Matching motor claims to accident reports
IIB Board AI Pack, page 23. Later. Built here on mock and synthetic data.
Today. Police now file digital accident reports through MoRTH's e-DAR portal, with photos and videos of the scene and the vehicles. Insurers use them claim by claim.
What would change. AI matches each motor claim to its accident report across all insurers, reads the report and compares the photos with the damage claimed. Claims with no matching accident, or damage that does not fit, are flagged. Genuine claims can move faster.
This page checks the 1,467 motor claims with a loss date in Jul to Sep 2026 where a police report would normally exist: every third-party claim (412) and every own-damage claim of ₹1 lakh or more (1,055). It looks for each one among 2,858 accident reports in a mock copy of e-DAR, the police's digital accident report system. 1,433 of those reports match no claim here, since many accidents never lead to an insurance claim. Matching and the photo check are plain rules. The AI's part is reading the officers' narratives, and a rules reader does that job wherever the AI has not read a report.
The typical question
1. Match rates by insurer and by state
17 of the 37 third-party claims with no report sit in 3 RTOs in Uttar Pradesh, where 49% of third-party claims have no report, against 5% everywhere else. The Motor team sees which RTOs.
Away from those RTOs, Insurer W and Insurer K stand out. 8 of Insurer W's 26 third-party claims have no report. 6 of Insurer K's 22 third-party claims have no report. Other insurers' third-party claims have no report 2% of the time.
An RTO or insurer stands out when at least 3 of its third-party claims have no report, and that many would turn up by chance less than 1 time in 100 at everyone else's rate.
By insurer
| Insurer | Claims | Matched | No report | Flagged | Match rate |
|---|---|---|---|---|---|
| Insurer W | 93 | 82 | 11 | 17 | 88.2% |
| Insurer K | 101 | 91 | 10 | 12 | 90.1% |
| Insurer V | 34 | 31 | 3 | 6 | 91.2% |
| Insurer N | 66 | 64 | 2 | 6 | 97.0% |
| Insurer R | 29 | 27 | 2 | 3 | 93.1% |
| Insurer D | 63 | 61 | 2 | 8 | 96.8% |
| Insurer A | 24 | 22 | 2 | 3 | 91.7% |
| Insurer X | 65 | 64 | 1 | 4 | 98.5% |
| Insurer M | 34 | 33 | 1 | 3 | 97.1% |
| Insurer I | 42 | 41 | 1 | 3 | 97.6% |
| Insurer G | 26 | 25 | 1 | 4 | 96.2% |
| Insurer P | 118 | 117 | 1 | 9 | 99.2% |
| Insurer J | 73 | 72 | 1 | 6 | 98.6% |
| Insurer O | 93 | 92 | 1 | 3 | 98.9% |
| Insurer L | 53 | 52 | 1 | 4 | 98.1% |
| Insurer E | 45 | 44 | 1 | 5 | 97.8% |
| Insurer Q | 39 | 38 | 1 | 5 | 97.4% |
| Insurer S | 85 | 85 | 0 | 2 | 100.0% |
| Insurer Y | 107 | 107 | 0 | 7 | 100.0% |
| Insurer H | 12 | 12 | 0 | 0 | 100.0% |
| Insurer F | 94 | 94 | 0 | 9 | 100.0% |
| Insurer B | 51 | 51 | 0 | 5 | 100.0% |
| Insurer U | 31 | 31 | 0 | 1 | 100.0% |
| Insurer T | 58 | 58 | 0 | 3 | 100.0% |
| Insurer C | 31 | 31 | 0 | 2 | 100.0% |
Insurers shown by alias, not by name.
Share of claims with no report, by state
Darker means a larger share of claims in scope found no report. Two-letter codes are states. Hover for the name and the share. States with fewer than 20 claims in scope are left grey, since one or two claims would swing the share.
By state, most claims with no report first
| State | Claims | Matched | No report | Flagged | Match rate |
|---|---|---|---|---|---|
| Uttar Pradesh | 179 | 161 | 18 | 30 | 89.9% |
| Maharashtra | 165 | 160 | 5 | 16 | 97.0% |
| Kerala | 84 | 80 | 4 | 8 | 95.2% |
| Madhya Pradesh | 123 | 120 | 3 | 11 | 97.6% |
| Andhra Pradesh | 48 | 46 | 2 | 5 | 95.8% |
| Bihar | 45 | 43 | 2 | 5 | 95.6% |
| Tamil Nadu | 75 | 73 | 2 | 6 | 97.3% |
| Assam | 37 | 36 | 1 | 2 | 97.3% |
| Delhi | 47 | 46 | 1 | 4 | 97.9% |
| Karnataka | 107 | 106 | 1 | 6 | 99.1% |
| Punjab | 49 | 48 | 1 | 3 | 98.0% |
| Sikkim | 1 | 0 | 1 | 1 | 0.0% |
RTO detail is for the Motor vertical. The Leadership role sees states.
2. Damage and dates that do not fit
One flagged case, shown without claim number, insurer or district
- The claim names the right side, which the report does not mention. The report shows the left side.
- The photo shows damage at the left side, which the claim does not mention.
The scene photo
A mock photo drawn for the demo: the vehicle seen from above, glass towards the front. The image check finds the vehicle, works out which end is the front from the windscreen, and finds the dark damaged patch. Here it sits at the left side.
Flagged claims by insurer
| Insurer | Claims | Flagged | No report | Damage | Dates |
|---|---|---|---|---|---|
| Insurer W | 93 | 17 | 11 | 5 | 1 |
| Insurer K | 101 | 12 | 10 | 1 | 1 |
| Insurer F | 94 | 9 | 0 | 5 | 4 |
| Insurer P | 118 | 9 | 1 | 4 | 4 |
| Insurer D | 63 | 8 | 2 | 5 | 1 |
| Insurer Y | 107 | 7 | 0 | 5 | 2 |
| Insurer N | 66 | 6 | 2 | 3 | 1 |
| Insurer J | 73 | 6 | 1 | 4 | 1 |
| Insurer V | 34 | 6 | 3 | 2 | 1 |
| Insurer B | 51 | 5 | 0 | 5 | 0 |
| Insurer E | 45 | 5 | 1 | 2 | 2 |
| Insurer Q | 39 | 5 | 1 | 2 | 2 |
| Insurer X | 65 | 4 | 1 | 3 | 0 |
| Insurer G | 26 | 4 | 1 | 3 | 0 |
| Insurer L | 53 | 4 | 1 | 3 | 0 |
| Insurer R | 29 | 3 | 2 | 0 | 1 |
| Insurer M | 34 | 3 | 1 | 2 | 0 |
| Insurer I | 42 | 3 | 1 | 2 | 0 |
| Insurer O | 93 | 3 | 1 | 2 | 0 |
| Insurer T | 58 | 3 | 0 | 2 | 1 |
| Insurer A | 24 | 3 | 2 | 1 | 0 |
| Insurer S | 85 | 2 | 0 | 0 | 2 |
| Insurer C | 31 | 2 | 0 | 1 | 1 |
| Insurer U | 31 | 1 | 0 | 1 | 0 |
| Insurer H | 12 | 0 | 0 | 0 | 0 |
Insurers shown by alias, not by name.
Individual claims, with their reports and photos, are for Data Operations only. This role sees totals.
3. How the reports and photos are read
Reading the report
Officers write the narrative in English, often with Hindi words or sentences. The page needs two things from it: the type of collision and the damaged sides of the claim's vehicle. A rules reader covers every report offline. It turns Hindi words into English ones, finds who hit whom and works out the sides that implies. The AI reader reads the same narratives in batches, only when someone asks, and keeps each read.
| Reader | Reports checked | Collision type right | Damaged sides right |
|---|---|---|---|
| Rules, every matched report | 1,425 | 100.0% | 92.1% |
| AI, reports it has read | 112 | 100.0% | 97.3% |
| Rules, on those same reports | 112 | 100.0% | 94.6% |
Scored against what each mock report was written to say. Collision types come out right so often because the mock reports use a small set of phrasings. Real narratives vary far more. The rules miss sentences such as "rammed into it", where the vehicle is a pronoun, and take a "left front door" as damage at the front as well as the left. On the reports both have read, the two readers differ on the claim's vehicle 9 times. The AI was right in 6 of those and the rules in 3. Where the AI finds no damaged side for the claim's vehicle, the rules read is used.
The AI sees only the narrative's sentences about the vehicles. Sentences about people hurt stay behind. Narratives carry no vehicle numbers or names, and every prompt is checked for personal data before it leaves. 110 of the matched claims use an AI read. 1,315 use the rules.
Looking at the photo
A simple image check, not AI. It finds the vehicle against the road, uses the windscreen to tell the front from the back, and finds where the dark damaged patch sits: front, rear, left or right.
It works on these synthetic photos, which are drawn from above. Real scene photos would need a trained vision model. This stands in for one.
How a claim is matched to a report
The claim's vehicle number comes from its policy. The page looks for it among the vehicles named in every report. A report matches when the number is the same, or one character off with the accident within 2 days and in the same state. Typing slips such as O for 0 or two digits swapped are common in police records. 69 claims matched that way. 100 matched reports were filed in a district other than the vehicle's home RTO district, which is normal when people travel.
An exact number with dates more than 2 days apart, up to 90 days, is taken as the claim's report with the dates flagged. A report that already fits another claim on the same vehicle by date is left to that claim. A claim with no match is flagged once it is more than 7 days old, since reports usually reach e-DAR within a week. No unmatched claim is younger than that.
The mock e-DAR copy also holds reports for accidents with no IIB claim, some with vehicle numbers one character away from a claimed vehicle, and some for the same vehicle months earlier. None of those should match, and the measure below checks that.
4. How we would measure it
Each kind of flag against the test data
| Flag | Flagged | Confirmed | Precision | Built in | Found | Recall |
|---|---|---|---|---|---|---|
| No accident report | 42 | 34 | 81% | 34 | 34 | 100% |
| Damage does not fit | 63 | 52 | 83% | 53 | 52 | 98% |
| Dates differ | 25 | 25 | 100% | 25 | 25 | 100% |
Precision is the share of flags that were real problems. Recall is the share of real problems that were flagged. The false flags are 8 genuine claims whose police report reaches e-DAR after 6 October 2026, and 11 genuine claims where the surveyor listed a corner side the police did not note. A reviewer would clear them. 1 built-in problem was not flagged because the photo was too unclear for the image check, and the report agreed with the claim.
Did the page find the cluster?
The test data hid a cluster of third-party claims with no report in 3 RTOs in Uttar Pradesh, and at Insurer K and Insurer W across the country.
The page found 3 of the 3 RTOs. It found 2 of the 2 insurers (Insurer W and Insurer K).
The problems were built into the test data, so these numbers show the method finds what it was built to find. The real test is a pilot with insurers' claims teams on real e-DAR reports.
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.