IBM Watson to replace salespeople and cold-calling telemarketers?

After conquering Jeopardy! and making inroads into the diagnosis of medical maladies, IBM's next application for Watson is the wholescale revitalization of two very lucrative markets: sales and customer support. Think about it: how many times have you asked a salesperson a question about a product, only to have a blank smile or glib response fired back? Even good salespeople can be stumped by tricky and application-specific questions. In both cases, a slow or weak response will result in the salesperson losing the sale. Now imagine if IBM Watson was there to help.

A question answering machine like Watson could be applied to almost every level of commerce. You could have a Watson on the floor at Best Buy, or the support departments of large companies could keep a pet Watson for answering trickier questions. Telemarketers, which are famously bad at handling off-script questions, could be saved by Watson. In a beautiful, self-fulfilling prophetic twist, the first application of Watson in a sales and support role will be internally at IBM, to help IBM sell other Watsons to other companies.

IBM Watson, which is powered by Big Blue's DeepQA software, is fundamentally a huge search engine that can be used to answer questions. You fill it with gigabytes or terabytes of raw data -- such as general knowledge for Jeopardy, symptoms for medical diagnosis, or product specifications for sales and support -- and DeepQA turns it into useful, actionable facts; it performs analytics(Opens in a new window), in other words. As you well know, though, search engines like Google have been successfully doing this for a long time. DeepQA, however, has one other trick up its sleeve: it's also exceptionally good at understanding natural language.

Natural language is everyday speech. It generally adheres to some basic syntactical rules, but not always. Now, it's not hard for computers to understand structured language ("go north", "open file"), or to literally translate what we say, but to actually understand natural language is hard. Not only must the computer understand each of the words individually, but it must also take into account the context. Is the question rhetorical? Ironic? Referential? DeepQA, as witnessed by its runaway victory in Jeopardy(Opens in a new window), seems to have natural language processing down to a fine art.

IBM makes it clear that Watson, at least in its current incarnation, is only designed to augment and assist humans -- but in reality, Watson could almost certainly operate without human intervention. Once you fill DeepQA up, it will answer questions all day long. You might be a little terrified the first time you pick up your phone to hear dulcet but unmistakably-computerized tones of Watson saying "Hello, can I interest you in cheap home insurance?", but when you realize that you're dealing with a computer that can answer all of your questions, and in a neutral way without emotional blackmail, you'll probably warm to the idea. The appeal of an entirely-computerized sales team must surely be an attractive proposition for many big companies, too.

The other application is e-commerce: instead of popping up a chat window with a human, you could use DeepQA instead. You could even hook Watson into Twitter: customers can tweet their problems, and prospective customers can tweet their questions about your products -- either way, DeepQA will merrily respond to the questions.

The only real problem, of course, is that DeepQA -- or rather the hardware behind it -- is incredibly expensive at the moment, both to purchase and to run. Despite its on-stage presence being fairly diminutive, IBM Watson(Opens in a new window) actually consists of no less than 10 server racks filled with Power 750 servers, with a total of 2,880 processor cores and 15 terabytes of RAM (see image). The exact cost of Watson isn't known, but considering a 32-core Power 750 server costs around $350,000, and Watson used 90 of these servers, the cost to install your own Watson would be around $32 million.