If your outbound AI voice agent is burning budget, voicemail detection is almost certainly the reason. Roughly 30 percent of answered outbound calls in 2026 reach a voicemail greeting or an iOS 26 Call Screening bot rather than a person, and an agent that cannot tell the difference will deliver its full pitch into a recording, or hang up on a real prospect who took a beat to say hello. Detection happens in the first three seconds of audio. Configure it properly and those calls either leave a clean message or exit cheaply. Leave it on defaults and you pay full conversation rates for calls no human ever hears.
We've deployed outbound voice agents for home services, medical practices, and B2B follow-up, and this is the setting clients most often ship wrong. Not the prompt. Not the voice. The three seconds after pickup.
What voicemail detection actually decides
Answering machine detection is a legacy name for something that now does considerably more than detect answering machines. The classifier listens to the opening audio and sorts the pickup into a category. The useful ones are:
- Human — a live person said hello. Greet normally.
- Voicemail — an outgoing greeting is playing. Stay quiet, wait for the beep, leave the message.
- IVR — an auto-attendant or phone tree. There is no message to leave. Exit and flag it.
- Screener — iOS Call Screening or a similar assistant asking who's calling. Answer that question in one sentence and nothing more.
Most platforms ship a binary human-or-machine flag, and that binary is the root of the problem. An IVR tree is not a voicemail box, but a two-bucket classifier has to call it one or the other. If it guesses voicemail, your agent monologues into a phone menu and the message goes nowhere. If it guesses human, your agent tries to have a conversation with a recording that says press one for sales.
Why a third of your calls never reach a person
Apple's Call Screening answers calls from unknown numbers on the user's behalf and asks the caller to state their name and purpose. The phone doesn't ring until that's done. From a classifier's point of view this is the hardest possible input, because the screener is doing an excellent impression of a person. It speaks conversationally. It pauses for a reply. It sounds human because it's designed to.
So the agent classifies it as human and runs the opening pitch. The screener collects thirty seconds of marketing copy, summarizes it as spam, and the call is over before the phone ever rings.
The practical effect is that outbound answered-call composition has shifted. Voicemail was always a chunk of it; screening bots are new, growing, and invisible in your dashboards unless you're specifically labeling them. If your reporting shows a healthy answer rate and a terrible conversation rate, you're probably looking at this. It's the same category of measurement problem we wrote about in why AI SDRs fail — the metric looks fine while the outcome quietly rots.
The three detection methods, and which one to use
Silence-pattern rules
The oldest approach. Measure how long the pickup speaks and how long it pauses. Genesys, for example, treats under 2,200 milliseconds of speech followed by 700-plus milliseconds of silence as a human. It's cheap and it works on stock greetings. It falls apart the moment someone has a long custom greeting or a receptionist who likes to talk, both of which get classified as machines.
Audio classifiers
Score the waveform directly without transcribing. Bland fine-tuned a Wave2Vec model on the first two seconds of audio and reports 98.5 percent accuracy; CNN approaches over Mel spectrograms land around 97 percent. Fast and accurate on the binary question. The limitation is that a waveform classifier can tell you machine, but it can't tell you which kind of machine, so you still can't branch between voicemail, IVR, and screener.
Transcript plus LLM
Transcribe the opening audio and let a small model read the actual words. This is where the industry has landed for agent use cases, because words are what distinguish the four categories. LiveKit publishes 95.7 percent F1 on human, 98.2 percent on IVR, and 97.3 percent on voicemail, with a median time to decision around 840 milliseconds.
For agency deployments, use transcript plus LLM. The accuracy is comparable, the latency is fine, and it's the only method that gives you the four-way classification you actually need to configure sane behavior. Every vendor benchmarks on its own audio, so treat all of those percentages as directional rather than comparable.
The settings that quietly cost you money
Three configuration values do most of the damage.
Beep wait. After detecting voicemail, how long the agent waits for the tone before speaking. Vapi defaults to about 30 seconds, and teams routinely cut it to "speed things up." That's a mistake. Beep wait tuned to a stock greeting will clip every message left on a long custom greeting, and long custom greetings are exactly what small businesses have. Leave it near the default. The cost of waiting is a few cents of connection time; the cost of clipping is the entire call.
Detection timeout. How long the classifier gets before it gives up and guesses. Twilio's AMD accepts 3 to 59 seconds and defaults to 30, and it also splits machine outcomes three ways — machine_end_beep, machine_end_silence, and machine_end_other. If you're consuming that signal, branch on all three rather than treating any machine result as identical.
Synchronous versus asynchronous. This is the one that matters most. Synchronous AMD holds the audio while it decides, which means a human who picks up hears several seconds of silence before your agent says anything. People hang up on silence. Async detection runs the classifier in parallel with the conversation model and gates only the TTS output, so a human hears a normal greeting while classification finishes in the background. Retell documents under 30 milliseconds of added latency on its path. If you're running synchronous AMD on outbound, change that before you touch anything else.
Every one of these has a direct line to unit economics — worth reading alongside what AI voice agents actually cost per minute, because a call that pitches into a recording still bills as a full conversation.
The six ways this fails in production
Every one of these we've seen on real client traffic:
- False positive on a human. A slow or non-native speaker crosses the silence threshold and gets classified as a machine. The agent either hangs up or starts reading a voicemail script at a confused person.
- False negative on voicemail. The greeting is missed, so the agent talks over it and the first ten seconds of the message — usually including the callback number — are lost under the outgoing greeting.
- Latency on human pickup. Synchronous detection inserts silence. A meaningful share of people hang up during it.
- IVR treated as voicemail. The agent waits for a beep that never comes, then delivers a message into a phone tree.
- Screener treated as human. The full pitch goes to iOS Call Screening, which summarizes it as spam.
- Language and accent bias. Classifiers trained mostly on English get worse on Spanish-language voicemail greetings and regional cadences. If you run bilingual outbound — as we do for several Dallas-area clients — test both languages separately, because the numbers are not the same.
Notice that four of the six are behavior problems, not accuracy problems. The classifier can be right and the deployment still fails because nobody defined what the agent does with the verdict. Write an acceptance rule per class before you tune a single threshold. The same discipline applies to handing a call to a human and to handling interruptions mid-sentence: define the behavior explicitly, don't let the model improvise it.
How to test before you turn on the dialer
Transcript-based testing will lie to you. Transcripts don't carry codec artifacts, packet loss, or the compression that makes a real cell connection sound the way it does, and those are precisely what breaks classifiers. Test over real telephony or don't bother.
Build a persona list and call each one deliberately. Ours covers a fast talker, a slow talker, a long business greeting, a Spanish-language greeting, a stock carrier voicemail, an IVR tree, an iPhone on iOS 26 with Call Screening enabled, and a number that rings out with no answer at all. For each call, score three things independently: did it classify correctly, did it classify fast enough, and did the agent then do the right thing. A call can pass the first two and still fail the third.
Then keep the failures. Every misclassification from production becomes a regression fixture, and after a couple of months you have a test set that reflects your actual traffic rather than a vendor's benchmark. Wire it into your release process so any change to prompts, models, or telephony config re-runs it.
One more thing worth checking while you're in here: your voicemail message itself. Keep it 15 to 20 seconds and state the callback number twice, once early and once at the end. That way a message clipped at either edge still does its job. And confirm your drop practice matches the disclosure rules covered in AI voice agent compliance — a recorded message is still an outbound communication, and the rules apply to it.
What we'd check first on your setup
If you inherit an outbound agent and have ten minutes, look at these in order: is detection on at all, is it async, does it classify into four buckets or two, is beep wait near default, and is there a defined behavior for IVR and screener. In our experience that ordering finds the problem on the first or second item most of the time.
Choosing between platforms before you build? We broke down the tradeoffs in Vapi vs Retell for agencies. Both handle detection competently in 2026; the differences show up in how much control you get over the branching.
If you're running outbound and you can't say off the top of your head what share of your answered calls hit voicemail, that number is worth finding this week. It's usually larger than people expect, and it's the cheapest thing to fix in the whole stack.
Get a Free Automation Audit — we'll review your outbound voice agent's detection config, listen to a sample of real call recordings, and tell you exactly where the budget is leaking. No pitch deck. If you want the broader picture first, here's how we build voice AI agents.