AI chatbot vs voice agent for New Zealand businesses.

How do AI chatbots and voice agents compare?

Both answer customers, look things up and hand over to staff. They differ in the channel, how they are billed, and what goes wrong.

Where is each one stronger?

A chatbot is easier to get right. Customers can paste an order number, staff can read every conversation in seconds, and a wrong answer is visible in writing for correction. It also builds on what you already have: our lesson on building customer FAQs from support tickets produces most of a chatbot’s knowledge.

A voice agent reaches customers who will not type. For a plumber, electrician or restaurant, a missed call after hours is often a lost job or booking. A voice agent can answer every call, take the details and book the slot. See our AI voice agents page for what one can and cannot handle.

How do you test a chatbot and a voice agent on the same task?

Script 30 conversations for each task and run them through both. Use real staff, with a mix of accents, and make a few of the voice calls from a noisy car or site.

1. After-hours booking. The customer gives a name, a street address with a unit number such as 14A, a suburb such as Paraparaumu, Whakatāne or Ōtaki, a job description and a preferred time. Count how often the suburb, street number and phone number are captured exactly.
2. Order or job status. The customer gives an order or job number and asks when it will arrive. Check that the agent looks up the right record and refuses politely when the number does not match.
3. Complaint and handover. An unhappy customer wants a person. Time how long the agent takes to hand over, and judge the summary the staff member receives.

Our engineering guide to voice agent quality evaluation covers latency, interruption and transcription tests in more depth.

What does each cost to run?

Worked example: 1,000 customer conversations a month. Assumptions: voice calls average three minutes; prices are vendor list prices on 19 October 2026 in US dollars, with Twilio’s NZ page showing “$” without a currency label.

Voice agent on Vapi with a Twilio NZ local number:

• Vapi’s own calculator estimate is US$82 to US$129 per 1,000 minutes, so 3,000 minutes is about US$246 to US$387.
• Twilio’s NZ inbound local rate is $0.0100 a minute: 3,000 × 0.0100 = $30, plus $3.15 for the number.
• Total: about US$279 to US$420 a month, or US$0.28 to US$0.42 a call.
• If callers ring an 0800 number instead, Twilio’s toll-free inbound rate of $0.2556 a minute adds 3,000 × 0.2556 = $766.80 a month, plus $40 for the number. Check your carrier’s rates before choosing a toll-free line.

Chatbot:

• Packaged, per outcome: if 600 of the 1,000 conversations are resolved by Intercom’s Fin at US$0.99 each, that is US$594 a month. Intercom counts an assumed resolution, where the customer leaves without asking for more help, as billable.
• Built on model APIs: assume eight turns per conversation at 3,000 input and 200 output tokens each on Claude Sonnet 5.5 (US$2 and US$10 per million). Each turn is US$0.008, each conversation US$0.064, so US$64 a month in model fees, plus build, hosting and maintenance.

So per conversation, an outcome-priced chatbot can cost as much as a voice agent, while a chatbot you build costs cents to run. Voice costs more to build and test well. Our AI chatbot development page explains what drives build cost.

Which should you choose?

Choose a chatbot if

• Most enquiries already arrive by web chat, email or messaging.
• Answers depend on order numbers, links or documents that are easier to share in writing.
• You want every conversation easy to review in the first months.
• Your support team works in a helpdesk such as Intercom or Zendesk already.

Choose a voice agent if

• Customers phone you, and missed calls after hours or at peak times lose work.
• The task is short and structured: take details, check availability, book.
• You can supply and test a +64 number and a reliable handover to a person.
• You have tested your callers’ accents and place names and the agent handles them.

Where do both fall short?

• Confident wrong answers. Both will state a price or availability that is wrong unless they look it up. Restrict them to verified sources, and see managing AI mistakes and escalation.
• Emergencies. Twilio does not allow calls to emergency services over its NZ numbers. A voice agent must tell anyone reporting an emergency to hang up and dial 111; a chatbot must say the same.
• Privacy notices. Under the Privacy Act 2020, tell people an automated assistant is answering and whether conversations are recorded or transcribed, and keep transcripts only as long as needed (IPP 9). If you use a caller’s voice to identify or verify them, the Biometric Processing Privacy Code 2025 may apply.
• No way out. An agent that cannot reach a person within one or two turns frustrates customers. Design the handover first, then the automation.

If you are weighing both for a support desk, our AI customer support work starts from your ticket and call history.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.