Artificial Intelligence in Home Insurance
AXA holds a large number of business partners worldwide, supplying both the insurance policies and the claims management. The mission: enable full automation of the coverage calculation, and provide automatic and vocalized responses in interaction with personal assistants of B2B2C insured end-customers.
More than 6000 insurance contracts have been analysed in order to see if we could determine some normalized patterns. Bottom line: not for all insurance contracts, no, but many could be rewritten in order to fit a framework.
To demonstrate the opportunity, a few IARD frameworks were elaborated based on a set of relevant Home Insurance Policies as well as Motor Insurance Policies. This POC involving personal assistants was carried out on Home Insurance only.
The technical difficulty to connect to a personal assistant such as Alexa or Google was rapidly overcome. Speech-to-Text for input & Text-to-Speech for output were sufficiently efficient to provide relevant data for our Coverage Rules Engine. The real knot of the project was, without the shadow of a doubt, the Coverage Rules Engine (CRE).
There were no software capable of providing a reliable technical foundation for our CRE. So we built it ourselves, from scratch. Thousands of General Exclusions and Specific Terms & Conditions were parsed and normalized, using Elastic Search technologies. A set of Policy Frameworks emerged, however we could not apply this framework to all policies. So we selected 15 home insurance contracts, which were very close to being executable.
These frameworks included fire clauses, water clauses, theft, storms, roofing problems, toilet issues, rats & other nuisances, and many other usual Home Insurance clauses. For each clause, contracts would differ in the details, making it difficult to have an executable numerus clausus. But, we did produce a framework, and coded a decision tree.
Extracting key words from the input speech, as well as meta-data such as geolocation when available, we could flow these words into a Rules-based engine of our making, which could determine as an output if an exclusion was reached or not, and if any limitation would apply to the claim. Is the end-customer covered ? Can he/she be relocated to a close-by hotel ? For how long ? Which kind of hotel ? What amount of money could be provided to fix the problem ? What restrictions would apply ?
The conclusion of this POC was that the insurance policies currently used by Insurance companies have tailored so much to their partners' specific needs that it becomes impossible to use a deterministic approach. Also, a human intervention, with kind empathy and benevolence, plays an important role in the relationship with the end-customer. This left us with two alternatives:
1. Wait until LLMs and Vocalization have reached a better maturity and create a whole new approach to Coverage Calculation and conversation with the end-customers.
2. Write new Insurance Policies in terms better executable, in order to fully automate the Coverage Calculation, and, subsequently, the Claim Management & Payment.
One thing is sure: be it by humans or by AI, Coverage Calculation is never a crystal clear process if contracts are too detailed. Interpretation remains paramount, and can be an important source of frustration.