Sam the Centipede wrote: ↑
Sat Mar 21, 2020 3:57 am
I don't think an AI (artificial intelligence) correspondence-handling program needs a multilayer convolutional neutral network to combat the AI (abnormal idiocy) of a sovcit disputant and issue the PUNOWCFYA (pay up now or we're coming for your ass) response. As soon as it spots
Notice to Principals is Notice to Agents; Notice to Agents is Notice to Principals.
it should be game over. That's all the logic it needs.
As you know, SovCits are not the world's best spellers (an arch euphemism for "they're a bunch of drooling illiterates"). I've seen quite a few different permutations of spelling in your example sentence, despite the fact that most people copy-pasta it from other people on the internet. So you'd definitely need to have some fuzzy logic to correct for the illiteracy factor alone.
It's a maxim among people who build high-reliability transaction processing systems that as systems get more reliable, the number of errors drops exponentially, but the cost to fix one of the remaining errors goes up exponentially. In other words, because you've automated all the easy problems, the only problems left are by definition much harder to fix. When I started in the money management business in the 1990s, settling trades was an intensely manual process, because probably 10% of trades came back with status "DK" ("Don't Know"), meaning that the counter party had no record of the trade. With a phone call or two, the problem would usually be resolved pretty straightforwardly, and it was just assumed that the system was screwed up and this was just what you had to do. The problem was that there were many thousands of people involved in repetitive aspects of trade clearing, which was an immense cost.
While I'm not anywhere near trade clearing these days, it's probably 99.9999% automated without problems. But chasing down the 1-in-a-million trade that goes wrong today is a much bigger problem because you need to understand why the settlement failed, and you need to validate that it wasn't a bug in the system. After all, if it's a system bug that calls into question the validity of settlement for the other 999,999 trades, and you have a big problem when you have a highly automated system that you can't trust. That's an immense problem but it's worth solving because it takes us from $1.30 per share in brokerage commissions 50 years ago (around $6.50 adjusted for inflation today), to approximately zero today. That greatly increases the amount of money that stays in the investor's pocket -- you had to pay that immense commission both when you bought and when you sold.
The point is that as you cut the number of exceptions, you have to continue to invest in automating the handling of those exceptions. And as SovCit arglebargle metastasizes when all those grumpy old grifters put their own stamp on the drivel, it's going to take ever more cleverness to parse all the nonsense. You have to do a good job because it's at least as important to avoid a false positive as it is to correctly ignore nonsense. You definitely don't want to dismiss as a crank a legit customer who who's mad and who thinks trying to "write like a lawyer" will win his case, especially in a regulated industry such as banking, insurance or brokerage.
The overall trend is for these systems to become more responsive to customers in a native format. Instead of making customers fill out a five-page complaint form with lots of irrelevant boxes and not enough room for narrative text, the trend is to make the interaction as close to natural language as possible, whether the issue is submitted in speech-to-text from a mobile device or whether it's submitted by e-mail.
A good example of this is how Target's returns process works. At Christmas, my job is to cook homemade lasagna for Christmas dinner (I have zero Italian blood in me but for pretty random reasons, this became a holiday tradition a few years ago). Target was the only source of 3" Pyrex baking dishes at a reasonable price (that's 7.5 cm, for Suranis and for those of you in the quaint, rural "Canada" region of upstate New York, located somewhere north of Buffalo) -- most dishes are only 2" (5 cm) and the lasagna bubbles over and makes a mess of the oven. I ordered four and they were poorly packed, so two were shattered. I got on the app, and the flow through the return process was extraordinarily well designed. I was able to get replacements sent in less than 60 seconds from logging on to order confirmation, including a nicely worded (but automated) apology and the instruction to just recycle the broken glass; no need to send it back or to send photos of the wreckage. The flow of the process led me to believe that they have different return processes for each different merchandise sub-category, so they asked questions appropriate to broken glassware and not to someone trying to return pants that didn't fit. Net result of this incredibly natural exception process is a very happy customer.
I do understand that your comment was tongue-in-cheek, by the way, but I wanted to give some perspective on where these systems are going and why.