Learning from disputes across the market
IIB Board AI Pack, page 27. Second phase. Built here on mock and synthetic data.
Today. Consumer commissions and the Insurance Ombudsman publish decisions on insurance disputes. They are long legal texts, so patterns across thousands of them are hard to see.
What would change. AI reads published decisions and records the product, insurer, issue and outcome of each, then links them to IIB's policy data. It shows which products, terms and sales channels lead to the most disputes, including mis-selling in life insurance.
In plain words. A ULIP (unit linked insurance plan) puts most of the premium into market funds, so its value rises and falls. Bancassurance means a bank sells the policy for the insurer. Mis-selling means a policy was sold on a misleading promise or to someone it did not suit.
Why rates. A product sold in large numbers draws more complaints just by being big. So this page divides the decisions by the policies in IIB's data and compares decisions per 10,000 policies.
The answer
What the writer was given
1. Complaints as rates, by product and sales channel
Decisions per 10,000 policies in force
Life policies only. Groups with fewer than 3,000 policies are left off the chart, because one decision moves their rate too much. They are still in the table.
Mis-selling, every group
Who buys bank-sold ULIPs, from IIB's policy data
| Policies | Buyers aged 55 or over | Lapsed within a year |
|---|---|---|
| ULIPs sold by banks | 33.1% | 41.0% |
| ULIPs sold in other ways | 5.0% | 17.7% |
| All other life policies | 2.2% | 15.7% |
Policies started from October 2024 to September 2026. The lapse share counts only policies that have had a full first year. Built into the test data: 33.1% of bank-sold ULIP buyers aged 55 or over and 41.0% lapsing within a year. The page reads the same from the data.
The decisions tell the same story. Of the 38 bank-sold ULIP mis-selling decisions read, 35 give the complainant's age and 25 of those were 55 or over (71%). For other life mis-selling decisions it is 25%.
Bank-sold ULIP mis-selling, by insurer
Only three insurers sell ULIPs in the demo data. Insurers shown by alias, not by name.
2. The decisions and what was read
Newest first. The list shows what was read: the AI's read where there is one, else the rules read. Click a decision to see it, with its date, forum and amount.
A decision
Pick a decision from the list.
3. How well the decisions were read
Right, field by field, on the checked sample
| Field | Rules reader | AI reader |
|---|---|---|
| Product type | 58 of 60 | 60 of 60 |
| Insurer | 60 of 60 | 60 of 60 |
| Sales channel | 49 of 60 | 60 of 60 |
| Issue | 60 of 60 | 60 of 60 |
| Outcome | 55 of 60 | 60 of 60 |
| Amount awarded | 51 of 60 | 60 of 60 |
| All six right | 44 of 60 | 60 of 60 |
The checked sample is 60 of the 260 decisions, picked at random. Here the checked answers are the facts each mock decision was built from. In a pilot, a person would check the sample by hand.
Read these scores with care. The mock decisions are clean typed text in one house format. Real decisions come as scanned pages, in several languages and many formats, so both readers would score lower in a pilot.
Read with AI
The rules reader covers every decision, offline, in under a second. The AI reads only when asked, six decisions at a time, and each read is saved by the decision's text, so no decision is sent twice. Before anything is sent, the complainant's name is removed and insurers and life products are given code names.
60 of 260 decisions have an AI read, 60 of them in the checked sample. The rest use the rules read.
The checked sample is read first, so the AI's score fills in quickly.
The planted pattern. 52 decisions were built as mis-selling of ULIPs sold by banks, and all of them say how the policy was sold. The reads in use find 38 of those 52 (73%). All 14 missed are rules reads. In 11 of them the rules did not read the sales channel as a bank, because the sale was described without the words they look for. No other decision was put in the group by mistake.
What a perfect read would show. With every decision read right, bank-sold ULIPs would show 25.5 mis-selling decisions per 10,000 policies, against 2.3 for all other life policies. The reads in use show 18.6 against 1.7. So the missed reads make the pattern look smaller than it is, not larger.
How the readers work
The rules reader takes the insurer from the parties at the head of the decision and the product from its name, or from words such as unit linked or floater. It takes the sales channel from phrases such as relationship manager or Point of Sales person, the issue from keyword counts and the point the forum decides, the outcome from the order, read from the side of whoever appealed, and the amount by adding the sums in the sentences that order a payment. It misses plain descriptions with no keyword, such as a neighbour who sold policies on commission, and it adds sums that should be left out, such as a surrender value already paid.
The AI reader gets the same text with the name removed and code names in place of insurers. It returns the six fields as JSON with a short quote for each. It lists the rupee sums that make up the award, and the page adds them up, so digit grouping such as 20,00,000 is never left to the AI. Anything outside the allowed values counts as wrong.
Which read is used. The AI's read where one is saved, else the rules read. Every table and rate on this page uses those reads. They are never taken from the answers the mock data was built from.
4. Health disputes
Decisions per 10,000 policies in IIB's health data (synthetic), from the reads in use. Senior citizen policies draw the most for their size. The most common issue is pre-existing disease exclusion, in 36 of the 82 health decisions read.
Personal data is masked before any person or the AI sees it. The AI flags and drafts. A person decides. Mock data stands in for outside sources that need an agreement.