Voice AI: The Dos, the Don’ts, and What’s Next

Voice AI has now evolved to the point that it provides workforce stability. Not the science fiction kind, but the kind that answers customer calls at 2 a.m. on a Saturday, listens carefully, assesses urgency, and either resolves the issue or escalates it immediately to the right person.

This new reality is due to recent developments in neural networks and large language models. Word-error rate has now fallen below 5 percent, while natural language understanding and generation are now commonly used as a writing partner (e.g. ChatGPT) across many professions.

Source: Artificial Analysis

However, AI is not magic. While AI can reliably replicate something already working, it’s not going to pull bunnies out of a hat. AI is great for routine jobs like scheduling appointments (tasks you can write a manual for), but it’s terrible at making an emotionally sensitive argument to close that new equipment install.

What’s Easy

Here are some great use cases for voice AI:

Routine scheduling. This is what voice AI does best—things that are repetitive and boring for humans. It gathers the caller’s information, assesses urgency, and schedules a visit with the customer. For emergencies, it dispatches the on-call tech immediately.

Abandoned calls. Call centers are not perfect. If you’re forced to put a customer on hold, that customer might just find a different vendor. AI will pick up no matter what. If it can handle it, it will. And if not, it takes a detailed message—like an answering service—and asks clarifying questions. No more abandoned calls. Ever.

Multilingual support. Many businesses serve areas with large Spanish-speaking populations. Hiring bilingual staff is hard and expensive. The AI handles both languages natively, 24/7, with zero additional cost.

What’s Hard

Unique situations. Every business has oddball situations: “The new install broke and the manager promised me a $1,000 rebate if it breaks within 30 days.” That’s a call the AI won’t know what to do with. Is the thing really broken? Did the manager really promise a rebate? Is there a way to make the customer happy without giving the rebate? These calls need a human. The AI’s job is to recognize when it’s too complicated and hand it off gracefully.

AI can be brittle. You’ve heard about “hallucinations” and since AI is probabilistic, getting it to work consistently for every scenario is hard. Configuring it is not like traditional software configuration. In traditional software, you can just check a box “Turn on Notifications” and expect to get a notification 100 percent of the time. But in AI, you write the instructions in English (aka The Prompt), and because AI is probabilistic, you can’t be sure it will do the right thing 100 times or if some other instruction is now being ignored.

What’s Coming

Far smoother voice interactions. Today’s voice AI is still largely turn-based—it speaks, then listens, then speaks. Future voice AI will handle full-duplex conversation: talking and listening simultaneously, just like humans do. It will detect frustration in a caller’s tone and slow down, soften its language, or escalate to a manager before the customer has to ask.

Giving voice AI its own computer. Today’s voice AI is limited not by its ability to think, but by the types of tasks it can perform. It can talk, but it can’t do a whole lot. Future voice AI will have access to its own computer—a browser, databases, a coding environment to write its own programs—and take actions to help the customer during the call. Imagine an AI that can simultaneously pull up a specific 1990s furnace manual, check the local distributor’s inventory for a capacitor, build a cost model, and present the homeowner with a “repair” versus “replace” choice all while on the call.

5 Steps to Get Started with AI

Now that you know what’s hard and what voice AI does well, here are a few tips for integrating voice AI into your business: 

Start small. Find a key use case where you’re not staffed properly. For most businesses, it’s those sudden spikes (after hours, for example) where it’s impossible to staff correctly. Let AI handle this one job during the spikes.

Monitor and develop trust. You will quickly see where your AI is working and where it needs better configuration. You can treat the AI as a new employee you just hired, and it needs some guidance before it can do its job right.

Scale up. When you’re happy with its performance with the task, you can expand its usage—perhaps that means adding more hours in the day, or you can have AI be the first touchpoint for your customers and promote your customer service reps to AI managers.

Measure what matters. AI has drift. AI models change underneath the hood all the time. These upgrades are mostly better, but sometimes an upgrade breaks a previously working scenario. So you should be diligent and measure the things that are important for this task. For trade companies booking appointments, it should be booking rate and escalation rate. For customer service, it should be resolution rate and escalation rate.

Voice AI is already very useful in 2026 and is capable of much more than transcription and simple tasks. It’s a practical tool that can free up employees, cut hold times, and provide consistent service. The space is moving rapidly and capabilities (task complexity, error rates, etc.) are all improving 2x every 6 months. Now is the ideal time for businesses to integrate voice AI into their operations.

Jeff Chen is CEO of Redcar AI, a San Francisco-based company that builds voice AI for call centers. Previously, he led product development at Google, where he helped launch the Google Voice Assistant. Chen studied computer science at Berkeley and Stanford University and is passionate about bringing enterprise-grade AI to every business.