How AI Voice Chatbots Cut Support Costs for SG Web

—

What SG Web Actually Pays Per Call

SG Web runs a 12-person agency in a Tanjong Pagar co-working space, managing roughly 140 active client sites. Before voice automation, they ran a 3-person support desk answering an average of 610 inbound calls monthly. Each agent earned SGD 3,400 monthly, plus 9% CPF. That totals SGD 10,200 in base salaries, or SGD 16.72 per call before factoring lost billable hours.

The voice bot runs on Vapi at USD 0.05 per minute for the base LLM plus Twilio telephony at USD 0.02 per minute. An average support call — Lumen AI or Vapi based — lasts 2.2 minutes. That’s roughly USD 0.154 per call, or SGD 0.21. The cost ratio is 80:1. SG Web moved from 610 human-handled calls monthly to 76 human-handled calls, with the bot absorbing the difference. Support payroll was resized to one full-time agent and one 4-hour daily part-timer.

Building the Intents Around L2 Support Tickets

SG Web analyzed six months of freshdesk tickets and call logs. They found 47% of calls were about the same four topics: forgot login, billing receipt attached to shopper (their WooCommerce care plans), SSL certificate expiry, and “my site is blank.” These are deterministic issues. A voice bot doesn’t need AGI to resolve a DNS propagation call — it needs a decision tree that handles the Singapore context.

The bot, called “Aunty Sue”, greets with “Hi, you have reached SG Web support. How can I help you?” It transcribes via Deepgram, passes the text to GPT-4o-mini with a system prompt that enforces three actions: resolve, log, or route. For SSL renewal questions, Aunty Sue asks for the domain, checks the client’s active SSL cert via WHM API, and if it is valid, says, “Your certificate is valid until 14 March. If your browser shows an error, try clearing your local DNS cache first.” That resolves 68% of SSL-related calls without a human touch.

Singlish and Short-Hand Vocabulary Handling

Generic American-accented bots fail in Singapore because clients speak a mixture of Mandarin, Hokkien, and English with specific quirks. SG Web seeded the Vapi tool with a lexicon of regional terminology:

– “Spot check” = checking the official website

– “The link spoilt” = broken URL or redirect error

– “My website kena hacked” = security compromise

– “Bill how much ah?” = invoice amount inquiry

The bot is prompted to respond in kind. If a client says, “Website down lah, can fix or not?”, Aunty Sue replies, “Can, can. Let me check now.” During initial testing, the bot failed 3% of Singlish queries, so SG Web added Vapi’s custom TTS voice — “Lilian” — a light Singaporean accent that significantly reduced repeat confirmations. The repeat-call rate dropped from 11% to 2.8% after tuning, because clients trusted the voice.

Hard Escalations: Sentiment-Based Handover

SG Web has a policy: if the client uses any of a set of trigger phrases — “I want to cancel”, “this is the third time”, “I’m reporting you to CASE” — the bot instantly transfers to a human agent with full transcript attached. This prevents burning goodwill for the sake of cost metrics.

The actual intercept rate on these triggers is 91%. The bot’s sentiment score, derived from transcription tone, identifies hesitation. If the client says, “I see… well… okay, but I need my site live by tomorrow,” the bot logs the uncertainty and shifts to a soft handover instead of ending the call.

This triage reduced escalation queue time from 8 minutes average to 30 seconds. The agent receives the transcript, the sentiment score, and a suggested action: “Client mentioned timeline: 2025-11-14. Escalate to dev team, KIV.”

Real Cost Numbers and Monthly P&L Impact

Here is the granular breakdown SG Web published in their internal September review:

– August 2025 (pre-bot): 612 calls, 612 human-touched, 237 support tickets logged. Agent costs: SGD 10,200. Estimated resolution cost (including waste time between calls): SGD 14,500.

– December 2025 (post-bot): 598 calls, 84 human-touched, 123 tickets logged. Bot compute + telephony: SGD 640 (including Studio credits and prompt iterations). Human agent cost (FTE + PT): SGD 5,400. Total resolved cost: SGD 6,040.

That is a 58% reduction in operating cost, not just per-call cost. SG Web also reduced ticket backlog by 45%, because the bot creates structured Freshdesk tickets that the remaining agent can process in batch instead of transcribing messy notes. The revenue side remains unaffected because the bot resolves what was previously unpaid support included in maintenance retainers, and two clients upgraded to their higher-priority care plan after positive voice-bot interactions.

SG Web is now testing the “Aunty Sue” workflow on WhatsApp voice notes, which is where local businesses increasingly send support requests. The target is to push the bot’s intent coverage to 85% of all inbound calls without increasing the escalation rate beyond 4%.