Posted by: Mike_R

Joined: Mar 2019

Posts: 847
so my company (mid size auto/home) just started pushing this "AI claims automation" thing hard. management keeps saying its gonna cut processing time by 40% and reduce fraud.
anyone else dealing with this? we got a pilot running on property claims and so far the system keeps flagging legitimate water damage as "suspicious patterns" because the photos look similar to other claims in the database. adjusters are spending more time overriding the AI than they used to just handling the claim.
also the customer facing chatbot is trash. people are getting pissed when it cant answer basic questions about their policy and just loops them back to "i can help with that! please describe your issue"
curious what other companies are seeing. real talk only please, not the marketing bullshit.

Posted by: claims_gal_23

Joined: Nov 2021

Posts: 312
ugh same. we rolled out something similar last year for auto claims. the AI is decent at the simple rear-end stuff where liability is obvious but anything with multiple parties or injuries and it falls apart.
our fraud model is constantly false positive-ing. had a guy with a legit theft claim get held up for 3 weeks because the AI thought his repair shop history looked "anomalous". he ended up complaining to the DOI.
management still calls it a success though because the numbers look good on the dashboard if you ignore the escalations.

Posted by: underwriter_ed

Joined: Jun 2016

Posts: 1,204
we're using it more on the underwriting side. the predictive models for risk scoring are actually pretty useful, especially for new business. catches some patterns humans miss.
but the black box problem is real. when an agent asks why a quote came back so high the system just says "risk factors identified" and thats it. underwriters end up having to reverse engineer it.
also the data quality is garbage half the time. garbage in garbage out. if the AI is trained on incomplete application data its just amplifying the mess.

Posted by: tech_skeptic

Joined: Feb 2020

Posts: 589
honestly most of what i've seen is just fancy rules engines with better marketing.
true generative AI for things like policy language or complex liability analysis still feels years away from being reliable enough for regulated industries. one wrong denial letter generated by AI and you're looking at a class action or regulatory fine.
the chatbots are the worst part. they make the company look incompetent. i'd rather just talk to a person who at least knows when they dont know something.

Posted by: progressive_rep

Joined: Aug 2018

Posts: 421
we use it pretty heavily and its mixed. the photo estimation tools for auto damage are legitimately good now. saves a ton of time on simple claims.
fraud detection has gotten better too but yeah the false positives are still annoying. you learn which ones to ignore after a while.
biggest issue is the people side. older adjusters hate it and refuse to trust anything the system says. younger ones over-rely on it and miss obvious things. training is a nightmare.

Posted by: Mike_R

Joined: Mar 2019

Posts: 848
@claims_gal_23 yeah the DOI thing is what keeps me up at night. we already had one close call.
@underwriter_ed black box is the biggest problem. compliance keeps asking for explainability and the vendor just says "proprietary model".
anyone know of carriers that actually pulled the plug on these systems after trying them? or is everyone just quietly living with the problems because the board loves the AI slides.

Posted by: old_school_adjuster

Joined: Jan 2014

Posts: 2,156
been doing this 28 years. every few years they come out with some new system thats supposed to replace half of us.
AI is just the latest. it helps with the easy stuff and makes the hard stuff harder because now you have to fight the computer on top of fighting the claim.
customers can smell when theyre talking to a bot and they hate it. some of them start lying more because they figure the computer wont catch it.
give me a good human adjuster with experience any day.

Posted by: data_nerd_42

Joined: Apr 2022

Posts: 167
the models are only as good as the training data and most carriers have terrible historical data. years of inconsistent coding, missing fields, different systems that never talked to each other.
then they feed that mess into a model and act surprised when it makes weird decisions.
also most of these "AI" solutions are just vendor demos that look impressive until you put them on real messy claims data. then reality hits.