AI Customer Inquiry Triage with Claude: A Practical Recipe for Singapore SMEs
Build a Claude inbox triage that classifies, drafts replies, and escalates urgent issues — set up in an afternoon. Singapore SME walkthrough with prompts and code.
Most SG SMEs have one shared inbox: hello@ or info@. Sales inquiries, support tickets, customer complaints, and spam all land in the same place. Someone — usually the founder or a senior ops person — reads everything before routing. That's 30–90 minutes of senior time per day that creates zero customer value.
Triage is exactly the kind of repetitive classification task Claude is good at. Build it once, save the time forever.
What the system does
For every incoming message:
1. Classify — sales / support / complaint / spam / other.
2. Tag urgency — low / medium / high.
3. Detect context — is this a returning customer, what's the topic.
4. Draft a reply — 60–120 words, ready to approve and send.
5. Route — sales goes to the salesperson, support to ops, complaints to the founder, spam to spam folder.
6. Log — every decision logged for the first month so you can verify accuracy.
The first version is one prompt. The polished version is a small Claude Code app. Either is shippable in an afternoon.
The classification prompt
*"You are an inbox triage assistant for [COMPANY NAME], a [INDUSTRY] business in Singapore.
Customer inquiry follows. Return JSON:
- category: one of [sales, support, complaint, spam, other]
- urgency: low / medium / high
- topic: 2–4 word summary
- requires_human: true if sensitive (legal threat, payment dispute, refund request, complaint about staff) else false
- draft_reply: 60–120 word polite professional reply, Singapore English, no clichés
If category is spam, set draft_reply to null.
Inquiry:
[PASTE EMAIL OR MESSAGE]"*
Test on 30 real inquiries from your inbox before going live. Tweak the category list and urgency rules to match how your team actually triages.
Wiring it up
Two paths.
Path 1: manual paste, fastest. Open Claude.ai. For each inquiry, paste the prompt + the message. Copy the JSON output. Approve or edit the draft reply. Send. Time per inquiry: 30 seconds vs. 3–5 minutes manual.
Path 2: automated, takes a day. Use Claude Code to build a small app:
1. Webhook receiver for inbound emails (Gmail forwarding rule, or polling Gmail API).
2. Each message goes through the prompt.
3. Output written to a Slack channel for human approval.
4. One-click "send" routes the reply via Resend.
5. Log everything to SQLite or Postgres.
Starter prompt for Claude Code: *"Build the system above. Use Vite + React + Express. Webhook receives Gmail forwards. Claude Sonnet for classification. Slack approval flow. Resend for sending. Log to SQLite. Deploy to Vercel + a small Render database. PDPA-aligned (Singapore-region hosting)."*
Tuning for your business
Three things to tune over the first month:
Categories. Your SME might need different categories (booking / quote / urgent / other for an F&B business; mandate-pitch / candidate / current-engagement / other for a recruiter). Adjust.
Urgency rules. Default urgency rules are generic. Replace with specifics: *"Mark urgent if (a) message contains 'refund', 'lawyer', 'cancel', 'angry'; (b) customer has spent over S$5,000 with us in the past year; (c) message arrived outside business hours from a known VIP."*
Reply tone. Train the prompt to match your brand voice. Add: *"Reply tone: warm, brief, no exclamation marks, Singapore English."* Or: *"Reply tone: enthusiastic, signs off with thanks-and-name."*
What this looks like in practice
An OTG client (mid-size SG-based service business, ~150 inquiries/week) deployed this last quarter. Results after 6 weeks:
- 96% classification accuracy
- 88% of draft replies sent without edit
- ~10 hours/week saved across the team
- Zero customer complaints about the AI-drafted replies (we asked)
The remaining 12% of replies that get edited are the cases where the customer's situation has nuance the prompt didn't see. That's expected — the system handles the routine, the team handles the unusual.
For more SME workflows like this, see Claude for Singapore SMEs: 10 Prompts. For the full Singapore-context prompt rules, see Singapore-Context Prompting.
Next steps
If you want to ship this yourself, start with Claude Code in Singapore. If you'd rather have OTG build it for you, the Functional App Sprint (from S$3,500, one week) includes deployment + handover.
Book a free 30-minute call — we'll diagnose your specific inbox patterns and recommend the right tier.
Frequently asked questions
Will this work with my existing email setup?
Yes — most providers (Gmail, Outlook, Zoho) support webhook forwarding or polling. WhatsApp Business API also supports webhooks. The triage logic is independent of the source.
How accurate is Claude at classifying inquiries?
On typical SG SME inboxes, we see 92–96% accuracy on the four-class classification (sales / support / complaint / spam). Urgency detection is harder (~85%). Always include a human-review step for the first 2 weeks while you tune the prompt.
Can it draft replies, or only classify?
Both. The prompt below classifies AND drafts a 60–120 word reply per inquiry. You approve before sending. We don't recommend full auto-send until you have a month of accuracy data.
What does it cost to run?
Claude API is approximately S$0.005–S$0.02 per inquiry depending on length. For 1,000 inquiries/month: S$5–S$20 in API spend. Versus 1,500 minutes of staff time at SG rates: easy win.
Is this PDPA-safe?
Yes when you use enterprise-tier Claude (no training on input). Customer email content is personal data; treat the system as a data processor and document accordingly. See [PDPA Prompting Checklist](/resources/pdpa-prompting-checklist).
