What is answering machine detection, and how accurate is it?

Answering machine detection (AMD) is dialer software that listens to the first moments of an answered call and decides who picked up: a person, a voicemail greeting, a phone menu or a call screener. Its accuracy depends on how long it listens. Listening longer means fewer mistakes, but the person who answered waits in silence, and US rules and UK regulator policy limit that wait to 2 seconds.

How answering machine detection tells a person from a machine

AMD starts when the call is answered and listens before your rep hears anything. In power and parallel dialing, it decides which answers reach a rep and which are dropped.

Older systems judge the shape of the sound. A person tends to say a short “Hello?” and then wait for a reply. A recorded greeting runs on in long strings of words and often ends with a beep. Amazon’s documentation for Amazon Connect Customer, its contact center service, lists these cues: background noise typical of a recording, long strings of words, and a short greeting followed by silence. Systems built on AI speech models can also weigh the words themselves, so “Press 1 for sales” reads differently from “Sorry I missed your call.”

Each answer type sounds different and needs a different action:

Answer	What the system hears	What should happen
Person	A short greeting, then a pause while they wait for a reply	Connect the rep at once
Voicemail	A long recorded greeting, recording noise, often a beep	Drop the line, or let the rep leave a message
Phone menu	“Thank you for calling”, then options or a dial-by-name prompt	Skip it, or let the rep work the menu for a key account
Call screener	A phone assistant asking who is calling and why	The rep gives their name and reason, then waits
No answer or a dead number	Ringing, tones or silence	Log it and move on

Phone menus matter most in B2B calling, where a company’s main number may answer with an automated attendant rather than a person.

Why faster detection is less accurate

AMD has to decide with partial information. The sooner it decides, the less it has heard and the more often it is wrong. The later it decides, the longer a real person sits in silence. No setting removes both problems; you choose which one you accept more often.

The mistakes come in two kinds:

• A person labeled as a machine (a false positive). The dialer hangs up or never connects the rep, so someone who answered hears nothing. Amazon’s documentation notes that a long greeting from a live person can be read as voicemail, and B2B greetings are often long: “Good morning, Northwind Traders, Priya speaking.”
• A machine labeled as a person (a false negative). The rep is connected to a recording and loses a few seconds. Amazon notes that a voicemail greeting that is a short “Hello”, or that includes a pause, is read as a person.

The first mistake is the expensive one. It costs a conversation, it annoys someone who picked up, and under the rules below it can count as an abandoned or silent call. The second costs a rep a few seconds.

This is why there is no single accuracy figure for AMD. It changes with the country and language, the mix of mobile and office lines, how people greet callers, and how long the system may listen. Amazon’s documentation says detection accuracy has limits, and it recommends routing answers the system cannot classify to an agent, as if a person had answered. Ask any vendor what their system does when it is unsure, and check the answer on your own calls.

The 2-second rule: where waiting for detection gets risky

Both the US and the UK look at how long a person waits after answering, and AMD listening time counts toward that wait. As of October 2026:

United States. The FTC’s Telemarketing Sales Rule (16 CFR 310.4(b)(1)(iv)) and the FCC’s rule (47 CFR 64.1200(a)(7)) count a call as abandoned when the person is not connected to a sales rep within 2 seconds of finishing their greeting. Someone AMD hangs up on is never connected, so that call is abandoned too. The FTC’s safe harbor (310.4(b)(4)) asks you to abandon no more than 3% of calls answered by a person per campaign over 30 days, let unanswered calls ring for at least 15 seconds or 4 rings, and play a recorded message with the seller’s name and phone number when no rep is free. The FCC rule makes the same 3% figure a limit rather than a safe harbor, and requires that message to say the call was for telemarketing and to offer an automated opt-out by voice or key press. The FTC’s rule exempts most calls to businesses (16 CFR 310.6(b)(7)); ask your counsel how each rule applies to your lists.

United Kingdom. Ofcom’s persistent misuse policy (published 20 Dec 2016, in force since 1 Mar 2017, and the current statement on Ofcom’s site as of October 2026) names AMD as a cause of silent calls: it can mistake a person for an answer machine and hang up. The statement makes action against silent calls Ofcom’s highest priority. Ofcom closed its enforcement program on silent and abandoned calls in 2023, but a reminder it published on 22 Jan 2026 says it can still take action.

Rule	United States (FTC and FCC)	United Kingdom (Ofcom)
The 2-second clock starts	When the person finishes their greeting	When the person starts to speak, if the system listens to classify the call; otherwise at pickup
If no rep is free	A recorded message with the seller’s name and phone number; the FCC adds a voice or key-press opt-out	A message that names the caller, says they tried to call, and gives a number to decline further calls, with no marketing
Abandoned-call rate	No more than 3% of calls answered by a person, per campaign, measured over each 30 days (an FCC limit, and part of the FTC safe harbor)	No fixed rate: Ofcom says the 3% figure in its 2010 policy was never a safe harbor
Ring time	At least 15 seconds or 4 rings	At least 15 seconds, and not left ringing for long periods
Calling back	Not covered in these sections	Calling again within 72 hours of an abandoned call counts against you unless a rep is certain to be free for that call
Records	Keep records that show compliance	Keep records for at least 6 months; Ofcom may ask for evidence of AMD accuracy

This is a plain summary, not legal advice.

Call screeners: a new answer on sales calls

Phones now answer some sales calls for their owners.

• iPhone. On iOS 26 and later, Call Screening can answer calls from numbers that are not saved in contacts. With Ask Reason for Calling turned on (Settings, then Apps, then Phone), the iPhone asks the caller for their name and reason, then rings and shows the response so the owner can decide whether to pick up. The other choices are Never, which rings as usual, and Silence, which sends unknown numbers to voicemail.
• Google Pixel. Call Screen answers and asks who is calling and why. As of October 2026, automatic screening of unknown numbers runs on Pixel phones in Australia, Canada, Ireland, the UK and the US. Calls it judges to be spam are hung up; the rest ring through with the caller’s answer on screen. Numbers saved as contacts are not screened.

For a dialer, a screener is neither a person nor a voicemail. Hang up and the call never reaches the person. Pass it on as a person and the rep must know they are speaking to an assistant, not the buyer. The right move is a short, plain answer: the rep’s name, the company and the reason for the call, in one or two sentences. That answer is what the person sees before deciding to pick up.

How to set up answer detection on your dialer

1. Decide what each label does. Use the first table: connect people at once, drop or leave voicemail, skip or work phone menus, and answer screeners.
2. When in doubt, connect the rep. An unsure answer sent to a rep costs a few seconds. An unsure answer hung up can cost a conversation and count as abandoned.
3. Keep the required message ready. If no rep is free within 2 seconds, the message the country requires should play, never silence.
4. Count AMD hang-ups in your abandonment rate. Include every person the system hung up on, per campaign.
5. Listen to a sample every week. Play back calls labeled voicemail and phone menu, and count the people among them. That is your false-positive rate on your own lists.
6. Write a screener line. Give reps one sentence with their name, company and reason, and have them practice it.
7. Check each country and language. Greetings differ by market, so review detection results for each one before you add more calls there.

How DialSavvy handles this

DialSavvy’s AI answer detection labels every answer as a person, voicemail, a phone menu or a call screener, and every outcome is logged. In parallel mode the first person who answers reaches the rep; if no rep is free, the line ends with the message the country requires, never an AI voice. The abandonment governor drops to 2 lines from 2.5% abandoned calls in a session, and to 1 line from 2.8%. Parallel dialing turns on in a country only after answer detection meets an accuracy gate on its languages and carriers (country coverage). Instant connect, where the person hears the rep at once while AI keeps listening in the background, is coming. See the DialSavvy dialer.

Start with last week’s calls: pull 20 labeled voicemail and listen for people. Then see how DialSavvy checks every call before it rings, or book a demo and bring a list to dial with us.