Written by an advisory that runs one
AI in Insurance in India, From Someone Actually Running One
Most writing on this subject is by firms selling chatbots or consultancies selling reports. This is from a twenty-two-year-old insurance advisory in Udaipur that put an assistant on its own website and has been reading what people ask it. The numbers are small and unglamorous, which is rather the point.
What we would tell another broker
- The corpus is the product; the model is a commodity. Every wrong answer we have had came from our own published material being wrong or missing, not from the model inventing things.
- It is the most used thing on the site by a wide margin, several times the WhatsApp clicks and far ahead of every form. Nobody predicted that, and we only found out because we finally measured it.
- We were not logging the questions. For months the single most used feature was also the only one we could not see. That is the mistake worth avoiding.
- The tone has to change when the visitor is in trouble. Offering a quote to someone whose godown burned down last night loses a client permanently.
- Draw the line at explaining versus advising. What a clause means is safe. Whether your sum insured is right is not, because it needs someone to look at the building.
What people actually ask
Not what we expected. Very little about price, which is what every brochure assumes people care about. Mostly comprehension: what does this term mean, is this thing covered, what do I do now. The single most common shape of question is a person describing their own situation and asking whether the policy they already hold responds to it.
That tells you something about the market rather than about the technology. People are not under-served on quotes. They are under-served on being told plainly what they have already bought.
The mistake that taught us most
For several months the assistant told people that insurance premiums in India attract GST. That stopped being true for individual life and health policies on 22 September 2025, when the rate went from 18 per cent to nil. Group policies still carry 18 per cent, so the honest answer had become two answers, and ours was still giving one.
The model was not hallucinating. It was faithfully repeating our own material, which nobody had updated. An assistant is a mirror of what you publish, including the parts that have quietly gone stale, and it will repeat your errors more consistently and more confidently than any human employee would. That is an argument for auditing what you publish, not for avoiding assistants.
Where it must hand over
Ours will not quote a premium, bind cover, or say a claim will be paid. Not because of caution for its own sake, but because it cannot see the policy. Any tool that answers those questions without reading the wording is guessing at the customer's expense, and the guess will be confident.
We also made it behave differently when someone reports a loss that has already happened. It stops offering quotes entirely, gives the two things that matter in the first hours, and puts a person in front of them. We enforce that in code as well as in the prompt, because asking a model nicely worked about two times in three, and the third time is the one that matters.
What is overstated
AI underwriting, in the Indian retail and SME market, is mostly still a slide in a deck. Pricing here was de-tariffed in 2007 and insurers already rate the same risk differently by wide margins; the constraint is appetite and reinsurance treaty, not computation.
Claims automation is real at the document-handling end and thin at the judgement end. A model can read a bill. It cannot stand in a burnt godown and decide whether the stock register is credible, and that is where claims are actually won and lost.
What is understated
Language. India's insurance material is overwhelmingly in English and a great many of the people buying it think, trade and argue in Hindi. An assistant that answers in the language the question was asked in removes a barrier that no amount of English content ever will, and almost nobody in Indian broking is doing it.
The other is patience. A person will ask an assistant the same question three times, and ask the embarrassing one they would never ask a human twice. That is worth more than any efficiency claim.
Frequently asked questions
How is AI actually being used in Indian insurance today?
Three ways that are real, and a great deal of talk that is not. Insurers use it for document handling and fraud triage. Distributors use it for lead qualifying. And a small number of advisors, including us, use it to answer the questions people are too embarrassed or too busy to ask a human twice. The underwriting revolution is mostly still a slide.
Does an AI assistant replace an insurance advisor?
No, and anyone selling that is selling something. An assistant is good at what a term means and what a cover normally does. It is useless at whether your sum insured is right, because that needs someone to look at your building. The line is between explaining and advising, and only the second one pays a claim.
What does an AI assistant get wrong in insurance?
It hedges when it should answer, and it answers when it should hedge. Ours told people all premiums attract GST for months after individual life and health were exempted, because the underlying material said so. The model was not wrong, the corpus was. That is the real lesson: an assistant is a mirror of what you publish, including the parts that are out of date.
Is it safe to ask an AI about my insurance policy?
Ask it what things mean, yes. Do not paste your policy number, Aadhaar or bank details into any chat window, ours included; we redact them before storage but not everyone does. And never act on a coverage answer from any assistant without checking the wording, because it has not read your policy.
What should a small brokerage know before building one?
That the corpus is the product and the model is a commodity. Every wrong answer we have had came from our own material being wrong or missing, never from the model inventing something. Budget for maintaining what you publish, not for prompt engineering.
Why publish this rather than keep it quiet?
Because the useful version of this subject is written by people running one, and almost all of what exists is written by people selling one. Our numbers are small and we say so. Small honest numbers are more useful to another broker than a market forecast.
Related: what Raksha does and refuses to do · how we source what we publish.