Companies That Automated Chat but Left the Phone Untouched
Over the past few years, many companies finished automating their text channels with web chatbots and messenger-based support. Yet more than half of all inquiries still arrive by phone. Call data from mid-sized Korean distribution and manufacturing firms shows that 50-70% of all customer contact is voice, and unless an agent manually takes notes, none of what is said in those calls ever reaches a system.
This is the cost of channel fragmentation. Chatbots leave logs that can be analyzed; phone calls remain a black box. A customer changes their delivery address over the phone, it never reaches the ERP, the shipment goes to the wrong place—and when that same customer opens the chatbot the next day, they have to explain everything from scratch.
The surge in production voice agent deployments through 2026 rests on a technical threshold being crossed. Streaming STT combined with low-latency TTS pushed response delay below one second, and past that point a conversation finally becomes tolerable.
What a Voice Agent Is Made Of
A voice agent has four layers.
What determines quality is not accuracy but latency. People cannot tolerate more than a second of silence on a call. STT at 300ms, LLM first token at 400ms, and TTS at 200ms already total 900ms, so each stage needs a budget measured in milliseconds. Add a tool call and your internal API response time lands directly on top of that.
Voice also brings problems that text never had: handling barge-in when a caller interrupts mid-response, detecting whether a turn has ended or the speaker is merely pausing, and coping with noise from stores and moving vehicles.
Korean adds its own burden. Transcription error rates for phone numbers, addresses, and alphanumeric part codes run far higher than for ordinary sentences. In practice you never use raw recognition output—you add a post-processing layer that cross-checks results against address databases and part master tables. Consistent formal register also has to be pinned down explicitly in the prompt.
Where to Start So You Don't Fail
Half the outcome is decided by which calls you choose.
Good candidates are calls where the answer is a structured data lookup: delivery status, stock checks, booking creation and changes, business hours. The questions are bounded and the answers live in a database.
Poor candidates are calls involving emotion and negotiation. Complaint handling, cancellation retention, and price negotiation only amplify frustration the moment you automate them.
Counterintuitively, ROI usually appears first on outbound, not inbound. Appointment confirmation calls, payment reminders, and delivery scheduling have a clear purpose, simple scripts, and low downside when they fail. Automating confirmation calls that agents used to make 200 times a day produces immediately measurable results.
The Real Difficulty Is Integration
Voice quality is largely solved by commercial engines. Integration with your internal systems is not—you have to design it.
Deployment Checklist and Cost Structure
Most voice platforms bill per minute, with STT, LLM, TTS, and telephony charges stacked on top. The per-minute rate looks cheap until you account for a three-minute average call and a 30% transfer rate. Calculate cost per call = (per-minute rate × average call duration) + agent cost on transfer to see actual savings.
Keep success metrics to three: containment rate, the share of calls resolved without human involvement; average handling time; and satisfaction on transferred calls. Chase containment alone and you will trap callers in the bot while satisfaction collapses.
POLYGLOTSOFT builds voice agent deployments in stages through a subscription model. We start with low-risk scenarios such as outbound confirmation calls, measure real cost per call and containment rate, then extend into ERP, order, and inventory integration along with human escalation flows. If you are considering turning your phone channel into data, we would be glad to talk.