You call a company. A robotic voice says: "Welcome to Acme Corp. For billing, press 1. For technical support, press 2. For sales, press 3. For all other inquiries, press 4. To repeat these options, press 9."

You press 2. "For software issues, press 1. For hardware issues, press 2. For network issues, press 3."

You press 1. "Please hold while we connect you to the next available representative." Hold music. Three minutes. Five minutes. Seven minutes.

IVR (Interactive Voice Response) has been the worst part of customer service since the 1990s. Everyone hates it. Companies know everyone hates it. They use it anyway because routing calls without it would be chaos.

Voice AI is changing this. Instead of button presses and menu trees, you talk to an AI that understands what you want and routes you accordingly. Or, increasingly, handles the issue entirely.

What Voice AI Actually Does

Modern voice AI agents (from companies like Retell, Bland, Parloa, PolyAI, and others) use speech recognition, natural language understanding, and text-to-speech to have actual conversations over the phone.

You call. The AI answers: "Hi, thanks for calling. How can I help?" You say: "I need to change my shipping address for an order I placed yesterday." The AI understands the intent, asks for your order number, looks it up, and either changes the address or connects you to someone who can.

The conversation is natural. The AI understands context, handles interruptions, and can ask clarifying questions. It doesn't say "I didn't catch that, please try again" (well, not as often as old IVR did).

Response time is under a second in most cases. The voice sounds human-like (some uncannily so). And it works 24/7 without hold times.

Where Voice AI Works Well

High-volume, simple interactions. Appointment scheduling, order status checks, password resets, account balance inquiries, store hours. These are the calls that IVR was designed to deflect, but IVR does it badly (because navigating menus is painful). Voice AI does it well.

After-hours coverage. A dental office that closes at 5pm can route after-hours calls to a voice AI that schedules appointments, answers common questions, and takes messages for the morning. No answering service needed.

Call routing. Even if the AI can't resolve the issue, it can understand what the caller needs and route them to the right department without a menu tree. "I have a question about my bill" goes to billing. "Something broke" goes to support. No "press 1, press 2."

Outbound calls. Appointment reminders, payment reminders, and satisfaction surveys. These are tedious for humans and perfect for AI. The AI calls, delivers the message, handles basic responses ("yes, I'll be there" or "I need to reschedule"), and logs the result.

Where Voice AI Struggles

Accents and speech patterns. Voice AI has improved dramatically, but it still struggles with heavy accents, speech impediments, background noise, and regional dialects. A caller from rural Alabama and a caller from Queens, New York use very different speech patterns. Models trained primarily on standard American English may misunderstand both.

Emotional calls. An angry customer who calls to complain doesn't want to talk to a machine. They want to vent to a human who can empathize and make things right. Voice AI can detect anger (through tone analysis), but it can't genuinely respond to it. The best approach: detect the emotion, say "I can hear this is frustrating. Let me connect you with a team member who can help," and route to a human immediately.

Complex troubleshooting. "My internet keeps dropping every 30 minutes but only on my laptop, not my phone, and it started after I updated my router firmware." This requires back-and-forth problem solving that voice AI isn't reliable enough for yet. It can gather the initial information, but a human needs to drive the diagnosis.

Elderly callers. Many older customers prefer phone support specifically because they're uncomfortable with digital channels. Voice AI that's too fast, too robotic, or too confusing defeats the purpose of phone support for this demographic.

The Cost Comparison

Traditional phone support with human agents:

An agent handling phone calls costs $40,000 to $55,000/year in salary and benefits. They handle about 40 to 60 calls per day (accounting for handle time, after-call work, and breaks). That's roughly $3 to $5 per call in labor cost.

After-hours coverage with an answering service costs $1 to $3 per call, but service quality is usually poor and they can't actually resolve anything.

Voice AI agents:

Pricing varies by vendor. Retell and Bland charge $0.07 to $0.15 per minute. An average call of 3 minutes costs $0.21 to $0.45. That's 10x cheaper than a human agent per call.

No overtime, no shifts, no hold times, no sick days. The AI handles 10 calls simultaneously or 1,000. Same cost per minute.

At 1,000 calls per month with an average 3-minute duration: voice AI costs $210 to $450/month. Human agents handling the same volume (50 calls/day, one agent) cost $3,500 to $4,500/month.

The Hybrid Phone Setup

Most companies won't go fully AI on phone support. The best approach mirrors text-based hybrid support:

AI answers the call. Understands the intent. Handles simple issues (appointment scheduling, order status, account updates) automatically.

If the issue is complex, emotional, or the caller requests a human, AI transfers the call with full context. The human agent knows what the caller wants before they pick up.

AI handles 40 to 60% of call volume. Humans handle the rest. Total phone support cost drops 30 to 50%. Wait times drop to zero for AI-handled calls and shrink significantly for human calls (because volume is lower).

What This Means for Text-Based Support

Voice AI's rise doesn't replace text-based support. It complements it. Customers who prefer phone get better phone support. Customers who prefer chat or email still use those channels.

The interesting convergence: whether the customer calls, chats, or emails, the underlying AI capability is the same. Understand the intent, route or resolve it, hand off when needed, and log the result. The channel is just the interface.

Supp handles the text-based side of this: classifying messages from chat, email, and messaging platforms into 315 intents with 92% accuracy. If you're looking for voice AI, pair a voice AI agent (for phone) with Supp (for text) and you have a unified classification layer across all channels.

The IVR phone tree is dying. Good riddance. What replaces it will be better for customers and cheaper for businesses. The transition just needs to be done carefully, with human fallbacks, accent sensitivity, and emotional intelligence that today's AI still can't fully provide.