose the single most urgent/purposeful label (prefer Action-High > Finance > Clients > Internal > Action-Low > Newsletters).• If unsubscribe/footer-heavy and no request → Newsletters.• If the message is a calendar invite, treat as Action-Low unless urgent language indicates Action-High.
Output ONLY this JSON schema:{“label”:”one label from the list”,”confidence”:0-100,”reason”:”short phrase”,”summary”:”one-line”,”escalate”:true|false}
Set escalate=true if label is Action-High or Uncertain. Do not include any extra text.
Email to classify: [PASTE EMAIL TEXT HERE]”
Pro tip: include 3–5 short examples in the prompt (few-shot learning) once you have them. Accuracy usually bumps meaningfully without code changes.
Example (what good looks like)
Input email: “Hi — please approve the attached invoice for $3,200 for Client X by Friday.”
Expected JSON:{“label”:”Finance”,”confidence”:92,”reason”:”Invoice approval with amount and deadline”,”summary”:”Approval request for $3,200 invoice by Friday”,”escalate”:true}
Rollout and expectations
- Day 1–2: Rules will catch 40–60% immediately.
- Day 3–5: AI in suggestion mode will route most of the fuzzy 30–40% at 70–85% confidence.
- Week 2: With corrections and 5–10 examples added to the prompt, high-confidence auto-move often reaches 80–90% precision for your core labels. Your results will vary; keep a light review cadence.
Mistakes to avoid (and the fix)
- Label sprawl: more labels = lower accuracy. Fix: stay under 8 and merge low-volume categories.
- Sender-only rules: forwards and aliases break them. Fix: add content keywords and negative keywords (e.g., exclude “unsubscribe” for Clients).
- No safety net: everything moves even when uncertain. Fix: use Uncertain + escalate=true for low confidence.
- One-shot prompts: never updated. Fix: add 3–5 real examples and one hard edge case per label after week 1.
- Privacy blind spots: sharing PII. Fix: tokenize sensitive data or use provider-native AI when required.
90-minute sprint
- List your 6–8 labels and write one sentence of purpose for each.
- Build 6–10 native filters for the obvious senders/keywords; auto-file Newsletters.
- Create TRAIN- folders; drop 20 emails into each.
- Set up the AI flow in suggestion mode; apply [SUGGESTED: Label] prefix.
- Review 25 suggestions; correct and note patterns to add to guardrails.
Week 2 tune-up
- Raise auto-move threshold to 80–85% for low-risk labels (Newsletters, Internal).
- Add VIP overrides and 5 few-shot examples to the prompt.
- Track: % auto-labeled, false positives, manual corrections, time-to-action on Action-High.
Closing thought: keep this lightweight. Rules do the heavy lifting, AI cleans up the messy middle, and your weekly tweaks make it feel like magic — without trusting your inbox to chance.
