Posted by: data_guy_19
Joined: May 2023
Posts: 156
most of these AI underwriting and generative AI tools are only as good as the training data. and most carriers historical data is a disaster - inconsistent coding, missing fields, different systems that never talked.
garbage in, garbage out. then everyone acts surprised when the risk scoring or fraud detection makes weird decisions.
the demos always look perfect. real messy claims data is where it falls apart.
Posted by: old_timer_adjuster
Joined: Feb 2013
Posts: 2,401
28 years in claims. every new tool is supposed to change everything.
AI claims automation and fraud detection help with the easy stuff. the complicated claims still need experienced people. now we just spend extra time fighting the computer recommendations.
customers can tell when an AI chatbot is talking to them and they get mad faster. some start stretching the truth because they think the system wont catch it.
give me a sharp human adjuster over another “smart” model any day.
Posted by: Sarah_K
Joined: Jan 2020
Posts: 613
@UW_Mike yeah the explainability problem is the same on our side. compliance and legal keep asking for audit trails and the vendor just says proprietary.
@claims_lead_88 the generative AI inventing details is my biggest fear. one wrong sentence in a denial letter and were in trouble.
anyone seeing carriers actually slow down or pause these generative AI and AI underwriting projects? or is everyone just pushing forward because the board wants AI in every slide deck?
Posted by: fraud_analyst_22
Joined: Nov 2019
Posts: 534
fraud detection side here. the newer AI models using graph analysis and behavioural signals are better than the old rules engines. we catch more staged accidents and provider billing patterns.
but false positives are still killing us. simple auto insurance claims get held for review because the repair shop or the claimant “looks similar” to previous fraud cases. then the customer complains and we look bad.
management only looks at the “fraud dollars prevented” number and ignores the customer experience cost.
Posted by: customer_exp_gal
Joined: Jul 2022
Posts: 278
can we talk about the AI chatbots for a second?
ours is supposed to handle FNOL and basic policy questions. customers hate it. they type “i need to file a claim” and it asks the same three questions over and over or dumps them into a form. half of them just hang up and call.
conversational AI is still not ready for regulated insurance. people want a human when something goes wrong with their claim.
Posted by: UW_Mike
Joined: Mar 2017
Posts: 890
AI underwriting is where the real pressure is. the predictive models and risk scoring tools are decent for high volume personal lines (car insurance and home insurance) but commercial is still a mess.
biggest issue is the black box. agents and brokers want to know why a quote came back high and all we can say is “the model identified elevated risk factors”. not good enough when the client pushes back.
we also tried generative AI for policy wording reviews. it hallucinated coverage that wasnt there twice. we shut that pilot down fast.
Posted by: claims_lead_88
Joined: Sep 2018
Posts: 1,043
weve been on AI claims automation for about 18 months. photo estimation on auto damage is solid now, especially for the easy fender benders. saves a ton of time.
fraud detection is mixed. it catches some organised rings that humans would miss but it also flags way too many legitimate home insurance water claims. adjusters are burned out overriding the system.
the generative AI part for writing claims summaries is hit or miss. sometimes it sounds professional, other times it invents details that arent in the file. that scares me for compliance.
Posted by: Sarah_K
Joined: Jan 2020
Posts: 612
ok so our carrier just announced were rolling out generative AI for underwriting and claims notes. management is hyped about faster risk scoring and automated decisioning.
anyone else dealing with this right now? our underwriters are already complaining the AI underwriting recommendations keep changing when the same application is run twice. and the explainability is still basically “trust the model”.
also hearing a lot about AI claims automation and fraud detection getting better but i still see way too many false positives on simple auto insurance claims.
whats the real experience out there? no vendor talk please.
