Early death claims
IIB Board AI Pack, page 17. First phase. Built here on synthetic data.
Today. Each insurer sees its own death claims soon after a policy starts. A cluster spread across insurers, regions or agents is invisible to any one of them.
What would change. Matches death claims to policy start dates across insurers, and flags unusual clusters by region, sales channel or agent. It is the Life vertical's weekly alert.
Early death claims in Bihar PoS cluster far beyond chance
In Bihar, over the last 26 weeks, there were 62 early death claims against 0.07 expected in the PoS channel. This is far beyond chance. The claims span 5 insurers: Insurer A, Insurer E, Insurer H, Insurer I, and Insurer O.
A median of 90 days from policy start to death and 64 days from death to intimation were seen. 95% of these early claims were on policies issued without a medical check. 90% were told to the insurer late. Twelve agents stand out in this cluster.
These are flags for review, not findings, and nothing is published. The next step is to ask the insurers for the claim files and check the agents' sales.
Drafted by Azure OpenAI from the figures below. Every number in it was checked against them. Saved from an earlier request with the same figures.
What the AI saw: the figures below, with insurers as code names. No claim, policy, agent or person.
The figures the answer was written from
{
"vertical": "Life",
"week_end": "2026-10-04",
"alert_week": "28 September to 4 October 2026",
"counts_are_over": "each item's own window, given in counted_over, not only the alert week",
"window_weeks": [
4,
8,
13,
26
],
"level": "state totals",
"clusters": 1,
"items": [
{
"where": "Bihar · PoS",
"state": "Bihar",
"districts_in_state_cluster": 2,
"channel": "PoS",
"weeks": 26,
"counted_over": "the last 26 weeks, 6 April to 4 October 2026",
"early_claims": 62,
"expected": 0.07,
"times_expected": "over 100",
"early_claims_this_week": 4,
"insurers": 5,
"insurer_labels": [
"Insurer A",
"Insurer E",
"Insurer H",
"Insurer I",
"Insurer O"
],
"agents_standing_out_on_their_own_here": 12,
"pct_of_these_early_claims_on_non_medical_policies": 95,
"pct_of_these_early_claims_told_late": 90,
"median_days_start_to_death": 90,
"median_days_death_to_intimation": 64,
"how_unusual": "far beyond chance",
"runs_by_chance_as_high": 0,
"runs": 999
}
],
"agg": {
"weeks": 26,
"early_claims": 62,
"expected": 0.07,
"pct_of_these_early_claims_on_non_medical_policies": 95,
"pct_of_these_early_claims_told_late": 90,
"amount_claimed": "₹10.0 cr"
},
"insurers_spanned": 5,
"agents_standing_out_on_their_own_anywhere": 12,
"book": {
"counted_over": "1 October 2024 to 4 October 2026",
"early_claims_elsewhere": 41,
"pct_of_early_claims_elsewhere_told_late": 7,
"pct_of_early_claims_elsewhere_on_non_medical_policies": 63,
"pct_of_all_death_claims_told_late": 18,
"pct_of_all_death_claims_that_were_early": 4.8,
"pct_of_first_year_policies_that_are_non_medical": 70
},
"late_means_days": 30,
"early_means_days": 365,
"min_claims": 3,
"published": false,
"agent_detail": "not shown at state level"
}Words used here. An early death claim is a death within 365 days of the policy start, counted in the week the insurer was told of the death (intimated). PoS means a Point of Sales person, a licensed seller of simple products. Non-medical means the policy was issued without a medical examination. Told late means more than 30 days after the death. Expected is the number of early claims the usual rate gives for the same first-year policies, by age, sum assured and underwriting.
Where early death claims stand out
| Where | Window | Early claims | Expected | Against expected | Insurers | Districts | Non-medical | Told late | How unusual |
|---|---|---|---|---|---|---|---|---|---|
| Bihar · PoS | last 26 weeks | 62 | 0.07 | over 100 times | 5 | 2 | 95% | 90% | flag far beyond chance |
Select a row for its detail. Each group is shown over the window where it stands out most, of the last 4, 8, 13 and 26 weeks to 4 October 2026. A flag needs 3 or more early claims and a result that fewer than 5% of 999 runs of the whole scan reach by chance. For this role each flagged district is rolled up to its state and sales channel.
By state
Early death claims inside flagged clusters, by state, last 26 weeks. Select a state for its totals.
Bihar · PoS
Last 26 weeks, 6 April to 4 October 2026
Month by month
Insurers
Agents selling these policies
The claims
Agents who stand out on their own
12 agents stand out. Agent detail is for the Vertical Heads, the Analyst and Data Operations. The Leadership role sees state totals.
Districts that stand out, all channels together
2 districts stand out. District detail is for the Vertical Heads, the Analyst and Data Operations. The Leadership role sees state totals.
How we would measure it
The problems were built into the test data on purpose (story L1: PoS policies in two districts of Bihar, sold by 12 new agents across five insurers, with early death claims from May 2026). So here we can score the alert against a known answer. That shows the method works on a known case. The real test is a pilot, where reviewers confirm or reject each flag and the share they confirm is the measure.
| Flags | Flagged | Planted | Flagged and planted | Precision | Recall |
|---|---|---|---|---|---|
| District and channel clusters | 2 | 2 | 2 | 100% | 100% |
| Districts | 2 | 2 | 2 | 100% | 100% |
| Agents | 12 | 12 | 12 | 100% | 100% |
| Early claims inside flagged clusters | 62 | 62 | 62 | 100% | 100% |
Precision is the share of flags that are part of the planted problem. Recall is the share of the planted problem the flags cover. Computed live on each build of the Data store.
How the alert works
1. First-year policies. Every life policy counts as in its first year for 365 days from its start, or until it lapses, is surrendered or the holder dies. The alert counts first-year policy days week by week.
2. The usual rate. A Poisson model (a standard model for counts of rare events) gives the early death rate per first-year policy year by age at entry, sum assured and medical or non-medical underwriting. It is fitted on the whole book from 1 October 2024 to 4 October 2026. A cluster would raise the average it is judged against, so the model is fitted again without the policies and claims of any cluster the previous pass flagged, until the flags stop changing. This run took 2 passes and left out 2 district and channel groups and 12 agents. The final fit used the other 41 early claims over 39,253 first-year policy years, about 1.04 per 1,000 years.
3. Expected claims. For any group of policies, the expected number of early claims is the sum of each policy's first-year days in the window times its usual rate.
4. The scan. Every district and sales channel (712 groups), every district (123) and every agent (2,892) is compared with its expected number over the last 4, 8, 13 and 26 weeks. A likelihood ratio says how far observed is above expected. Testing so many groups would throw up false alarms, so the whole scan is run again 999 times on made-up claims drawn at the usual rates, and a group is flagged only when fewer than 5% of those runs produce a result as extreme anywhere. The score that 5% of runs exceed was 7.5 for district and channel groups, 6.0 for districts and 8.2 for agents. A flag also needs 3 or more early claims.
5. What the flags show. For each cluster, how many insurers it spans, the agents behind it and what share of its claims were non-medical or told late. Across the whole book 103 of 2,159 death claims (4.8%) were early, and 18% of all death claims were told late.
| Model group | Level | Early death rate against the first level |
|---|---|---|
| Age at entry | 18 to 34 | 1.00 times |
| Age at entry | 35 to 44 | 2.13 times |
| Age at entry | 45 to 54 | 2.72 times |
| Age at entry | 55 to 64 | 5.57 times |
| Age at entry | 65 and over | 9.34 times |
| Sum assured | under 5 lakh | 1.00 times |
| Sum assured | 5 to 10 lakh | 1.65 times |
| Sum assured | 10 to 25 lakh | 1.35 times |
| Sum assured | 25 lakh to 1 crore | 1.17 times |
| Sum assured | 1 crore and over | 0.33 times |
| Underwriting | medical | 1.00 times |
| Underwriting | non-medical | 0.87 times |
The base rate, for ages 18 to 34, sum assured under 5 lakh and medical underwriting, is 0.47 early deaths per 1,000 first-year policy years. Group effects are pulled slightly towards 1 so a group with few claims does not get an extreme rate. The scan ran in 0.95 seconds and is kept until the Data store is rebuilt.
Policyholder names, contact details and policy numbers never leave the Data store, so none can reach this page or the AI. The policy number is used only inside it, to match each claim to its policy. Agent ids are registry numbers, not personal data. The scan flags, the AI only drafts the words, and a person decides. This use needs no mock data. Both of its sources are synthetic insurer-side data.
Flags are for review only. Nothing is published.