Smart Systems Without Rocket Science: Q&A With Fair Isaac's James TaylorSmart Systems Without Rocket Science: Q&A With Fair Isaac's James Taylor

"Smart (Enough) Systems" suggests that technologies are available for more intelligent IT systems but that thinking has to change. Can you explain?

The point is to focus on decision making and to think about it as something you can automate. Decisions are different and require a new mindset. For instance, there's something we call "adaptive control," which is the idea that one of the challenges with decisions is that they're not static. How you make decisions evolves over time, yet most people, when they think about information systems, think of maintenance as a bad thing. People constantly think about ways to make better decisions as they learn about their environment and then their environment changes as business conditions and regulations change. You need to have a different attitude toward code that implements decisions because maintenance is not a bad thing in this case; it's inevitable and good because it means you can improve the decision over time.

Two of the technologies you point to are predictive analytics and business rules management systems and you say they're headed mainstream. What's the evidence of that?

If you follow the discussions around business process management (BPM), for instance, there's a lot more talk about business rules as a complementary peer technology rather than as just part of the process engine. There's a lot more talk about predictive analytics as well, and it's moving away from just "predictive reporting" to actually embedding the prediction into applications to drive better decisions.

We're still at the stage where most organizations are adopting these technologies because they have one particular problem that really justifies the investment, but once it's adopted it quickly spreads. One of the biggest growth areas has been in insurance, which in the last couple of years has adopted rules and analytics in underwriting, claims and marketing, and much smaller insurance companies are investing because the big players at the top of the market have made life so difficult. For the smaller firms it's a make-or-break proposition. If they can't get their underwriting process to be competitive, it almost doesn't matter what else they do.

Part of the growth question is who's driving adoption at the top of the industry? Entertainment and leisure companies are investing because they're scared about what Harrahs is doing with the technology. Large telcos are embracing rules for CRM applications, so the smaller companies are following along. It requires a certain amount of fear to drive adoption, quite frankly, and we're seeing that fear as big players get onboard.

Can you cite examples of the predictive analytics being embedded into applications?

If you swipe your credit card, there's a high probability that a product Fair Isaac makes called Falcon Fraud Manager will check for fraud. It does that by running an analytic model that comes up with a predictive score. The score describes the risk of the transaction based on your past behavior, the behavior of other people like you, known fraud patterns and a bunch of other factors. Your bank sets up a set of rules to decide how to react to that score based on your customer profile and that type of transaction. Should they reject it, pay it but follow up with you, or should they take some other course? The prediction itself is not an action; the rules determine how to react to the prediction.

We had a discussion with a telco customer recently to discuss best next actions — what's the best thing to do next with a particular customer? Typically that involves some combination of predictions about how much of a retention risk they are, how profitable they are, how much of a credit risk they are and so on, and then rules can be used to decide what to do next. Instead of just presenting a bunch of information to a call center rep, which they then have to interpret very quickly and then apply whatever the current policies might be, why not figure out what the best treatment is and present the top two or three choices?

That sounds like the kind of thing a CRM system could handle based on customer segments or transaction parameters.

It's something that's often attempted, but most implementations lack analytic depth —particularly predictive analytics — and most of them don't make it as easy as a rules environment to change aspects of the decision making. Many solutions don't even separate out that decision, so it's hard to make it consistently across all channels. By abstracting the decision, you can regard it as a separate opportunity for improvement and management. When someone uses an ATM or interacts with your IVR system or calls your call center, or goes to your Web site or walks into a branch, you don't want the same decision being made differently across different systems. It's also important to have the rules involved controlled by the business and executable in real time. You often see some elements of these things, but the key thing is that other solutions usually don't regard decisions as reusable corporate assets.

Is there more to it, technologywise, than analytics and business rules?

I think there's a role for optimization and simulation as well. If you're in an environment in which all the decisions are made by people, you can start by modeling those decisions rather than having them stuck in people's heads. Next, maybe you can automate some of those decisions with business rules. While you're at it, maybe you could make better decisions or better rules by interpreting the available data with prediction or simulation and optimization. Gradually you get to the point where you have a more formal, modeled and mathematical view of how a decision is made and you can improve it over time.

So you would suggest people are ready to move in this direction?

We think people underestimate their ability to make their systems more usefully intelligent with proven technology… Plenty of vendors have these technologies and they're not rocket science, but they've been cast into these very narrow segments. Any credit card issuer will tell you that deciding whether a transaction is fraudulent or not is a crucial decision, but the need for the technology is less obvious in other situations.

If you're creating a marketing campaign, for example, most people would say, "I have to decide what the content of the campaign is going to be and who I'm going to target." When I look at that scenario, I don't think of it as just two decisions; I think of it as sending a particular offer to a particular customer, so it's really 100,000 decisions.

What if you gave yourself the opportunity to make each one of those 100,000 decisions differently? Would you send everyone the same letter? Maybe you wouldn't. Would they all get the same IVR options? Maybe they wouldn't. Some of these things are more practical now than they used to be, but people just aren't used to thinking about the possibilities.