Delayed-playback FAS: the route fraud your dial tests cannot see

Classic false answer supervision is crude: a route signals answer before the callee picks up, or plays fake ringback into a call that never connected, and the billing clock starts on a conversation that does not exist. Operators learned to catch that years ago, because it fails the most basic check: place a call and listen to what comes back.

The delayed-playback variant is the patient one. The call connects for real: genuine ringback, a genuine answer, often a real human who opens a real conversation. Every early signal you would check is legitimate. Then, roughly a minute or two in, the vendor swaps the callee leg for a recording, a loop, canned prompts, speech that has stopped responding to anything the caller says. The signalling never changes. The call stays up, the caller keeps talking, and every one of those minutes is billed as conversation.

Wholesale voice is billed by the minute, and a terminating vendor is paid for every minute it reports as answered. Delivering a real conversation to a real callee is the expensive part. A recording played into an already-answered call is nearly free. Delayed-playback FAS converts one genuine answer into an open-ended stream of billed minutes with almost no delivery behind them. A caller who was mid-conversation when the switch happened stays on the line far longer than one hearing dead air, so the most profitable victim is a caller engaged enough to keep talking.

CCaaS and UCaaS providers

Your customers' minutes ride on wholesale routes you buy, and so does your reputation. When an agent's call quietly turns into a one-sided conversation with a recording, the complaint lands on your platform, not on the carrier three hops down.

CPaaS platforms

Programmable voice runs at a scale where nobody listens to calls: thousands of API customers, traffic spread across multiple carriers, routing chosen by rate and measured by connect statistics. A poisoned route inflates duration across an enormous call volume while every dashboard metric, answer rate, connect rate, looks healthy.

Carriers

An interconnect partner misbehaving on a destination costs you twice: you pay the padded minutes, and the customers you resell that route to blame you for the quality. Without a recording of what the callee leg was doing, the dispute is your word against your supplier's.

Test calls are short

A test call confirms answer supervision and audio, then hangs up within seconds. The switch to playback lands after one to two minutes of conversation, comfortably outside the window any affordable test campaign covers.

Test numbers are known

Test campaigns dial known destination numbers from recognizable caller IDs, on schedules. A vendor running playback fraud can learn those patterns and behave for them: the test numbers get clean termination, and the fraud runs on everything else.

There is no conversation to betray it

Even a long test call carries a script, a tone or silence. Delayed playback only exposes itself against a real conversation, a callee leg that stops responding to what the caller is saying. A synthetic call has no such exchange for the switch to break.

The structural answer is to stop probing the route with synthetic calls and start reading a sample of the real ones. A dial plan that can split traffic across carriers can route a percentage of live calls, one in ten, one in twenty, through an analysis hop on the way to their normal destination. The sampled call is ordinary traffic: a real caller, a real reason to stay on the line, and nothing for the vendor to recognize and behave for.

• Both legs, separate channels. The hop records the caller on one channel and the callee on the other, so what the callee leg was doing in minute three is readable on its own.
• Per-channel transcription. Each channel is transcribed separately and merged by time, so every line is attributed to caller or callee, the raw material for spotting a leg that switches from answering to looping.
• The switch becomes a route alert. A callee channel that turns from interactive conversation into repeated phrases or canned playback flags the call as a route alert naming the route to pull, rather than as a suspect caller, with the recording and transcript attached as the evidence for the vendor dispute.

Open Voice Shield is that analysis hop. It is an ordinary SIP hop authenticated by IP whitelist: your switch routes a sampled share of calls through it, it relays them onward to your own destination, and every call that passes through is recorded on separate channels, transcribed per channel and analyzed after hangup. The FAS detection page describes the callee-channel check in detail, and how it works shows where the hop sits in the call path.

Sampling is also what makes the economics workable. On pay-as-you-go there is no monthly fee: a call is charged for its minutes plus one flat analysis fee, so sampling a tenth of your traffic is charged on a tenth of the calls. Continuous route surveillance costs about as much as a spot check.

Detection is not a guarantee. A sampled deployment catches patterns, such as a route that keeps swapping calls to playback, rather than every individual fraudulent call. A route with no alert has not been certified clean. We also publish no detection-rate figure for this check, because none has been measured against a labelled corpus. What an alert gives you is a recorded, transcribed, channel-attributed call: enough to judge the route yourself and to put in front of the vendor when you pull it.

No card is needed to create an account, and there is nothing to install. A route-quality question this post does not answer goes to the team.