ServiceNow's free voice AI can take a customer's request and act on it

The short version

• ServiceNow has released a free voice AI model that listens to a caller and uses company systems to act on the request.
• In standard tests it completed 65.51% of customer service tasks on average, from 81.58% in retail down to 45.61% in telecom.
• It is built mainly for English, and ServiceNow reports no results in Indonesian. Try it, but don't count on it for your call center yet.

What ServiceNow released

ServiceNow, the workplace software company, published a preview of a model called Apriel Voice on October 29. Any company can download it. Its license generally allows commercial use.

Most voice assistants work like a relay. One system turns speech into text, a second decides what to say, and a third reads the reply aloud. Apriel Voice joins the first 2 steps. The same model hears the person directly. It also decides when to use a tool, such as opening a support ticket. A separate voice system still speaks the answer, so a company can swap voices without retraining the model.

ServiceNow's paper opens with an example from everyday office life. An employee with a locked account explains the problem aloud and shares a screenshot. They then ask the assistant to open a ticket. The model is meant to follow all of that as one conversation.

How well it works

The main test covers customer service tasks in 3 industries. Apriel Voice's scores show what you might get today:

• Retail: 81.58% of tasks completed.
• Airline: 69.33% of tasks completed.
• Telecom: 45.61% of tasks completed.

On average, the first spoken response took 1.91 seconds. For comparison, Google's best live voice system completed 84.19% of tasks with a 1.50-second delay. OpenAI's reached 69.82% at 1.55 seconds. Among free models that also hear speech directly, the next best managed 26.80% at 3.95 seconds. A free relay system matched Apriel Voice on tasks, at 67.78%, but took 3.01 seconds to respond.

A second test is less flattering. It checks whether the whole conversation stays accurate and faithful to the facts of the case. Apriel Voice passed 51.27% of the time, against 54.82% for Google's system. It scored 74.13% on how the conversation felt, including how it took turns and kept replies short. So it handles the conversation better than it gets the facts right. On long, multi-step office tasks with the right tools already supplied, it completed just 23.0%.

What it means for your company

Customer service is an obvious place for Indonesian companies to try this. A free model you can run on your own servers has 2 appeals. Your call recordings don't have to go to an outside provider, and you avoid paying per minute for someone else's system.

You would still pay for the computers and the people to look after them. Language is also a serious gap. ServiceNow says the model prioritizes English and that its other languages are uneven. English results don't predict quality elsewhere, in the company's own words. The paper's translation tests cover European languages. It reports nothing for Bahasa Indonesia or for callers who mix it with English or local languages.

The telecom score is also worth a second look. Results vary widely by industry, so a good average may hide a poor result in your own line of business. These tests were also scripted scenarios. Real callers interrupt and call from noisy streets.

What to do next

• Try it first on internal English-language requests, such as IT help for staff. A mistake there costs less than one with a paying customer.
• Build your own test set from real calls in the languages your customers use, with proper consent.
• Measure completed tasks, and check every action the AI took against what the customer needed. A pleasant conversation can hide a wrong result.
• Keep a quick route to a human agent, especially for payments and complaints.
• Compare the full cost, including servers and staff, against paid services. Free to download doesn't mean free to run.