The new Gethuman.com website's front page says:

The "gethuman" movement is powered by over one million consumers who are fed up with bad customer service. We demand high quality, prompt, human customer support, by friendly, qualified people who we can understand, and who can handle our call without putting us on hold or transferring us again and again.

And believe it or not, most companies want that too. Early on, the primary motivation for implementing an IVR was to save money and take pressure off the call center agents. That's still true, but increasingly, the IVR and its newer cousin, the speech recognition application are giving consumers quick, convenient service for basic needs -- store location finders, bank balance checks, flight information, package tracking. When the application works, these simple tasks are executed efficiently and no one will ever hear about it, least of all the media. Since when is good service a story?

But when it goes wrong, it's a months-long media odyssey. If you call the US Postal Service to track a package, it's simple. If you call them to ask why your packages never reached their destination, it's a nightmare. Go ahead, mash the keypad; eventually the system will connect you with an agent. And that's what Paul English's story is all about. We've all had trouble like that. Why?

The first person I talked to about speech rec was Peter Leppik of VocaLabs, a company that tests speech apps for the companies that make them. Before Leppik started VocaLabs, he was an industry analyst who tracked call center technology companies. I asked him why he started his firm, and he said:

I made an observation that while I was doing that, there was a lot of really impressive technology going into call centers, but despite that, very few customers thought they were actually getting better service. I realized after doing this for a number of years, after seeing the same pattern happening over and over again, that a big part of the problem was that the ways people had to measure how well a call center was performing from the customer's point of view were very limited.

He noticed that while there were lots of extremely precise measures for how well the call center performed from the enterprise's point of view, there weren't any effective measures for how well call centers serve customers.

VocaLabs does quarterly reports on the call center-based customer service satisfaction level of the mobile phone and financial sectors. When VocaLabs released their preliminary findings for last quarter on their blog (which is excellent, by the way -- interesting call center reading without sales pitches. Check it out.), they pointed out that when T-Mobile got a new IVR system in late 2004,

The IVR picked up about 10% more of their calls, but caller frustration also rose by 10% and 24% of callers now have to call back more than once when a year ago it was 15%. So whatever savings they hoped to realize in agent overhead was lost since the added 10% of IVR handled calls is offset by a 9% increase in multiple calls from the same customer.

So the technology they used to save money did nothing. Nothing except irritate customers. I asked Leppik about it. It seemed like it was a trade between good (agented) service and automation. Leppik:

A lot of companies think of customer service automation as being a trade-off between cost and service level, and it's really not. But when you think about it that way, it becomes a trade-off. We've done some fairly extensive research on the subject, and we've found the most powerful factor in determining the satisfaction level with the company's customer service is just how easy people perceive it is to get to an agent if they need to get to one. But the interesting thing is that making it easier to get to an agent doesn't really push down the automation rates any. The fact is, most people know when they pick up the phone and dial whether they have a problem that can be solved with self service, or whether they need to talk to a person. Creating those barriers really has no effect on the overall automation rate, but it does make people have to call back more than once to get through to an agent. Making it hard to get to an agent pushes the satisfaction level way down.

That's key. Let's look at that again: "Creating those barriers really has no effect on overall automation rate." Trust your callers. They know what they need when they call. Or at least they know when the IVR can't help them. And: "Making it hard to get to an agent pushes the satisfaction level way down."

When I mentioned the Paul English circus to Leppik, he said:

That illustrates very nicely another point that we keep trying to pound over and over about making it hard to reach an agent, which is that people are much smarter than machines. Your customers are going to figure out a way to get to an agent anyway, regardless of the barriers you put up. All you're really doing is increasing the frustration and decreasing the satisfaction, forcing people to call back multiple times.

And here's a lesson for call centers who use simple metrics to measure the success of IVRs: "You're not actually keeping a whole lot of calls away from your agents, although the fact that people are calling back multiple times makes the IVR statistics look good, because every time someone calls and hangs up before they hit an agent, it makes their containment rate look better. It's a misleading statistic at that point." Leppik says it's important to track callers through multiple calls to really find out how well you're helping them.

Next, I asked Leppik about accuracy. I often ask vendors about accuracy in speech recognition, because I keep hearing that it isn't an issue anymore. We have a video lecture from last year's ACCE demo by analyst Art Schoeller called "ASR Architectures and Application Development" (Free, but registration required. Worth looking at for more background on speech systems.) in which Schoeller shows a graph comparing the U.S. GDP to the number of words speech rec engines have in their vocabularies. It's a fascinating graph: the vocabulary size follows GDP growth upwards after 2001. In 2003 the vocabulary size hit 200,000 words. The dictionary I keep on my desk has 120,000 words. Vocabulary isn't an issue anymore.

But implementation and application design are definitely still issues. Leppik said that accuracy hadn't been an issue for about eight years, which was when Art Schoeller's graph started. But it's all in the way you build the application. "There are still applications being built where because of bad design, accuracy becomes an issue. We still see bad implementations of good technology."

Leppik said that it was an issue of making the vocabulary adequate for what you're doing. "We have found that it's possible to build an application that functions very well from the user perspective, even when the speech recognition accuracy is relatively poor. It's a question of paying attention to what the users want to do, and good design, and good error recovery. In a lot of cases, when you've got good error recovery, the callers won't even perceive that there was an error. With a bad design, accuracy can make or break an application, and with good design, you can have relatively poor accuracy and the system still works."