Claims

AI adjudication

Claude-class reasoning on Amazon Bedrock (apac.amazon.nova-pro-v1:0, ap-south-1) reviews only the claims the rules engine could not settle. Every recommendation is validated, graded against hidden truth, and sealed in the audit chain.

Operating mode

Assist: recommendations are pre-filled on the examiner's decision panel; a person records every outcome.

Change in claims policy
3,575 claims waiting for review · assist mode
adj-mvi71s41done50/50 processed0 applied0 errors$0.0912315 ms avgapac.amazon.nova-pro-v1:0
14:14:54.098Starting: 50 claims, model apac.amazon.nova-pro-v1:0, mode assist, concurrency 4
14:14:59.96310/50 processed, 0 applied, 0 errors, 0.018 USD
14:15:05.48120/50 processed, 0 applied, 0 errors, 0.036 USD
14:15:10.32830/50 processed, 0 applied, 0 errors, 0.055 USD
14:15:19.14440/50 processed, 0 applied, 0 errors, 0.073 USD
14:15:24.42050/50 processed, 0 applied, 0 errors, 0.091 USD
14:15:24.420Finished: 50 processed, 0 applied, 0 errors
AI recommendations
89
3,664 gray-zone claims in total
Accuracy vs truth
93.3%
89 graded
Validated
100%
evidence, codes, confidence, ceiling
Average confidence
0.93
policy floor 0.8
Latency
2.2s
average per claim
Spend
$0.16
122,998 in / 19,765 out tokens

Confusion matrix

Rows: AI recommendation. Columns: expected outcome.

AI saysApprovedDeniedPendedReferred to SIU
Approved43000
Denied02600
Pended42140
Referred to SIU0000
A recommendation that sends the claim to a person counts as a pend. Rules engine alone: 90.7% on the 30,982 claims it decided.

Precision and recall

Per outcome, against hidden truth

OutcomePrecisionRecallExpected
Approved100%91%47
Denied100%93%28
Pended70%100%14
Referred to SIU——0
Calibration
  • confidence 0.95 to 1.00100% correct of 80
  • confidence 0.90 to 0.95—
  • confidence 0.80 to 0.9033% correct of 9
  • confidence below 0.80—

Accuracy by scenario

What kind of gray-zone claim the model handles well

ScenarioGradedCorrect
Dx mismatch3294%
Unbundling13100%
Missing info12100%
Unbundling supported9100%
High dollar8100%
Replacement6100%
Clean40%
Oon emergency3100%
Auth unverifiable2100%

Recent recommendations

Latest recommendation per claim

ClaimAIConf.CodesHuman?ValidExpectedBilledStatusWhen
CLM-2026-029399Approved0.95NoYesApproved$782.00Needs review2026-10-21 14:15
CLM-2026-024408Approved0.95YesYesApproved$1,206.00Needs review2026-10-21 14:15
CLM-2026-025496Approved0.95NoYesApproved$108.00Needs review2026-10-21 14:15
CLM-2026-028245Pended0.95CO16NoYesPended$166.00Needs review2026-10-21 14:15
CLM-2026-032646Approved0.95NoYesApproved$3,864.00Needs review2026-10-21 14:15
CLM-2026-016985Pended0.95CO16NoYesPended$197.00Needs review2026-10-21 14:15
CLM-2026-029037Approved0.95NoYesApproved$315.00Needs review2026-10-21 14:15
CLM-2026-030098Denied0.95CO50NoYesDenied$2,667.00Needs review2026-10-21 14:15
CLM-2026-028277Pended0.80CO16YesYesPended$1,935.00Needs review2026-10-21 14:15
CLM-2026-019207Approved0.95NoYesApproved$229.00Needs review2026-10-21 14:15
CLM-2026-023429Approved0.95NoYesApproved$86.00Needs review2026-10-21 14:15
CLM-2026-021043Denied0.95CO97NoYesDenied$470.00Needs review2026-10-21 14:15
CLM-2026-015936Pended0.80CO50NoYesApproved$2,069.00Needs review2026-10-21 14:15
CLM-2026-017415Approved0.95NoYesApproved$93.00Needs review2026-10-21 14:15
CLM-2026-019170Denied0.95CO97NoYesDenied$460.00Needs review2026-10-21 14:15
CLM-2026-017072Approved0.95YesYesApproved$32,757.00Needs review2026-10-21 14:15
CLM-2026-011954Approved0.95YesYesApproved$2,654.00Needs review2026-10-21 14:15
CLM-2026-015028Pended0.80CO50NoYesApproved$2,822.00Needs review2026-10-21 14:15
CLM-2026-012810Denied0.95CO97NoYesDenied$503.00Needs review2026-10-21 14:15
CLM-2026-030093Pended0.95CO16NoYesPended$278.00Needs review2026-10-21 14:15
CLM-2026-029085Approved0.95NoYesApproved$2,420.00Needs review2026-10-21 14:15
CLM-2026-027610Approved0.95NoYesApproved$1,862.00Needs review2026-10-21 14:15
CLM-2026-027283Denied0.95CO50NoYesDenied$2,600.00Needs review2026-10-21 14:15
CLM-2026-027894Denied0.95CO97NoYesDenied$375.00Needs review2026-10-21 14:15
CLM-2026-020258Approved0.95YesYesApproved$48,120.00Needs review2026-10-21 14:15

Batch runs

RunDoneAppliedCostStatus
adj-mvi71s41
2026-10-21 14:14 · assist
50/500$0.091Completed
adj-mvi6pofs
2026-10-21 14:05 · assist
38/400$0.069Completed

Guardrails on every recommendation

  • Only claims the rules engine left for review reach the model; hard denials and pends never do.
  • The model answers through a strict schema; free text cannot change a claim.
  • Every cited evidence ID must exist in the pack it was given, and every reason code must be in the catalogue.
  • Confidence below the policy floor, or a billed amount above the ceiling, routes the claim to a person.
  • Member name, date of birth and member ID are withheld from the prompt unless the policy allows them.
  • The exact input is hashed and sealed with the model ID, prompt version and policy version, so any decision can be replayed.
  • If the model is unavailable, the circuit opens and the platform continues on rules plus examiners.