The Payer Operations Squeeze

Health plan operations teams are caught between forces that all move in the wrong direction simultaneously.

On one side: regulatory pressure is intensifying. CMS finalization of the Interoperability and Prior Authorization Rule means Medicaid, CHIP, and qualified health plans on the exchanges must meet specific electronic prior authorization timelines: 72 hours for urgent requests, 7 days for standard. State regulators are layering on additional requirements for grievance response times, appeals turnaround, and network adequacy reporting. The margin for missing an SLA is serious: a corrective action plan, a public report, or a contract at risk.

On the other side: administrative budgets are under a microscope. The Administrative Loss Ratio (the percentage of premium dollars consumed by non-clinical operations) is the number that keeps payer CFOs up at night. Every dollar spent on manual claims processing, authorization reviews, and member services is a dollar that doesn't flow to medical spend or margin. For Medicaid managed care organizations (MCOs) operating on thin margins, hitting ALR is about survival.

In the middle: volume keeps growing. More members, more claims, more prior authorization requests, more appeals, more calls. And the labor market for experienced claims examiners, utilization management nurses, and member services representatives isn't getting any easier.

AI agents, autonomous workflows that can read clinical documents, apply policy logic, make determinations, and route exceptions, offer a way to break the linear relationship between volume and headcount. Not by replacing clinical judgment on complex cases, but by handling the 60–70% of operational volume that follows well-defined rules and currently consumes the most staff time.

The KPIs that matter: cost per authorization, first-pass claims rate, appeals cycle time, backlog aging, and per-member-per-month (PMPM) administrative cost.

1. Prior Authorization Review and Approval

The problem: Prior authorization is the single largest operational bottleneck in most health plan operations. A mid-size regional plan might process 15,000–40,000 authorization requests per month. Each request requires a reviewer to:

• Verify member eligibility and benefit coverage

• Confirm the requesting provider's network status

• Determine whether the requested service requires authorization under the member's plan

• If authorization is required, retrieve and review the submitted clinical documentation

• Apply the plan's medical policy criteria (often hundreds of pages of clinical guidelines per service category) to the clinical information

• Render a determination: approve, deny, or request additional information

• Document the rationale

• Communicate the decision to the provider and member within regulatory timelines

For straightforward cases, this process takes a trained reviewer 15–25 minutes. Multiply that by thousands of requests per week, and you have a department of 30–80 people doing work that is largely rule-application, not clinical reasoning.

How AI agents help: An agent can process the full authorization workflow for cases that meet defined criteria:

• Intake and triage: The agent ingests the authorization request (whether submitted via portal, fax, or EDI transaction), extracts the key data elements (member ID, provider NPI, requested service codes, diagnosis codes), and verifies eligibility and benefit coverage in real time.

• Documentation extraction: For requests that include clinical attachments (chart notes, lab results, imaging reports, letters of medical necessity) the agent extracts structured clinical data from unstructured documents. It identifies the relevant clinical findings, prior treatments, and contraindications.

• Criteria application: The agent maps the extracted clinical data against the plan's medical policy criteria for the requested service. For a lumbar MRI, does the patient have documented conservative treatment failure of 6+ weeks? For a biologic medication, has the patient failed the required first-line therapies? The agent evaluates each criterion and produces a structured determination rationale.

• Auto-approval: Cases that clearly meet all criteria are approved automatically, with a complete audit trail documenting the clinical basis for the decision. The approval notification is generated and transmitted to the provider and member.

• Intelligent routing: Cases that don't clearly meet criteria due to missing documentation, borderline clinical findings, or complex comorbidities, are routed to a human reviewer with a pre-assembled case summary: here's what was submitted, here's what the criteria require, here's what's missing or ambiguous. The reviewer starts at the decision point, not at the beginning of the review.

Why it matters: The math is compelling. If 55–65% of authorization requests are clinically straightforward and can be auto-adjudicated, a plan processing 25,000 requests per month removes 13,750–16,250 cases from the manual review queue. At 20 minutes per case, that's 4,500–5,400 hours of reviewer time per month, roughly 27–33 FTEs. Even accounting for the human oversight layer (spot audits, exception handling), the net staffing reduction is substantial.

But the bigger win may be turnaround time. An auto-approved authorization is returned in minutes, not days. For providers, that means the patient gets scheduled faster. For the plan, it means fewer phone calls from providers checking on authorization status, which is itself a major call center volume driver.

The regulatory angle: With CMS timelines tightening, plans that can't process standard authorizations within 7 calendar days face compliance risk. Auto-adjudication of straightforward cases ensures the plan meets timelines on the easy cases, freeing reviewer capacity to focus on complex cases that actually need clinical judgment.

2. Claims Routing and First-Pass Adjudication

The problem: Claims adjudication is the industrial core of health plan operations. The claims processing system handles the bulk of adjudication automatically through rules engines, but a significant percentage of claims fall out of auto-adjudication and land in manual review queues (often called "pend" queues).

Claims pend for many reasons: missing or mismatched member information, coordination of benefits questions, authorization discrepancies, unusual code combinations, pricing exceptions, or provider contract ambiguities. Each pended claim requires a claims examiner to investigate, often across multiple systems, and make a determination. Industry surveys indicate that each claim submission requires an average of three to four minutes of provider staff time, not including time spent creating the claim itself,1 and related transactions such as eligibility verifications and prior authorizations can take over 10 minutes each on average. At scale, pend queues represent hundreds of thousands of dollars in monthly labor cost — and every day a claim sits in a queue adds to the plan's days-in-accounts-payable and provider abrasion.

How AI agents help: An agent sits between the claims processing system and the manual review queue, intercepting pended claims and attempting resolution before they reach a human:

• Data reconciliation: For claims that pend due to member or provider data mismatches, the agent queries across enrollment, credentialing, and provider databases to resolve discrepancies. A claim that pended because the member's name on the claim doesn't exactly match the enrollment record can be resolved easily by an LLM.

• Authorization matching: Claims that pend because no matching authorization is found can be investigated by the agent: Was the auth issued under a different member ID? Was the service date outside the auth window by a small margin? Was the auth issued for a closely related procedure code? The agent applies configurable matching logic to resolve near-misses.

• Coordination of benefits (COB): For claims that pend due to potential other insurance, the agent can query COB databases, check the member's coverage history, and determine primary/secondary liability, resolving a category of pends that is notoriously time-consuming for human examiners.

• Intelligent AI routing: Claims the agent can't resolve are routed to the appropriate examiner queue with a structured summary of what was investigated and what remains unresolved. This eliminates the examiner's investigation time and lets them focus on the actual decision.

Why it matters: Improving the first-pass adjudication rate by even 5–10 percentage points has an outsized impact on operations. For a plan processing 1 million claims per month with a 25% pend rate, a 10-point improvement means 100,000 fewer claims hitting manual queues each month. At $3–$5 per manual touch, that's $300,000–$500,000 in monthly savings, and a measurable reduction in provider payment cycle times.

The payment integrity connection: Faster, more accurate claims routing also reduces improper payments. Claims that sit in pend queues are sometimes bulk-released under time pressure, increasing the risk of overpayment. An agent that resolves pends accurately and quickly reduces both the backlog and the error rate.

3. Call Center and Member Services Compliance

The problem: Health plan call centers handle a staggering volume of interactions: member inquiries about benefits, claims status, provider directories, ID cards, appeals processes, and grievances. For Medicaid plans, the member population often has complex needs and limited health literacy, making calls longer and more nuanced.

These interactions are heavily regulated. CMS and state regulators specify requirements for call answer times, abandonment rates, first-call resolution, and, importantly, the accuracy and completeness of information provided to members. A member services representative who gives incorrect benefit information, fails to document a verbal grievance, or doesn't offer required notices (like the right to appeal) creates compliance exposure.

Manual quality assurance typically covers 2–5% of interactions. For a plan handling 200,000 member calls per month, that means 194,000–198,000 calls go unreviewed.

How AI agents help: An agent can review every call via transcript analysis against a comprehensive compliance and quality framework:

• Regulatory compliance: Did the representative verify member identity per HIPAA requirements? When a member expressed dissatisfaction, was it documented as a grievance per CMS/state requirements? Were required notices provided (appeal rights, state fair hearing rights for Medicaid members, language access services)?

• Information accuracy: Did the representative provide correct benefit information? When quoting cost-sharing, did the figures match the member's actual plan? When describing network status, was the information current?

• Process adherence: For calls involving claims inquiries, was the representative following the correct script for explaining EOBs? For authorization inquiries, was the status communicated accurately? For provider directory requests, was the member directed to the correct resource?

• Sentiment and escalation detection: Did the member express frustration or confusion that should have triggered a supervisor escalation? Were there indicators of a potential quality-of-care concern that should have been routed to the medical management team?

• Grievance capture: One of the most common compliance findings in Medicaid plan audits is failure to capture oral grievances. An agent can identify calls where a member expressed a complaint that meets the regulatory definition of a grievance but was not documented as such, catching a compliance gap that manual auditing almost never detects at scale.

Why it matters: For Medicaid managed care plans, call center compliance is a top audit finding category. CMS and state regulators routinely cite plans for inadequate grievance capture, inaccurate benefit information, and failure to provide required notices. Each finding can result in corrective action plans, financial penalties, or in extreme cases, enrollment sanctions. Automated monitoring transforms compliance from a sampling exercise into a census.

Beyond compliance, the operational intelligence is valuable. Aggregate analysis of call topics, member sentiment, and resolution patterns reveals systemic issues: a benefit design that's confusing members, a provider group generating excessive claims inquiries, a recent policy change that wasn't adequately communicated. These can all be addressed proactively with the help of agentic AI.

4. Care Management: Member Stratification and Outreach

The problem: Health plan care management programs are a core component of both Medicaid and commercial plan operations. Care managers are responsible for identifying high-risk members, conducting outreach, developing care plans, coordinating services, and monitoring outcomes. The challenge is scale and targeting.

A Medicaid plan with 500,000 members might have 50 care managers. Even with risk stratification models, the volume of members who could benefit from outreach far exceeds the team's capacity. Care managers spend significant time on administrative tasks: reviewing claims and utilization data to identify outreach candidates, assembling member profiles, drafting outreach letters and call scripts, documenting interactions, and updating care plans. The actual member engagement is a mere fraction of the workday.

How AI agents help: An agent can automate the administrative scaffolding around care management, letting care managers focus on what they do best: talking to members.

• Dynamic member stratification: Beyond static risk scores, an agent can continuously monitor claims feeds, pharmacy data, ADT notifications, lab results, and social determinant indicators to identify members whose risk profile is changing right now. A member with diabetes who just had an ED visit for hyperglycemia, hasn't filled their insulin in 45 days, and missed their last endocrinology appointment is in active crisis. The agent surfaces these members at the top of the worklist with a structured summary of what's happening.

• Outreach preparation: For each prioritized member, the agent assembles a complete profile: recent utilization, active medications, open authorizations, assigned PCP, specialist involvement, gaps in care (overdue screenings, missed preventive visits), social determinant flags, and prior care management interactions. It drafts a personalized outreach script that references the member's specific situation: "I'm calling because I noticed you visited the emergency room last week for your blood sugar, and it looks like your insulin prescription may need to be refilled."

• Outreach execution: For lower-acuity outreach (appointment reminders, gaps-in-care notifications, wellness check-ins) the agent can execute outreach directly via text, email, or automated call, reserving care manager time for complex, high-touch interactions.

• Documentation: After a care manager completes a member interaction, the agent can generate the care management note from the call summary, update the care plan, and trigger any follow-up actions (referral to community resources, authorization for additional services, notification to the PCP).

Why it matters: Care management is one of the few levers health plans have to directly influence medical costs and member outcomes. But the ROI depends entirely on reaching the right members at the right time with the right intervention. An agent that automates the identification, preparation, and documentation workflow can effectively double the outreach capacity of the existing care management team without adding headcount.

For Medicaid plans, there's an additional regulatory dimension: many state contracts specify minimum care management contact rates for specific populations (e.g., members with serious mental illness, pregnant members, members transitioning from institutional to community settings). Automated outreach for routine touchpoints helps plans meet these contractual requirements while preserving care manager capacity for complex cases.

5. Member-Facing Chatbot and Self-Service

The problem: Members call health plans for predictable reasons: "What's my copay for this?" "Is this doctor in my network?" "Where's my ID card?" "What's the status of my claim?" "Do I need a referral?" These inquiries are high-volume, low-complexity, and expensive to handle through a live representative, typically $5–$8 per call. For Medicaid plans serving populations with limited digital literacy, call volume per member tends to be higher than commercial plans, making the cost pressure even more acute.

At the same time, member experience scores (like CAHPS for Medicare and Medicaid) are increasingly tied to financial incentives. Long hold times and slow resolution directly impact these scores.

How AI agents help: A member-facing AI agent deployed via the plan's website, member portal, mobile app, or SMS can resolve the most common member inquiries autonomously:

• Benefits and cost-sharing: "What's my copay for a specialist visit?" The agent queries the member's specific plan design and provides an accurate answer.

• Provider directory: "Is Dr. Smith in my network?" The agent checks the current provider directory, confirms network status, and can provide office locations, phone numbers, and availability.

• Claims status: "I had a lab done two weeks ago, has the claim been processed?" The agent queries the claims system and provides a plain-language status update, including any member responsibility.

• ID card and document requests: "I need a new ID card." The agent triggers the request and confirms delivery method and timeline.

• Authorization status: "My doctor submitted a prior auth for my MRI, what's the status?" The agent checks the UM system and provides a current status, including any pending information requests.

• Grievance intake: For members who want to file a complaint, the agent can collect the necessary information in a structured format, confirm the grievance category, provide the member with a reference number and expected timeline, and route the grievance to the appropriate team, ensuring regulatory capture requirements are met.

Multilingual capability is particularly important for Medicaid plans serving diverse populations. Modern AI agents can conduct conversations in Spanish, Mandarin, Vietnamese, Arabic, and dozens of other languages without requiring separate language-specific builds.

Why it matters: A well-implemented member-facing agent can deflect 35–50% of inbound call volume. For a plan handling 150,000 member calls per month, that's 52,500–75,000 calls resolved without a live representative. At $6 per call, the annual savings are $3.8–$5.4 million: a direct reduction in PMPM administrative cost.

But the member experience impact may be more strategically important than the cost savings. A member who gets an accurate answer to a benefits question in 30 seconds via chat at 9 PM on a Sunday has a fundamentally different experience than one who calls during business hours, waits on hold for 12 minutes, and gets transferred once. For plans competing on CAHPS scores or bidding on Medicaid contracts where member satisfaction is an evaluation criterion, this matters.

6. Member and Provider Intake Processing

The problem: Health plans process enormous volumes of inbound documents — member enrollment applications, provider credentialing packets, grievance and appeal submissions, clinical documentation for authorization requests, coordination of benefits questionnaires, and more. Much of this arrives as unstructured data: faxed forms, scanned PDFs, email attachments, and portal uploads.

The intake process (receiving the document, classifying it, extracting the relevant data, entering it into the appropriate system, and routing it for action) is labor-intensive and error-prone. A misclassified document sits in the wrong queue. A data entry error in a member enrollment creates downstream claims processing failures. A grievance submission that isn't recognized as such triggers a regulatory timeline violation.

How AI agents help: An agent can automate the full intake pipeline:

• Document classification: Incoming documents are automatically classified by type (authorization request, appeal, grievance, enrollment application, COB questionnaire, provider credentialing document) regardless of format or how they were submitted.

• Data extraction: The agent extracts structured data from unstructured documents: member demographics from enrollment forms, clinical findings from chart notes, provider credentials from applications, grievance details from member letters. It handles handwritten forms, poor-quality faxes, and multi-page documents with mixed content.

• Validation and enrichment: Extracted data is validated against existing records. A member enrollment application is checked against the eligibility file. A provider credentialing document is matched to the provider's existing record. Discrepancies are flagged.

• System entry and routing: Clean, validated data is entered into the appropriate system (enrollment, credentialing, UM, grievance tracking) and routed to the appropriate team for action. The agent applies business rules to determine priority and routing: a grievance involving quality of care is routed differently than one involving billing, and both are timestamped for regulatory timeline tracking.

Why it matters: Intake processing is invisible work: no one notices it when it's done well, but errors cascade through every downstream process. For Medicaid plans subject to enrollment processing timelines, grievance acknowledgment requirements, and authorization turnaround SLAs, the intake step is where the regulatory clock starts. An agent that processes intake in minutes rather than hours or days gives the plan more time to meet every downstream deadline.

The Payer-Specific Considerations

Health plans evaluating AI agent platforms face a set of requirements that differ meaningfully from provider-side deployments:

Regulatory Auditability

Everything a health plan does is subject to audit by CMS, by state regulators, by accreditation bodies (NCQA, URAC), and by employer group clients. Any AI-driven determination (authorization decision, claims adjudication, grievance classification) must produce a complete, auditable trail: what data was considered, what logic was applied, what determination was reached, and why. "The AI decided" is not an acceptable audit response. The trail must be as detailed and defensible as a human reviewer's documentation.

Medical Policy Integration

Health plan medical policies are complex, frequently updated, and vary by line of business (commercial, Medicare, Medicaid). An AI agent handling authorization reviews must be able to ingest and apply these policies accurately and update its logic when policies change, which can happen monthly. The platform needs to support rapid policy updates without requiring re-engineering of the underlying workflow.

Multi-System Integration

Payer operations span a constellation of systems: claims processing (e.g., QNXT, Facets, HealthEdge), utilization management, enrollment, provider credentialing, member services CRM, grievance tracking, pharmacy benefits, and more. AI agents need to read from and write to these systems reliably. The integration layer is often the hardest part of the implementation.

PHI and Security

Health plans handle PHI at massive scale. Any AI platform must meet HIPAA requirements, support BAA execution, and demonstrate appropriate security controls: encryption, access controls, audit logging, data retention policies. For Medicaid plans, there may be additional state-specific security requirements.

Delegation and Oversight

Many health plans delegate functions: utilization management to an IRO, claims processing to a TPA, care management to a vendor. AI agents operating within delegated functions need to respect the delegation boundaries and reporting requirements. The plan remains accountable for delegated functions, so oversight and monitoring capabilities are essential.

Measuring Impact: The Payer Operations Scorecard

A Practical Sequencing Approach

Phase 1: Highest Volume, Clearest Rules (Months 1–3)

Prior authorization auto-approval for clinically straightforward cases. Start with 3–5 high-volume service categories where the medical policy criteria are well-defined and the auto-approval rate is expected to be highest (e.g., routine imaging with clear clinical indication, standard DME, medication refills meeting step-therapy requirements). This delivers immediate volume relief to the UM team and establishes the audit trail and oversight processes.

Member-facing chatbot for the top 5–8 call drivers (benefits inquiries, claims status, ID card requests, provider directory, authorization status). This reduces call center volume quickly and improves member experience metrics.

Phase 2: Revenue Protection and Compliance (Months 3–6)

Claims routing and pend resolution for the highest-volume pend categories (data mismatches, authorization matching, COB). This directly improves first-pass rates and reduces days in A/P.

Call center compliance monitoring across 100% of interactions. This addresses a persistent audit vulnerability and generates operational intelligence about systemic issues.

Phase 3: Operational Transformation (Months 6+)

Intake processing automation across all document types. This is the most complex implementation (many document types, many source systems) but creates the foundation for end-to-end workflow automation.

Care management automation across member stratification, outreach preparation, and documentation. This requires integration with clinical data sources and careful calibration of risk models, but delivers the largest long-term impact on medical cost and member outcomes.

The Bottom Line

Health plan operations have been “optimized” through traditional means (process improvement, offshoring, system consolidation) for decades. Those levers are largely exhausted. The plans that are still running 25% pend rates, 3-day average auth turnaround times, and $25+ PMPM admin costs aren't doing so because they haven't tried to improve. They're doing so because the work is genuinely complex, the volume is genuinely large, and the tools available until recently couldn't handle unstructured clinical data, apply nuanced policy logic, or operate across fragmented system architectures.

AI agents change the capability frontier. They don't eliminate the need for experienced claims examiners, UM nurses, or care managers, but they dramatically reduce the volume of work that requires those expensive, scarce resources. The result is an operations model where humans focus on the cases that actually require human judgment, and everything else moves at machine speed with machine consistency.

Want to see these use cases in action? Book a demo with StackAI’s healthcare team, or learn more about StackAI for healthcare here.