HEARTCOUNT Decision Layer Insurance application

Overpayments found.
What should the claims team do next?

Connect claims data with the judgment of experienced reviewers.

When the overpayment rate exceeds its target, a Decision Agent can follow the claims team’s review criteria to organize the evidence: where to investigate first and which individual cases warrant an audit.

Explore the analysis report Watch the investigation Coming soon

Illustrative synthetic claims-review data. Approval, audit-workflow integration, and execution are future directions.

Overpayments in non-life insurance claims review

Look beyond the 5.66% overpayment rate.
Investigate 22 adjusters and two individual cases.

A breached target does not tell the team who to review or which cases to audit. Monthly outliers, recurrence, and individual payment amounts help distinguish a recurring pattern from a high-value incident.

Synthetic-data analysis example

Investigating an overpayment-rate breach

Unit of analysis: claims adjuster × month
July 2025–July 2026 · July data through the 19th

The operational question

Why did the overpayment rate exceed target, and where should we investigate first?

How the agent investigates

  1. Check the metrics

    Compare the 5.66% overpayment rate with the 4% target, alongside the supporting metrics.

  2. Compare like with like

    Examine adjuster-month differences using peers in the same center and period.

  3. Separate review targets

    Distinguish 22 recurring outliers from ADJ0177’s two high-value cases.

  4. Prepare the report

    Set out the priorities for recurrence reviews and case audits, with evidence and open questions.

What the analysis found

Recurring outliers and individual audits need different reviews.

Across 130 outlier-months, 22 adjusters appeared at least twice. Four are active in July-to-date data and warrant priority review. High-value individual cases are examined separately.

Two cases account for most of ADJ0177’s overpayments.

For ADJ0177, KRW 225m of KRW 234m (96%) is concentrated in two CT and MRI cases. Review the source records for possible base-amount entry errors, unapplied coverage limits, or mistaken payments. Concentration alone does not establish misconduct or a cause. Audit findings and recoverability require separate decisions.

Recommendations based on synthetic data, not completed audits or recovered payments. Planned approvals and execution are described separately below.

Inspect the data and evidence

What changed?

Overpayment rate
5.66%Target ≤4% · ▲4.2% vs. prior month
Average processing time
3.86daysTarget ≤5 days · ▲29.5% vs. prior month
Review reversal rate
49.5%▲17.4% vs. prior month

Loss ratio 6.92% · target ≤40%Anomaly ratio 10.56%Non-covered service share 43.90%Overpayment amount: KRW 422.94m

The overpayment rate is the metric exceeding its stated target. Processing time and the review reversal rate are also presented at their highest levels in the example, warranting investigation of a possible link between backlog and review quality. July is incomplete; confirm comparisons with closed months after the data is finalized.

How broad is the exposure?

Outlier-months
130Adjuster × month · January–July · rate >10% and overpayment >KRW 5m
Repeat outliers · at least twice
22adjustersCandidates for a review watchlist
ADJ0177 · two June cases
KRW 225m of KRW 234m (96%)Review individual cases separately from recurring behavior

Top 12 on the watchlist

  • Recurring in July to date
  • Earlier recurrence

Cumulative overpayments (KRW) · two occurrences each

ADJ0123 · Capital LT 3 128.94m
ADJ0011 · Capital LT 1 82.94m
ADJ0070 · Specialist Review 82.92m
ADJ0101 · Capital LT 1 80.72m
ADJ0139 · Specialist Review 2 80.56m
ADJ0049 · Specialist Review 2 76.82m
ADJ0027 · Long-Term Claims 74.18m
ADJ0134 · Yeongnam LT 71.55m
ADJ0080 · Specialist Review 68.45m
ADJ0113 · Capital LT 3 68.05m
ADJ0198 · Specialist Review 1 60.85m
ADJ0042 · Capital LT 2 47.55m

Prioritize ADJ0123, ADJ0139, ADJ0042, and ADJ0091 for July-to-date review. ADJ0199 is the only three-time repeat outlier, appearing in February, April, and June—an alternating-month pattern that a consecutive-month rule would miss.

Which individual cases warrant an audit?

ADJ0177 · June · KRW 225m of KRW 234m (96%)
Item × provider categoryOverpayment (KRW 10,000)Flagged / cell totalShare
CT × hospital11,3811 of 148.5%
MRI × hospital11,0991 of 347.3%
Manual therapy × tertiary institution5521 of 12.4%
Manual therapy × clinic3861 of 31.6%
Shockwave therapy × clinic291 of 20.1%

Shares retain the rounded values from the supplied table. Counts show flagged cases out of all cases in the same item/provider-category cell. On a small screen, scroll the table horizontally.

What still needs checking?

Rerun the watchlist after July closes. Adjustments to the data could reduce some adjusters’ recurrence count to one, so the intervention list must be confirmed again.

Align the supplied January–July summary with the proposed operational rule of two occurrences within six months. Examine whether ADJ0199’s alternating-month pattern relates to case mix or rotating assignments.

Retrieve claim identifiers, payment dates, and coverage details for ADJ0177’s two cases, then reconcile base amounts, coverage limits, and payment records.

See the analysis process in the video · Coming soon

Source: the supplied synthetic claims-review analysis summary, as of July 2026. Adjuster codes and center names are fictional and do not identify real employees or customers.

Give the agent the claims team’s review method

Add the review sequence
to the claims data.

A metric formula cannot tell an agent where to investigate first. Define the trigger, investigation order, and comparison conditions.

KPI

Define an exception

Overpayment rate ceiling: 4%. Supporting thresholds: loss ratio ≤40% and average processing time ≤5 days. An adjuster-month outlier requires both a rate above 10% and overpayments above KRW 5m.

Driver

Prioritize the investigation

Recurrence (proposed operational rule: at least twice within six months), then cumulative overpayments, individual cases above KRW 50m, and finally item × provider category.

Lever

Choose a fair comparison

Narrow from adjuster to center and month, comparing peers in the same center and period. For an individual case, compare other cases in the same item/provider-category cell.

Structure this investigation path and its comparisons in a Decision Graph. Connect the agent to data and metric definitions through Semantic MCP, retaining the steps and evidence behind the analysis.

Explore the role of Decision Layer

After the analysis report

Decide what to do.
Then verify what actually changed.

Investigate possible causes and audit candidates, review proposed actions, and extend the workflow toward approval, execution, and outcome verification.

Illustrative workflow: steps 1–2 are the supplied analysis and recommendations. Steps 3–6 are future directions. Expected benefits are evaluation targets; no post-action improvement is reported.

01 / 06

Supplied analysis example

Move from adjuster-months to individual cases.

Distinguish recurring outliers, consecutive occurrences, and high-value individual incidents. A single detection rule will not capture all three patterns.

Investigation scope

Unit of analysis
Claims adjuster × month
Recurrence and continuity
22 watchlist candidates · a three-consecutive-month pattern
High-value cases
ADJ0177 · KRW 225m of KRW 234m (96%)

Check consecutive patterns even when amounts are small, and use separate criteria for alternating-month recurrence and large individual payments.

02 / 06

Proposed actions

Prioritize case audits and recurrence prevention separately.

Consider source-case audits, recoverability reviews, watchlist interventions, and detection rules. The timing and benefits below are illustrative proposals, not completed work.

01 · Audit ADJ0177’s two cases and assess recoverability

Illustrative timing: one week after kickoff. Review KRW 225m of KRW 234m (96%), presented in the supplied proposal as 18% of June’s company-wide overpayments. Actual recovery depends on the audit.

02 · Review interventions for 22 watchlist candidates

Illustrative timing: within two weeks. Prioritize the four July-to-date repeat outliers. Moving from 5.66% into the 4% range is a proposed interim target; it does not meet the final ≤4% threshold by itself.

03 · Configure recurrence, consecutive-month, and high-value rules

Illustrative timing: within one month. Reducing detection lag from monthly to same-day is a target to validate. Set the high-value alert threshold and data-refresh cadence separately.

Decisions before implementation

Confirm the population
Rerun the watchlist on finalized July data
Assign ownership
Name the audit and intervention teams and accountable owners
Set detection conditions
Agree high-value thresholds, refresh cadence, and review rules

Recovery and KPI improvement remain expected benefits. Confirm the evidence and permitted scope before taking action.

03 / 06

Future direction

A claims-review owner approves the scope and conditions.

The authorized claims-review manager checks the evidence, target cases, proposed interventions, and permissions before approving.

Approval information

Approver
An authorized claims-review manager
Scope
Two audit candidates · 22 intervention candidates
Conditions
Approval validity · stop and reconsideration conditions

Business-action approval requires development beyond the existing ability to ask a person for analytical input or a selection.

04 / 06

Future direction

Pass approved requests to claims and audit systems.

Connect approved audit requests and watchlists to the claims-review system and audit workflow.

System responsibilities

Decision Layer
Manage decision evidence, chosen actions, and constraints
Execution module
Check authority and approval, then form a system-specific request
Claims system
Perform the permitted workflow and return its status

Claims-system execution connectors are not included in the public framework.

05 / 06

Future direction

Distinguish an accepted request from work actually performed.

Record the audit request, track its processing status, and check separately whether audit or recovery procedures have started.

Execution evidence

Request record
Target · time · approval · request identifier
Status
Received · in progress · complete · failed
Applied action
Source-case audit progress · recovery procedure initiated

Sending a request does not establish that an audit or recovery has completed.

06 / 06

Future direction

Check implementation separately from performance.

Compare overpayment rates, actual recoveries, detection lag, and repeat outliers. Consider automated alerts only for rules whose effects and conditions have been validated.

Outcome criteria

Implementation
Were the approved audits and interventions applied to the intended targets?
Performance
Entry into the 4% range · actual recovery · detection lag · repeat outliers
Final threshold
Check the ≤4% target and unintended effects separately

This example contains no post-action outcome figures. Operational automation remains subject to validation.

Connecting claims decisions to operations

Turn the claims team’s judgment
into a coordinated workflow.

Look across overpayments, processing time, review outcomes, and loss ratio to evaluate the next action.

Combine claims data, reviewer assignments, coverage and item definitions, and review rules. The goal is to weigh recoverability, processing speed, complaints, and loss ratio together.

HEARTCOUNT’s direction is to extend claims analysis toward integration with existing review and audit systems, checking actual implementation separately from its results.

Available today.
What needs configuration. What comes next.

The scenario connects claims data and review criteria to the public framework’s reusable analytical procedures. Metric definitions, MCP access, and report formats require configuration for the organization.

Public framework · available today

Reuse reviewed analysis and inspect its evidence

  • Query connected data through defined metrics
  • Reuse registered procedures, including configured watchlist logic
  • Inspect input, query, result, and validation records
  • Call analysis from Claude and other agents or applications
  • Organize analytical evidence for a report

The seven metrics, Semantic MCP connection, and executive-report format in this example require organization-specific configuration.

Explore the public code and current scope
Environment-specific configuration

Use the claims team’s data and review criteria

  • Overpayment targets and outlier thresholds
  • Coverage, item, and provider classifications
  • Claims data and MCP access
  • Audit criteria, permissions, and report configuration

The 4%, 10% plus KRW 5m, and KRW 50m thresholds belong to the synthetic example. Each organization must adapt and validate them.

Planned open-source release

From proactive investigation to audit-workflow integration

  • Scheduled monitoring rules with the Always-on Decision Agent
  • Approval integration for audit and intervention requests
  • Execution requests to claims systems
  • Outcome checks before expanding automation

The Always-on Decision Agent and approval and execution interfaces are planned for open-source release and are not part of the current public framework. System connectors and approval conditions are configured during implementation.

Start with one overpayment problem
where recovery opportunities may be missed.

Agree on the KPI, claims data, and review criteria first. Then explore the audit procedures and claims systems you may want to connect.

Discuss an insurance application (KO) Example overpayment action report Coming soon