Artificial intelligence is moving out of the lab and into the office. Several programs now track customers' buying habits to predict market trends
For Plow & Hearth, a $30-million catalog company that sells yard and garden wares, traditional marketing strategies just weren't working. John Popowski, the marketing manager until 1995, and Pete Rice, the circulation manager, tried the recency, frequency, and monetary (RFM) method, which classifies people according to the date of their most recent purchase, the frequency with which they buy, and the amount they usually spend. They'd jam customers into categories, in an attempt to determine which ones would prove most valuable. But too often those categories were completely random. Rice recalls sessions in which a few managers would sit down and say, "OK. Let's divide the monetary categories into ranges of $20." "The whole process of trying to determine which customers were valuable was so arbitrary," he says.
At the time many of Plow & Hearth's bigger competitors were using models built by statisticians to predict consumer spending. But that alternative was cost-prohibitive to a small company. Even hiring statisticians part-time was out of the question. So when Popowski got a call from Advanced Software Applications (ASA) asking if he wanted his company to be a beta test site for a new product that used artificial intelligence to track buying trends, he jumped at the chance. Soon he was predicting customer behavior down to the smallest purchase.
Artificial intelligence-- neural networks, or neural nets in the jargon--can be glamorous. These computer applications were designed to approach problems the way the human brain does: by trying to recognize the patterns that underlie a complex set of data. Neural nets, like people, can be "trained." They start off with wild guesses, but over time they learn to refine their guesswork until they can pick out even the most subtle patterns--in much the same way that you pick out the face of an old friend at a high school reunion. The technology's power lies in its ability to analyze nearly endless combinations of variables very quickly--more quickly, some say, and with greater accuracy than traditional statistical models.
Law-enforcement agencies use neural nets to pick out patterns in travel data (one-day trips to the Cayman Islands on the first of each month, for example), hoping to pinpoint likely drug dealers. Some hospitals are using them to determine a patient's likelihood of contracting cancer or other diseases. But they are also making their way into marketing and other business arenas. According to Bob Barrie of Global Management Technologies, a small software-development and consulting firm in Atlanta, neural nets work particularly well when it comes to analyzing down-to-the-minute trends. Barrie, whose clients include Walt Disney and several large banks, uses a neural net from BioComp Systems to figure out staffing needs based on customer traffic--a task that requires precision analysis, down to the half hour. Increasingly, even businesses a fraction of the size of many of Barrie's clients are firing up neural nets to help them navigate ever-thicker forests of data and make sense of myriad customer traits and buying habits.
Brain in a Box
Plow & Hearth is a case in point. Based in Madison, Va., today the 80-employee company is definitely on the cutting edge of marketing research. But when it started using ASA's ModelMAX, nearly three years ago, most catalog companies three times its size didn't even know the technology existed.
Unlike some neural-net products, which have to be built up by trained programmers to run useful models, ModelMAX is a polished, ready-to-use marketing tool with a simple Windows interface. Users don't have to know how neural nets work (ModelMAX's $25,000 price tag reflects the product's simplicity); they just have to know how their company works and have access to its information on purchasing behavior. Each time Pete Rice decides to do a catalog mailing, for example, he pulls up the information about customers and their buying habits gathered from previous mailings and purchases. He enters all the key variables, for example, the time of year the customer made his or her first purchase, where (by zip code) the customer lives, and which products the customer buys most often. ModelMAX works through the information to determine which combinations of customer characteristics are most important in predicting customer value. Rice has learned, for instance, that customers whose first purchase is clothing tend to stay customers. That fact was surprising, given that Plow & Hearth sells clothes only as a peripheral product. The information has been extremely helpful in targeting which customers to pursue aggressively.
The hard part, explains Rice, is learning how to sort through and organize the company's information (feed for the neural-net models) so that the program runs easily. He was able to use the product to churn out decent reports within a week, he says. But it took about a year before he could use it to its fullest. His efforts have paid off, though. In the past three years, Plow & Hearth's per-catalog sales have gone up 32%, and its hit rates have risen by 19%.
Neural Nets Built from Scratch
Unlike ModelMAX, most neural-net packages require a lot of set-up time--not to mention big consulting fees if you don't know how to build and feed the nets yourself. That's where the services of businesses like Atlanta-based Marketing Arsenal come in. A consulting and marketing firm with sales of $200,000 a year, Marketing Arsenal uses neural nets to give small auto-body shops the kind of information they need to target customers before insurance agencies get involved. Most people who've had a collision get recommendations for repair shops from their insurance companies, says Marketing Arsenal's founder, Marc Wright. And small shops have to "kiss a lot of butt," he says, to get on the referral lists. With Marketing Arsenal's neural-net services, however, the owners can bypass those lists entirely and appeal directly to the drivers--even before accidents happen.
Marketing Arsenal outsources the model building to various vendors, but it provides all the necessary information itself, largely in the form of a proprietary database of more than 40 million accidents and information gathered from its clients. It then interprets the models' output and writes up the data as profiles its clients can target. A profile might show, for example, that the most likely people to get in an accident are 26-year-old females who own sports cars, have three corporate credit cards, and spend at least $4,000 a month on each card.
So far Marketing Arsenal's test runs have yielded highly accurate results--about 25% more accurate than statistical modeling. And those results come relatively cheap: the company charges just $3,000 for its profile services. If Marketing Arsenal actually pulls together a mailing list of people who fit the profiles and produces a direct-mail piece for an auto-body shop, it adds a $3 charge per name, plus the cost of postage. The enterprise has been so successful that Wright is extending the service to chiropractors. Though it may look as if Wright has taken ambulance chasing to new levels, he's drawn a moral line in the sand. "I refuse to market to lawyers," he says.
Netting a Net
Neural nets aren't for everyone. If your market niche is small and uncomplicated, you probably don't need the power of neural-net modeling; RFM may work fine. And if your historical information is a mess, you'll have trouble feeding the models anything useful (though most neural nets can compensate for some holes in your data). The more complicated your marketing information and the more variables involved--and the better shape your files are in--the more helpful neural nets are.
If you decide that modeling can give you an edge, be patient as you search for the program that's right for you. The technology is just finding its way into the mainstream. Until it does, and until more off-the-shelf packages are developed, pricing is likely to remain erratic--and comparison shopping is going to be difficult. Right now neural-net products cost anywhere from $149 (for just the software) to $25,000 (including consulting services to build the net). According to Global Management's Barrie, hiring a consultant to help you choose a program and apply it is usually a good idea. "Most businesspeople just want solutions," he says. "They don't want to play around."
Finding a good neural-network product for your business takes a lot of research. Here's a list of several neural-net companies, along with their main products, to help you get started. Unless you plan to buy an off-the-shelf model, you'll most likely end up working with a consultant. Expect consulting fees to drive up the price by at least a few thousand dollars.
Advanced Software Applications Pittsburgh
Z Solutions Atlanta
BackPack Neural Network System (starts at $2,000 with three days of training)
Sarah Schafer is a reporter at Inc. Technology.