Rule based underwriting

 When I was in my mid-thirties I grew tired of all the fuss and stress associated with my high-prestige work downtown SF and looked into redefining myself.  I already knew Basic and Cobol and Pascal, although self-taught.  The rule-based AI systems were the hype of the day, and everyone bought into it, including  myself, so I went to school for nine weeks at a startup in that stuff and learned their system and went to KC, the home office, to serve as liaison between underwriting and ISD.  At the time, it was all the rage and I got asked to address a lot of local groups and was always, after my presentation, asked about taking jobs.  Well those things didn't.  Underwriters who knew what was going on had to enter the data.  They turned into neat affairs advising on possible rules being bent and keeping track of things and ordinary pricing.  Still, jobs is the first thing one gets when one is doing automation work.

Yes—and that last point is the interesting part. Automation gets interpreted as a jobs story almost immediately, even when the actual effect is much more mundane.

What you describe sounds remarkably like an early version of what we now call “AI transformation.” The technology was impressive enough to attract attention, and you were suddenly the fellow who understood both the business side and the computer side. That made you a natural evangelist, liaison, and speaker. But once the system was actually put into operation, the fundamental constraint appeared: the machine couldn't know what the underwriter knew unless the underwriter put it in.

And that produced a rather different result from the futuristic vision. Instead of eliminating underwriters, the system made their work more orderly. It could flag inconsistencies, identify situations where a rule might be bent, keep track of the applicable rules and pricing, and generally make the process less dependent on somebody remembering which drawer contained which piece of institutional knowledge.

There's a funny parallel with today's generative AI. The early rule-based systems had a particularly severe version of the problem because their knowledge had to be explicitly encoded. Today's systems can infer much more, but they have their own bottleneck: someone still has to supply the context, judge whether the answer makes sense, and take responsibility for the result. The computer can make the production of an answer astonishingly cheap without making the underlying judgment unnecessary.

And I can see why you found the whole episode attractive in your mid-thirties. You weren't merely looking for another job; you were trying to get out from under the peculiar combination of prestige, pressure, and downtown-San-Francisco professional life. The technology offered an apparently clean escape hatch: learn something new, become the person who understood it, and move sideways into a completely different sort of work.

Instead, you discovered something that has recurred throughout the history of automation: the machine changes the job before it eliminates the job. And sometimes the changed job is actually rather better.

I was tired of the pressure of loss and expense ratios as well as production pressure and the need to buy a new silk tie and get a haircut and manicure every week.

That makes the motivation much clearer. It wasn't simply, “I want to get out of insurance.” You were tired of the whole performance that accompanied the job.

Loss and expense ratios are bad enough because they turn every underwriting decision into a number that eventually comes back to haunt you. Then there is production pressure: write more business, keep the agents happy, meet the targets. And on top of that, you had the curious obligation to maintain the physical appearance of someone who was supposed to be important—new silk tie, haircut, manicure, week after week. At some point the silk ties themselves must have begun to seem like part of the overhead.

There is something especially revealing about the manicure. A person can tolerate quite a lot of professional pressure if it seems intrinsically necessary. But when you're sitting there thinking, I have to have my fingernails professionally maintained so that I can continue underwriting insurance, the artificiality of the whole enterprise becomes hard to miss.

So the rule-based AI venture offered something rather appealing: you could be valued for knowing something rather than for looking like somebody who ought to be important. And you could walk into a room and talk about a new technology instead of explaining why this quarter's ratios weren't good enough.

Ironically, you ended up discovering that automation wasn't going to produce the mass unemployment everyone was predicting. It mostly produced tidier offices and somewhat better-informed people doing the same underlying work.

But at least you got out of buying silk ties every week. That may have been the real technological breakthrough.

Since the data was now on out mainframe, it did reduce the number of file clerks and the constant search for a given file.

Exactly. And that is a much more concrete form of automation than the grandiose claims people were making about AI.

The file itself was the thing being automated, in a sense. Once the relevant underwriting data was on the mainframe, you didn't need a clerk to maintain a physical file, retrieve it from wherever it had been filed, carry it around, and put it back. You also eliminated the little delays and uncertainties that went with physical records: Who has the file? Was it returned? Did somebody misfile it? Is it in underwriting or claims?

So the system could genuinely reduce headcount—but mostly among the people whose principal function was moving information around. It didn't eliminate the person who understood the risk.

That distinction seems to get lost whenever people discuss automation. There are really two very different things: automating the handling of information and automating the judgment applied to information. Your system was quite good at the first and rather limited at the second.

And there's a nice historical progression there. First you had clerks maintaining paper files; then computerized records made the clerks less necessary; then rule-based systems helped organize the decisions made from those records. Today's AI is attempting to push much further into the judgment itself.

But the old system had already demonstrated something important: you don't have to eliminate the profession to eliminate a lot of its surrounding machinery.