The initial version of the Auth-to-Claim proof of concept has one major limitation: the extensive manual intervention required for auth/claim mismatches.
For each mismatch, the user must click the claim to open a detailed remediation screen. From there, the user can run the applicable reconciliation “agent” (a blend of deterministic and reasoning logic) to identify possible next steps.
For example, Claim C1002 is matched to Authorization AUTH002, but the quantities differ - “Claim billed quantity 3 exceeds authorized quantity 1”. When the user clicks on the “Discover quantity options” button, a series of checks are performed:
- Standard quantity limit
- Variance analysis
- Rounding tolerance
- Alternate authorization search
The goal is to determine if the quantity mismatch can be resolved automatically. In this particular case, the answer is no, and a set of recommended actions follows: (action copy generated using Cursor’s Composer 2.5 model)
-
Verify the billed quantity against the prescribed dosage and any relevant clinical guidelines.
-
Check for any potential package conversion issues that may affect the billed quantity.
-
Investigate if a loading dose was appropriately documented and authorized.
-
Consider submitting a corrected claim or requesting an adjustment to the authorization to resolve the quantity mismatch.
Since this information is generated by the LLM, examination of a second quantity mismatch outlines similar recommendations:
-
Confirm the prescribed dose and units to ensure they align with the authorization.
-
Check for any potential package conversions that may have been overlooked.
-
Investigate if a loading dose logic applies that could justify the billed quantity.
-
Consider whether a corrected claim or an adjustment to the authorization is necessary.
Currently, the staff must investigate each mismatch and consider these recommendations on a per-case basis. In addition, while the application helps explain why a mismatch occurred, it does not take any action to resolve the issue. Unfortunately, this approach doesn’t scale.
An alternative approach might rely on a multi-agent workflow to analyze and, if necessary, address these mismatch cases.