Automation

AI Voice Agents

An AI voice agent can answer a call at midnight, take the details and book an appointment. It can also frustrate a customer badly if it is used for the wrong calls. This page is about which calls it handles well, which it must hand to a person, and what works in Nepali.

  • Handles the repeatable calls
  • Hands off cleanly
  • Honest about language support

Who this is for

Businesses losing enquiries because nobody answers the phone: clinics after hours, service businesses whose staff are on site, hotels at reception-light hours, and anyone whose missed-call rate is high enough to measure. It works best where a large share of calls are the same three or four questions. It is a poor fit where every call is different, where the caller is likely to be distressed, or where the conversation is the sale rather than a step towards it. This sits within marketing and business automation.

What a voice agent does well

  • Answering every call. No queue, no engaged tone, at 2am and during a festival.
  • Repetitive information. Opening hours, location, whether a service is offered, what documents to bring, price ranges.
  • Qualification. Establishing what the caller wants and whether it is something you handle, then routing accordingly.
  • Appointment booking, against a real calendar, with a confirmation sent afterwards: see appointment and reminder automation.
  • Capturing the details of an enquiry it cannot resolve, and creating a record with a transcript so the person who calls back has the context.

What it does badly, and where I would not use it

Being direct about this, because the failures are expensive.

  • Nepali language support is the practical constraint. Voice recognition and synthesis quality in Nepali is behind English, and mixed Nepali and English speech, which is how most people here actually talk, is harder still. I test with recordings of your own callers before recommending it, and for many Nepali-facing businesses the honest answer today is that a WhatsApp automation serves better than a voice agent.
  • Distressed or urgent callers need a person immediately. Health emergencies, complaints and anything involving money going wrong should route straight to a human.
  • Complex or negotiated conversations are not a fit. If the call is where the deal is made, the agent should only be qualifying and handing over.
  • Accents and poor lines. Real calls come over patchy mobile connections, and accuracy drops. The agent needs a graceful failure route rather than repeating a question three times.

The rule I apply: the agent handles the calls that would otherwise go unanswered, and a person handles everything the agent cannot resolve in two exchanges.

What the work includes

  • Call analysis first. Recordings and volumes reviewed to establish what proportion of calls are actually repeatable, which decides whether this is worth building.
  • Conversation design. The questions, the branches, and the explicit handoff triggers.
  • Platform selection and configuration, with a phone number, and the honest cost per minute set out before you commit.
  • Calendar and CRM integration so a booking is real and an enquiry is recorded.
  • Handoff paths. Transfer to a person during hours, a callback request outside them, and an immediate escalation route for urgent calls.
  • Testing with real recordings and real callers, including deliberately awkward ones, before it takes a live call.
  • Monitoring. Transcripts reviewed weekly at first, so the failures are found by you rather than by a customer.

How a build runs

  1. Assessment of call volumes, patterns and language mix, with an honest recommendation which may be not to build it.
  2. Design of the conversation and the escalation rules.
  3. Build and integration with the calendar, CRM and phone system.
  4. Testing including edge cases, poor lines and mixed-language callers.
  5. Soft launch on after-hours calls only, where the alternative is nobody answering.
  6. Review and widen or withdraw, based on transcripts rather than on optimism.

What you receive

  • A call analysis with the share of calls suitable for automation.
  • A configured voice agent with documented conversation flows.
  • Calendar and CRM integration, with transcripts stored against each enquiry.
  • Escalation and handoff paths.
  • A monitoring routine and the per-minute running cost.

Build pricing is on request, and the platform’s per-minute cost is paid by you directly. See packages.

Next step

Tell me how many calls you miss in a week and what most of them ask on the contact page. If a voice agent is not the right answer, WhatsApp automation or AI chatbots often are.

Frequently asked questions

Will it work in Nepali?

Test before you trust it. Nepali voice quality lags English, and mixed Nepali and English speech is harder still. I test against recordings of your own callers first, and for many Nepali-facing businesses WhatsApp automation is the better answer today.

Will customers know it is not a person?

Yes, quickly, and the agent should say so rather than pretend. Callers accept an AI that is clearly an AI and handles their request; they resent one that pretended and then failed.

What happens when it cannot help?

It transfers to a person during hours or takes a callback request outside them, and always creates a record with a transcript. An agent without a clean handoff is worse than no agent.

What does it cost to run?

The platforms charge per minute, so the running cost scales with call volume and is paid by you directly. I set out the expected monthly figure from your call volumes before anything is built.

Ready to talk about your project?

A free 30-minute call, a straight answer about what would move the numbers, and a written proposal within 48 hours if we are a fit.