Connected Claims USAFind us at booth 502, Signia by Hilton Orlando, Nov 3-4Booth 502 · Orlando, Nov 3-4Booth 502 · Connected ClaimsBook a meeting

Why 99% Accuracy Isn’t Enough for AI in Insurance

Large language models are correct most of the time. But getting a decision wrong on a claim can lead to many downstream issues like lawsuits and lower customer satisfaction.

We tested 9 models on 400 claim pages, all of them were 99% accurate. This sounds like a problem solved but for critical workflows we need to understand what 1% the model gets wrong and whether the model can tell which 1% it is.

What we tested

We ran 9 models on 2 tasks for our insurance claims work.

Three small decision models - Jev, Gemini Decisions and OpenAI Decisions. These models answer a question with the fixed options and give out a confidence score

6 large reasoning models - GPT-6 Astra, GPT-6.1 Sol and Terra, and Gemini 4 Argon, Carbon and Hydrogen. None of these models return a confidence score. As a workaround we used sampling confidence. We run the same prompt multiple times. We can count how often a model answers the same. If the answer is the same 4/5 times. We can say that it is 80% confidence

Experiment 1: what kind of page is this? (easy)

We use 400 pages from insurance claims. 200 are pdfs and 200 are photos or scans. We use the same instruction and the same seven options across the models. Aim of this experiment is to classify each page into one of seven categories (invoice, receipt, estimate, medical record, claim form, other document, not a document)

DecisionsJevGemini DecisionsGPT-6 AstraGPT-6.1 SolGPT-6.1 TerraGemini 4 ArgonGemini 4 CarbonGemini 4 Hydrogen
Correct, out of 400397398396395395394395395398
Mistakes made while sure001556452
Pages held back for a person6610110100
Answer changed when options were reversed537331112
Cost per 1,000 pages~$0.28~$0.22~$0.15~$51~$9~$17~$21~$12~$22

"Sure" means 0.8 or higher. For GPT-6 Astra and the GPT-6.1 and Gemini 4 models, that is 4 or 5 of their 5 answers agreeing. Their costs cover all five answers. Gemini Decisions sorts options alphabetically by name whatever order they are sent in, so for the reversed run we added letter prefixes to the option names to flip that order.

All of the models are right 99% of the time so accuracy doesn’t separate them. But when we look at the mistakes we see that Jev and OpenAI Decisions make mistakes while being unsure about the answer (0.59 to 0.75) so a threshold of 0.8 is a good enough filter for review. Gemini Decisions lets 1 through at 0.83. So only a handful of documents will be routed to a human

Reasoning models make 2 to 6 mistakes and asking them repeatedly doesn’t help. They gave same answer 5 out of 5 times

We also ran each page with options in reverse order. Every model changed its answer on 1 to 7 of the 400 pages when the options were in a different order.

Experiment 2: is the billed visit in the medical records, and what was the urgency? (hard)

For this experiment we will use 95 claims with 1 invoice and a full medical history and a third of the histories are more than 60k characters

Accuracy changed when the task changed.

Page type · 400Urgency · 95
Decisions99.2%81.1%
Jev99.5%77.9%
Gemini Decisions99.0%85.3%
GPT-6 Astra98.8%96.8%
GPT-6.1 Sol98.8%97.9%
GPT-6.1 Terra98.5%98.9%
Gemini 4 Argon98.8%89.5%
Gemini 4 Carbon98.8%91.6%
Gemini 4 Hydrogen99.5%92.6%
Share of answers correct on each task. The small models drop by 14 to 22 points on urgency and the Gemini 4 models by 7 to 9; GPT-6 Astra and the GPT-6.1 models stay within 2 points.
DecisionsJevGemini DecisionsGPT-6 AstraGPT-6.1 SolGPT-6.1 TerraGemini 4 ArgonGemini 4 CarbonGemini 4 Hydrogen
Visit in records: correct of 95908585949594909192
Urgency: correct of 95777481929394858788
Urgency mistakes while sure381321986
Urgency answers held back363444021211
Long histories: visit found of 31302325313131293131
Cost per 100 claims$0.17$0.08$0.16$24.37$4.20$8.52$5.95$8.88$14.98

The GPT-6.1 and Gemini 4 models are newer reasoning models, each asked five times like GPT-6 Astra. The reference answers were set partly by GPT-6 Sol, a close relative of GPT-6.1 Sol, so its scores may be flattered. Their costs use list prices from third-party listings, and the GPT-6.1 Terra, Gemini 4 Carbon and Gemini 4 Hydrogen costs are upper bounds because we had no cached-input price.

In this experiment the ranking flips, the models fall into 3 groups

  1. The decision models struggle, they hold back more than a third of the cases for human review. Jev made 8 mistakes while sure so its signal was not safe to rely on
  2. GPT-6 models are the most accurate, they get 92 or more right.
  3. Gemini 4 models are the worst. They make 6 or more mistakes while sure

GPT-6 Astra costs around $25 for 100 claim documents against $0.16 for a decision model. But GPT-6.1 Sol costs around $5 while maintaining the same level of accuracy. This means that accuracy on harder tasks is getting cheap

Accuracy against cost on the urgency question.

100%90%80%75% $0.10$1$10 cost per 100 claims, log scale GPT-6.1 Terra $8.52 · 98.9% GPT-6.1 Sol $4.20 · 97.9% GPT-6 Astra $24.37 · 96.8% Gemini 4 Hydrogen $14.98 · 92.6% Gemini 4 Carbon $8.88 · 91.6% Gemini 4 Argon $5.95 · 89.5% Gemini Decisions $0.16 · 85.3% Decisions $0.17 · 81.1% Jev $0.08 · 77.9%
Each point is one model on the 95 medical-record claims. GPT-6 Astra costs about 140 times as much as OpenAI Decisions or Gemini Decisions; the GPT-6.1 models match its accuracy for a sixth to a third of its cost.

What do both the experiments say

  1. Accuracy is a wrong number to compare models for our usecase. All the models were 99% accurate for the first experiment
  2. A newer model doesn’t help. GPT-6.1 Sol repeated exactly same five mistakes GPT-6 Astra made
  3. Neither kind of model wins at both the tasks - smaller models perform great at tasks which require quick judgement. Larger models are better suited for search and reasoning tasks
  4. Human reviewers need to know which answers to review. Larger models don’t give any signal for that

In practice, I think we should use smaller models for quick checks and route their unsure answers to human review. Larger models should be used for complex reviews with safeguards

Limitations