Artificial evolution: the fourth generation of artificially intelligent underwriting systems produces an instant policy.

Key Points

* Fourth-generation artificially intelligent underwriting requires only a name and one identifying piece of information from an agent or an applicant.

* Virtually all underwriting and policy information is developed from independent sources of information, and it uses a sophisticated black box to do 90%+ of all underwriting.

* This type of system can work for nearly all personal lines and small commercial insurance and even individual practitioner malpractice.

Imagine a young man sitting in a coffee shop with his laptop computer, sipping a $5 mocha-deluxe. Since the coffee shop has provided wireless Internet access, he is surfing the Internet and decides to get a quote on his auto insurance.

He goes to the Web page for the Internet Insurance Co., which asks him for his name and address. Alternatively, it may require a driver's license number instead of an address. All that is needed is some way to identify him, John Smith, from all other John Smiths.

He gives the information, and in a few seconds he receives a quote on his two automobiles. He decides to buy and enters his credit card number. A policy can be printed out immediately if he wants. This process is the result of interactive artificially intelligent underwriting systems, and while no company can do this fight now, it is closer to reality than one might think.

Behind the Scenes

What happens behind the scenes in such a system is possible only with advanced computer logic programs and the nearly universal access to information provided by the Internet. The first step the artificial intelligence system takes is to verify the applicant's identity using available name and address databases, such as Intellius.com. Then it accesses the state Bureau of Motor Vehicles to get vehicle information on the applicant's cars and queries the same BMV for information on any violations on the applicant's driver's license.

It then goes to several other databases, such as the Comprehensive Loss Underwriting Exchange, provided by Choicepoint Inc., which gives it information on any losses the applicant has reported to any insurance company. It also checks the applicant's credit score to see how well he or she pays bills. It even will check the applicant's criminal record if there is one.

Up to now, the process has been very similar to what insurance companies do today except that, currently, underwriters and agents often gather the information manually. Here is where things start to get interesting, though, as, in addition to this information, the artificial intelligence system will access a lot of other, more obscure, databases to pick up tidbits of information about the applicant, such as:

* Employment,

* Age,

* Marital status (plus all the same information on the applicant's spouse or significant other),

* How far the applicant lives from work,

* How many times the applicant has moved,

* Whether the applicant owns or rents,

* Hobbies, and

* Shopping habits (including whether the applicant has bought any radar detectors).

Data mining of this sort will fill in many of the old insurance application questions, such as annual miles driven, but also much, much more. This information is all more available than most people might think, much of it volunteered by consumers as they fill out product registrations, warranties and numerous other applications.

Once all the information is gathered, all of the traditional insurance application questions have been answered more credibly than if they were filled in by the applicant. For example, most people don't remember exactly when their last accident or violation occurred, unless it was in the past year. This process is called "independent indexing," as it is independent of input from the potential insured.

The "Black Box"

At this point the underwriting process could probably involve a human underwriter, but that would delay the process, be more expensive, and potentially involve a bad underwriting decision if something is missed. Instead, the Internet Insurance Co. will have constructed a "black box." The black box is overtly an algorithm that weighs various pieces of information about the applicant, and decides whether he or she should be accepted by the company and/or into which rate tier.

The construction of a black box is a relatively complex statistical and actuarial exercise, but can be well worth the effort. First the company will pick at least a dozen or more pieces of relevant data that can be identified for a high percentage of their current book of insureds. For auto insurance, these data elements could include accidents, violations, miles driven, credit score, age of vehicles, type of vehicle, occupation, hobbies, whether a property owner or renter, number of jobs, etc. Relative values in each of these elements will then be statistically correlated to the actual loss experience of the current book of insureds. For example, a correlation to loss ratio will be performed using one violation, two violations, three violations, etc. Removing policyholders with one violation will change the loss ratio by a specific number of percentage points. Removing policyholders with two violations will lower the loss ratio by a different specific number of points.

This kind of activity will be performed with each of the selected elements of data, but the real benefit will be derived when all or some groups of data elements are combined, and the effect on loss ratio is measured in the current book of insureds. Often these combinations are much more predictive of lower loss ratios than going through all the individual data elements as a human underwriter might do.

As might be expected, though, performing all these combinations and correlations requires a significant amount of computer power and some actuarial and underwriting instinct on the part of the people putting the black box together. Often these efforts are so complex they are led by outside consultants, for example, from the major accounting firms. For that reason, the current carriers who actually use black boxes are very protective of their formula. Using a black box approach to select new and renewal business can improve loss ratios significantly, with 20%+ improvement commonplace. Currently, only a very few larger companies are reported to be developing or using early forms of a black box system. Allstate and Progressive are identified as practitioners in the personal lines arena, and Safeco with small commercial business.

As long as the risk-of-loss characteristics are relatively uniform, this kind of underwriting system can work for nearly all personal lines and small commercial insurance, and even individual practitioner malpractice. It promises much more profitable results for companies and lower costs for consumers. Even those consumers who will be moved to a higher rate tier may eventually benefit as expenses are wrung out of the system.

Generational Development

Once the black box is designed, the main task of the insurance company is to allow easy access to the consumer. Using the Internet as a distribution system is probably the easiest way to accomplish this, although the limited information required from the potential insured would make a telephone or personal contact just as easy, but more costly, due to staffing expense. The system described in the coffee-shop scenario is totally interactive and artificially intelligent and can be classified as a true fourth generation automation system.

To explain the capabilities of a true fourth-generation automation system, and how it fits with what is available today, the following history reviews the progression of automation systems as they have been designed to use the Internet.

First Generation Systems: These were some of the first electronic interfaces used by insurance companies. Many smaller and specialty companies are still using these types of electronic systems right now. They allow rating of policies online, either over the Internet or on proprietary company networks. Often, these systems will have some underwriting guides as "read only" files. They will have little or no upload and/or download capabilities, so once the policy is rated, it still must have an application completed and sent in for underwriting approval. This means there is still a lot of paper handling being performed by highly trained underwriters and agents. This classic model is short-term easy, but long-term very expensive.

Second Generation Systems: These systems are currently in use by most insurance companies, at least for personal lines. They allow rating automation, capture policy issuance information and incorporate a few underwriting filters. They use the Internet or proprietary company networks for transmission among insured (sometimes), agent and company. They usually allow upload of the information to the company, but not download of the policy, which is usually mailed. As with first generation systems, they still have a very high degree of human involvement in underwriting and processing the policies. Very few of these systems allow for sending the same information to multiple companies. Known as single-entry multiple-company interface, this is the Holy Grail for most agents and probably insureds, but may not be as universally accepted by insurance companies, as they might perceive they would be constantly "shopped." On the other hand, several larger insurance agents have actually built their own systems to allow them to enter information only once. Their system then converts the information to what several of their companies need so they can get multiple quotes. These have proven to be expensive undertakings so far.

Third Generation Systems: Very few insurance companies are currently using systems this advanced. The system integrates information entry with processing of the policy so the agent (or applicant) inputs information, which is translated into a format for immediate policy issuance. This generation of systems usually involves the use of at least a rudimentary black box, so only those applicants who fall outside the parameters of the system are actually touched by human underwriters. The black box could entail automatic credit scoring combined with CLUE information and standard application responses to such questions as "have you been cancelled or nonrenewed by any insurance company?" The underwriting logic used is still mainly dependent on complete insurance-application type entries from the agent or the prospective insured. Upload of the information is standard but download is available only on a case-by-case basis, with policies still usually mailed from a central location. SEMCI is still a goal for most. Since fewer policies are actually reviewed by human staff, expenses are lower, allowing for lower rates.

Fourth Generation Systems: How this type of system works has already been described in the scenario above. It requires the input of very little information from an agent or an applicant. Virtually all underwriting and policy information is developed from independent sources of information, and it uses a sophisticated black box to do 90%+ of all underwriting. It is tied instantaneously to a policy issuance system so upload of information and download of policies are available. Because the amount of information required is so small, SEMCI is no longer a major issue. Internet access will be standard for the companies with this capability. (A proprietary network would mean going to a dedicated computer or workstation to enter the information.) This type of system drastically reduces the underwriting and agency expenses, which will result in much lower rates.

Market Expectations

While there are benefits of upgrading to artificially intelligent underwriting and the advanced systems discussed above, recent disclosures of confidential information from the information consolidation companies certainly brings up some serious privacy issues. This quick-and-easy service vs. privacy equation will probably be debated for years to come. Such systems may not even be possible under some state laws, but that may not always be the case.

Additionally, systems such as these must be put into the macro-economic insurance environment. Development of these systems is expensive short term, and events such as the Sept. 11 terrorists attacks and the medical-malpractice crisis resulted in a hard market and back-to-basics underwriting, which work against implementation of innovation at the company level.

The overall insurance marketplace, however, always feels pressure from competition and consumerism to drive rates downward and squeeze expenses out of the system. The days of property/casualty companies having expense ratios of 25, 30 or even higher are numbered if even one company can find a better way. Using the most advanced systems, expense ratios of 5 to 10 can be achieved and will eventually be demanded by the market/competition. A few Internet-based and/or highly automated insurance companies such as the "Internet Insurance Co." will be able to achieve those numbers.

Contributor Paul B. Nielander is a principal with Malecki Deimling Nielander &Associates LLC in Erlanger, Ky.