IVR menus are cheap and predictable. AI phone agents handle the calls that don't fit the tree. An honest comparison of cost, containment, and TCPA rules.
You know this call. Someone dials your business, hears "press 1 for billing, press 2 for scheduling," picks the closest match, gets three more options that don't fit, and starts mashing zero. Forty seconds in, they're holding for a human anyway, more annoyed than when they dialed.
That menu tree is an IVR (interactive voice response). It has been answering business phones since the nineties, and for one narrow job it's still the right tool. The problem is every call that doesn't fit the tree, and those are usually the calls that cost you the most.
We've shipped 30+ projects to production since 2024, including voice and call workflows, and this is the comparison we walk prospects through before anyone spends a dollar.
| IVR | AI phone agent | |
|---|---|---|
| How it works | Deterministic menu tree. Keypresses or rigid voice commands walk a finite set of paths. | Speech-to-text, a language model that reads intent, tools that hit your calendar or CRM, text-to-speech back. |
| Cost | Cheap. Usually bundled with your phone system, plus configuration time. | Build: $8k-$20k for a single-purpose agent. Inference typically runs $50-$2,000 per month, and good engineering (caching, model routing, prompt design) cuts that 3-10x. |
| Setup | Days. Draw the tree, record the prompts. | 3-4 weeks for a scoped agent, including the evaluation suite. |
| Containment rate | High for pure routing. Collapses the moment a caller's problem isn't on the menu, and they zero out. | Handles the long tail of intents, but only as reliably as the eval suite behind it. Unevaluated agents contain demos, not calls. |
| Failure modes | Fails loudly and predictably. Wrong branch, abandoned call. Easy to spot in metrics. | Fails quietly and creatively. Wrong data read back, invented policy answers. You need transcripts, guardrails, and review. |
| Compliance | Trivially auditable. Every possible path can be printed on one page. |
| Needs logging, transcript retention, and consent handling. Outbound use sits under the TCPA (more on that below). |
An IVR is a finite state machine wearing a phone voice. That's its strength. Every path is enumerable, nothing is improvised, and it will never invent a refund policy at 11pm. If a regulator asks what your phone system can say, you hand them the tree.
It's also the weakness. Real callers don't sort themselves into eight buckets. The practical ceiling is two menu levels. Past that, each layer is another chance for the caller to give up, and the ones who give up are disproportionately the ones with billable problems. They don't file a complaint about your menu. They call your competitor.
Here's a real shape of call: "I got a bill for the visit my insurance was supposed to cover, and I also need to move Thursday's appointment." No menu branch exists for that. An AI phone agent handles it as one conversation, because it reads intent instead of matching keypresses. It checks the billing record, flags the claim for review, moves the appointment, and confirms both back in plain language.
The catch is real, so we'll say it plainly: prototypes lie. A voice demo that nails five hand-picked calls tells you nothing about five thousand real ones with accents, background noise, interruptions, and half-formed questions. Moving from 90% to 99% reliability is where the actual engineering lives. That means an evaluation suite: recorded scenarios, scored transcripts, regression runs on every prompt change. An agent without one is a liability with a phone number.
We tell some prospects to keep their IVR. The cases are consistent:
We take on two engagements per quarter, so we have no incentive to oversell this. If your call problem is routing, keep the IVR and spend the budget on a workflow that actually hurts.
Technically, yes. Any modern voice platform can dial out. The real question is legal, and the answer is "only with consent, and carefully." What follows is an informational overview, not legal advice. Run any outbound calling plan past a TCPA attorney before you dial.
The TCPA basics as they stand in mid-2026:
Beyond the law, there's trust. Robocall-style outbound is a brand tax even when it's technically legal, for the same reason fake urgency banners are: people can smell manufactured pressure, and they remember who applied it. The outbound that works is consent-based and expected. Appointment reminders. Order updates. A callback the person requested an hour ago. The caller wanted the call. That's the line, and we don't build on the other side of it.
Two verticals come up in almost every scoping conversation.
The workhorse use cases are appointment reminders, scheduling, and refill-status routing. Reminders alone move the no-show number, because the agent can actually reschedule on the spot instead of leaving a voicemail into the void.
HIPAA shapes the whole build. Every vendor touching protected health information (telephony, transcription, the model provider) signs a business associate agreement, or they're out. Messages follow minimum necessary: confirm the time, not the condition. Transcripts are logged and retained under the practice's policy. Reminders are generally permitted as treatment communications without separate authorization, but keep the content thin and the audit trail thick.
This one is about speed-to-lead. Agents miss calls during showings, and a buyer who hits voicemail dials the next listing within a minute. A voice agent answers instantly, qualifies the lead (budget, timeline, pre-approval, target area), books a viewing against the live calendar, and texts a summary to the human agent. Nothing clever, just answering while intent is hot, every time, including Sunday afternoon.
A pattern we keep meeting: an ops team where somebody is on call at 2am doing phone triage by hand, deciding whether tonight's call is a real emergency, because the phone system doesn't talk to the ticketing system or the on-call rota. We've seen it in property management, field services, and clinics. That gap is usually one integration and one agent away from gone. The agent answers, asks the same structured questions the on-call person would ask, pages a human for genuine emergencies, and books everything else into the morning queue. Nobody's sleep should be an integration layer.
Every voice agent needs a way out. When confidence drops, when the caller is in a loud environment, or when the answer involves things you shouldn't read aloud (links, addresses, confirmation codes), the right move is "I'll text you the details." An SMS chatbot picks up the thread, sends the booking link, and confirms in writing.
It's also cheaper per interaction than a voice minute, it leaves a record both sides can scroll back through, and plenty of callers prefer it for anything with details in it. One caveat: outbound texts live under the same consent logic as outbound calls. Don't text people who didn't ask.
Is it legal for AI agents to make outbound calls?
Yes with the right consent, and expensively risky without it. The FCC treats AI voices as artificial voices under the TCPA, so marketing calls need prior express written consent and informational calls need prior express consent. Statutory damages run $500 to $1,500 per call. This is an overview, not legal advice. Involve a TCPA attorney before launching any outbound program.
What does an AI phone agent cost?
For our builds, a single-purpose agent (after-hours triage, appointment scheduling) runs $8k-$20k fixed price and ships in 3-4 weeks. Running inference typically costs $50-$2,000 per month depending on call volume, and caching, model routing, and prompt design cut that 3-10x.
Should I replace my IVR entirely?
Usually not on day one. The common pattern is a hybrid: keep a one-level front door for the two or three high-volume routes, and hand everything else to the agent. Or run the agent first with the IVR and a human queue as fallback. Measure containment before you delete anything.
What containment rate should I expect?
Distrust any vendor who quotes a number before seeing your call mix. Containment depends on how many of your real intents the agent covers with working tools, and on the eval suite keeping it honest. Record and categorize a week of real calls first, then set targets per intent.
If you're weighing an IVR against an AI phone agent for your line, book a free 30-minute scoping call. We'll look at your actual call mix, tell you honestly if the boring option is the right one, and get back to you within two business days. We take on two engagements per quarter, and every client owns 100% of the code.
Book the call or read how we build and evaluate these systems on our AI agent development page.
