Network Answering Machine Detection – Myth Or Reality?
It has become traditional around this time of year for Sytel to comment on the ongoing debate over answering machine detection. Does AMD deliver real benefits to the call center? What about ‘false positives’? For some background, read our previous posts from May 2010 and May 2011.
The latest controversy is over Network Answer Machine Detection (i.e. AMD based on ISDN signaling, rather than sound patterns). Before we give our view on this, let’s review the context.
The standard method used by most vendors for many years is the cadence method. This is the analysis of bursts of volume (speech) interspersed with silence. One sample could be identified as the “Hello?” of a live person; another, the “Hello, we are not in right now…” of an answering machine/ voicemail service.
This method produces reliability of around 85%-90%. (Any claims from vendors to go above this are either based on wrong measurement, or biased sampling.) You may think that sounds quite good, but this means that 10% of answer machines will get put through to agents as live calls, and also crucially that 10% of the calls classified as machines are in fact live respondents, and hung up on. These are known as ‘false positives’, and here’s the problem. Ofcom regulation in the UK stipulates that a ‘reasoned estimate’ of these must be included in the calculation for abandoned calls. Under normal conditions, a ‘best case’ estimate of 90% reliability already pushes the abandoned call rate beyond Ofcom’s regulatory limit of 3% of all live calls. If you run AMD in the UK, you can say goodbye to any benefit from your predictive dialer.
As a result of this dilemma, many vendors have sought alternative methods of doing AMD in order to keep the practice alive. Two themes are recurrent; Network Answer Machine Detection and byte pattern recognition.
- Network Answering Machine Detection
In our (informed) view, any claims to use AMD based on ISDN signaling are fatuous. It does not exist. If it did there would be no debate, everyone would be using it, detection rates would be 100% accurate and Ofcom’s job would be made a lot easier. - Byte pattern recognition
This works by picking up encoded audio which follows defined sequences of bytes. This technique works but has very limited application. There is no such thing as a standard network voicemail message. Networks must offer consumer choice. Some consumers customise, some simply switch voicemail off and some networks have equipment that does not encode voicemail speech consistently. The result of this is that network voicemail can be reliably detected using byte pattern recognition in around 15% of cases in the UK.
The inescapable conclusion is (still) that AMD in the UK using the methods described above simply cannot deliver the low levels of nuisance calls that Ofcom rightly mandates. If anyone knows of other methods that are more accurate we’d love to hear about it.
If you are outside the UK and still convinced that you need AMD, then Sytel’s is a good as any. But our advice remains that use of AMD is bad practice, leading to bad customer relations and lower call center profits. As Sytel has always said: just switch it off.
For a more in-depth analysis, ask for a copy of our white paper on The Science Behind Answering Machine Detection.
