How Do Healthcare and Insurance Payers Use Process Intelligence? Key Findings From Skan AI

McKinsey estimates AI could save payers $150 million to $300 million in administrative costs per $10 billion of revenue (McKinsey, 2025). Realizing those savings is an operations problem. Most payers cannot yet capture them because automation runs on documented rules and event logs while staff work exceptions across mainframes, portals, and spreadsheets.

Process intelligence shows healthcare payers where claims, prior authorization, payment integrity, servicing, underwriting, and member services work actually stalls. The eight findings below pair Skan AI deployment results with patterns in that work.

What does process intelligence reveal in payer operations?

Operations leaders use process intelligence for payers to measure where claims, prior authorization, and servicing work stalls and to find rework loops, denial root causes, and workarounds.

• Skan AI captures human-system interactions at the desktop level and feeds that record into the Context Graph of Work. This continuously updated model of work connects people, applications, tasks, decisions, policies, and outcomes, including cross-system handoffs that systems of record miss.
• Process mining instead reconstructs a process from event logs, the timestamped records instrumented systems write. An adjudication platform logs a status change instead of the minutes an examiner spent in a provider portal and an Excel tracker.

Finding 1: Stalled claims adjudications trace back to upstream data issues

Auto-adjudication rates (i.e., rate of claims the rules engine pays or denies without an examiner) stall for reasons upstream of it. Fallout claims carry missing provider data, benefit configuration mismatches between product lines, and clinical edit exceptions configuration teams never anticipated, often rooted in provider data maintenance or benefit setup in a different system.

Examiners work each fallout claim across the adjudication platform, provider data tools, and configuration screens, so every repeat of an unfixed cause adds those touches to cost per claim.

A denial reason code records the outcome, not the upstream step behind it, so teams must trace denials back through those cross-system steps. Skan AI’s Intelligence shows operations leaders where fallout work accumulates by step and system. Rank fallout causes by claim volume and fix the largest in configuration before adding examiner capacity.

Finding 2: Manual second and third touches inflate claims cost

In a typical case, an examiner pends a claim, a colleague re-keys the data into a correspondence queue, and a supervisor reopens it after a provider call. Adjudication audit trails miss manual rework that happens between status changes, so second and third touches inflate claims cost beyond the process intelligence insurance claims data that teams receive.

Those touches are often where cycle-time targets slip. Cost-per-claim models under-count them, while a claims leader who tracks the second-touch rate per claim type gets a continuous leakage indicator that cost-per-claim reporting cannot provide on its own.

Finding 3: Prior authorization turnaround breaches hide in non-clinical work

Turnaround reports show whether a case breached, not whether the hours went to clinical review or administrative waiting. Observation of work separates clinical determination from documentation chase: faxing providers for missing notes and toggling between the care management platform and provider portals. Breaches can cluster where cases wait on documentation or peer-to-peer scheduling (the physician-to-physician review call).

Under CMS-0057-F, the Interoperability and Prior Authorization Final Rule, Medicare Advantage payers and payers serving Medicaid and CHIP must decide expedited requests within 72 hours and standard requests within 7 calendar days starting in 2026, and the core API provisions, including the Prior Authorization API, take effect January 1, 2027 (CMS, 2024).

A prior authorization leader who observes reviewer work under privacy by design can see how many hours of the 7-day window go to fax and portal work an API or an agent could absorb. The “CAQH 2024 Index,” based on calendar year 2023 data, puts a medical plan’s per-transaction labor and overhead cost, excluding system costs, at $3.41 for a manual prior authorization versus $0.05 for a fully electronic one (CAQH, 2024). Observed reviewer hours show which transactions still run manually, where administrative cost can fall.

Finding 4: Payment integrity shifts from post-payment recovery to prevention

Examiner screens show overpayment signals at adjudication: an unchecked secondary coverage flag, an unsurfaced prior claim for the same service, and a hand-keyed pricing override. Payment integrity teams later use them for coordination of benefits (which payer pays first when a member has two coverages) or duplicate-claim recoveries. GAO states that prepayment reviews can better protect funds because not all identified overpayments are collectible (GAO, 2016).

Prevention depends on knowing which examiner steps produce the leakage. Lower-variance steps, those that run the same way each time, are stronger candidates for prepayment edits (i.e., automated checks that stop a questionable claim before it is paid) because they carry fewer exceptions that break rules-based automation.

Finding 5: Policy servicing runs the same transaction many different ways

One endorsement type can run as many process variants, the distinct paths staff take through the same transaction. For example, one processor completes a mid-term endorsement in a few screens across the mainframe policy administration system and the imaging and billing applications; another takes far more because a system gap forces an offline rate calculation in Excel.

This variation stays invisible to process mining because it relies on event logs.

A servicing leader who observes every variant can isolate the fastest observed path and measure the handling-time gap to the rest; priced as cost per transaction, that gap is the standardization opportunity.

Finding 6: Underwriting judgment lives in untracked Excel overrides

Celent describes many insurers still relying on Excel spreadsheets for pricing support. Senior underwriters’ override judgment lives in Excel using supplemental third-party data rather than in any tracked system.

For an underwriting leader facing retirements, observed senior-underwriter override patterns capture institutional knowledge that would otherwise leave with those underwriters, and they define what a future agent may price alone.

With its private-by-design desktop agent, the Skan AI platform captures those decision patterns as agentic operating procedures (AOPs): machine-usable representations of observed work that encode workflow steps, decision rules, exception handling, and escalation thresholds.

Finding 7: Member services navigation waste drives up handle time

Telephony and customer relationship management (CRM) systems record talk time and disposition codes, not the work between them: navigation across mainframe eligibility records, Citrix-hosted claims views, and the CRM, with repeated lookups and copy-paste on every call. Manual quality assurance sampling misses it too, often covering “less than 5 percent” of conversations (McKinsey, 2024).

On a typical member call, the agent moves between the CRM, mainframe eligibility records, and a Citrix-hosted claims view, repeating lookups and copy-pasting between them, and every repeated step adds to handle time.

Capturing every interaction’s on-screen path lets a member services leader see which navigation sequences produce lower handle time and higher first-call resolution, and train toward them.

Finding 8: Agents piloted on documented processes stall when they hit the exception paths payer workflows actually run on

Agents piloted on documented processes stall at exception paths. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 (Gartner, 2025). It names escalating costs or unclear business value as reasons; inadequate risk controls are another reason.

The pattern is consistent with agents grounded only in documented rules and event logs, because each undocumented exception path forces a human handoff, which adds cost and leaves the agent’s risk controls untested.

Skan AI Agents uses the Context Graph of Work and AOPs to ground and govern agent execution using how the payer’s own examiners handle the work. The observation-to-agent pipeline turns observed work, captured under privacy by design, into agent instructions, and teams configure each agent’s scope and escalation paths in Agent Studio before launch.

Each agent is evaluated against the human baseline for the same workflow. That comparison is the basis for governed autonomy, meaning the agent acts alone only inside limits the payer has set and tested. A claims leader moves volume only where the agent matches examiner decisions and human fallback is defined.

How payers protect PHI while observing claims work

Examiner screens carry protected health information (PHI), so payers build privacy controls into the capture architecture itself. Skan PRISM is an on-premises privacy and information security gateway that applies pixel-level redaction before protected information can leave the customer environment; raw screen data remains inside the customer-controlled perimeter, and only abstracted metadata reaches analytics. Skan AI never captures email content or keystrokes.

Skan AI’s security documentation lists SOC 2 Type 2 alongside the ISO 27001 and ISO 42001 standards as third-party audited and describes deployments as HIPAA-ready.

Which payer workflow should you start with?

Start with the workflow where cost per claim, turnaround compliance, or (for member services) handle time is furthest off target and the audit trail is thinnest.

• Claims operations: Findings 1, 2, and 4 give the auto-adjudication, second-touch, and leakage evidence a COO expects.
• Prior authorization and care management: Finding 3 ties reviewer time to the CMS-0057-F clock.
• Policy servicing and underwriting: Findings 5 and 6 quantify variant cost and preserve underwriting judgment.
• Member services: Finding 7 replaces QA sampling with continuous coverage.
• Revenue cycle: Findings 4 and 7 tie prepayment leakage and member-call navigation to cost.
• AI program leads: Finding 8 sets the behavioral baseline an agent needs before launch.

The gated whitepaper, Process Intelligence for Healthcare and Insurance, works through each finding’s measurement method, and the payer process intelligence roadmap blog sequences a first deployment.

Request a Skan AI demo and we’ll show you how the platform supports your organization’s process and evidence requirements and business goals.

Example.