IVR handles inbound VoIP or PSTN calls to a business DID. After the call is answered at the platform, the IVR plays audio (recorded or text-to-speech), waits for input, then transfers, queues, takes a message, or continues a script. The classic input channel is DTMF (touch-tone). Many systems also accept limited speech recognition ("sales," a zip code, an account number).
IVR is a routing and self-service layer, not a phone number and not a queue. A number still has to land somewhere. The IVR decides the next hop: another menu, a hunt group, a call queue, call forwarding to a cell, voicemail, or an automated transaction (balance, appointment confirm, password reset).
Auto-attendant vs multi-level IVR vs conversational agents
Vendors use the word IVR loosely. Three designs are worth keeping separate, even when one product implements all three.
| Design | What the caller hears | Typical input | What it is for |
|---|---|---|---|
| Auto-attendant (single-level) | One greeting, a short list of departments or people | One DTMF digit, sometimes a name-dial | Replace a live receptionist for simple "press 1 for sales" routing |
| Multi-level IVR | A tree of menus, often with hours, language, and account prompts | Several DTMF steps, optional speech, lookup against a CRM or billing system | Self-service plus finer routing (billing vs new order vs existing ticket) |
| Conversational / AI agent | A spoken dialogue instead of a fixed tree | Natural language, with DTMF as fallback | Tasks that do not map cleanly to a digit map: "I need to move Tuesday's appointment" |
A conversational agent (see AI receptionist) still needs the same telephony primitives: answer the SIP INVITE (or WebRTC session), play media, collect input, and REFER or re-INVITE the call to a person or queue. The difference is the decision engine (intent, dialogue state) rather than a static digit map.
DTMF and speech input
DTMF remains the reliable path through a PSTN or SIP trunk. Digits travel as RFC 4733 telephone-events (preferred), as SIP INFO, or in-band in the audio. In-band DTMF breaks when a codec compresses the tones. RFC 4733 avoids that. Speech recognition adds a second channel: the IVR must stream audio to a recognizer, handle barge-in (the caller speaks over the prompt), and recover when the recognizer is unsure.
Timeouts, invalid digits, barge-in
- Timeouts: if the caller enters nothing, replay the prompt or route to an operator.
- Invalid digits: replay, then overflow after N failures rather than looping forever.
- Barge-in: stop the prompt when DTMF or speech starts, so the tree feels responsive.
- Language and hours: a first menu for language, then open/closed behavior, is a multi-level tree, not a separate product.
Where IVR sits in the call path
- A caller dials a DID, local number, or toll-free number.
- The terminating provider delivers the call over SIP (or from the PSTN through a gateway).
- The platform answers and connects the caller to the IVR media server.
- The IVR collects input, optionally looks up a record, then transfers: queue, hunt group, extension, cell via forwarding, or disconnect after a self-service result.
Transfers may be a SIP REFER, a re-INVITE, or a new outbound leg (tromboning). The caller should keep hearing ringback or hold audio. If the transfer target is a mobile, E911 and caller ID on the outbound leg are separate problems from the inbound menu.
Design rules that affect completion
- Keep the first menu short. Long spoken lists lose callers before digit 1.
- Offer a person early (0 or a stated operator option). Dead-end trees raise abandonment.
- Do not collect payment card data in an IVR unless the path is PCI-scoped and recording is paused. See call recording.
- If the next hop is a queue, say so. Silence after a digit feels like a drop.
- Match hours to routing: after-hours should not ring an empty hunt group.
IVR, hunts, queues, and forwarding
| After the menu | Mechanism | Caller experience |
|---|---|---|
| One person or one cell | Call forwarding (CFU / CFB / CFNR) | Ring the target. If no answer, voicemail or next rule. |
| A small team, no wait | Hunt group / ring group | Members ring in order or at once. No position in line. |
| A team that can be busy | Call queue / ACD | Wait, position or estimated wait, overflow, SLA. |
Standards and limits
- Media
- Prompts are RTP audio, subject to the same codecs and call quality limits as any VoIP leg.
- Signaling
- The IVR is a SIP user agent (or a CPaaS application controlling one). It is not a substitute for STIR/SHAKEN on outbound legs.
- Messaging counterpart
- SMS short codes and A2P keywords are a text analog of a menu. They are not IVR.