Years ago I led the development of a deep reporting system for WellPoint, the Fortune 500 health insurance company that, at the time, was sitting at around $12.4 billion in revenue and number 204 on the Fortune list. I was the technical architect and project manager, running twelve-plus offshore resources against the design. The interesting thing about that build was not the reporting layer itself. The interesting thing was what it took to make the source data trustworthy enough to report on at all. Member files. Claims history. Provider network records. Eligibility periods. Each one had been shaped by decades of mergers, regulatory cycles, and reclassifications of what counted as what. The reporting layer was the visible deliverable. The forensic data work underneath was the actual job.
That underneath part is what nobody puts in an AI vendor pitch deck, and it decides whether a payor’s AI investment earns its keep or quietly becomes a cost center.
flowchart TD
A[Inbound claim] --> B{Clean claim?<br/>Coded correctly,<br/>eligible member,<br/>contracted provider}
B -->|Yes| C[Auto-adjudication path<br/>Rules engine + AI confirmation]
C --> G[Paid in seconds<br/>Cost per claim falls]
B -->|No - missing data| D[AI repair queue<br/>Suggest corrections]
D --> C
B -->|No - clinical judgment| E[Specialist review<br/>Medical director, coder]
B -->|No - fraud signals| F[Special investigations<br/>Human-led]
class G good
class E warn
class F warn
classDef good fill:#163a26,stroke:#44cc77,color:#d7ffe6;
classDef bad fill:#3a1620,stroke:#ff5555,color:#ffd9d9;
classDef warn fill:#3a2e16,stroke:#ffaa33,color:#ffe9c7;
classDef accent fill:#15233b,stroke:#4488ff,color:#dce9ff;
What a Fortune 500 payor reporting system actually looks like
The WellPoint engagement was not glamorous. To the outside it looked like dashboards and pivot tables on top of a warehouse. To the team building it, it was a multi-year exercise in resolving how membership records, claims records, provider records, and benefit configuration records reconciled to each other across systems written, acquired, and re-platformed at different points in the company’s history. The hardest meetings were the ones where two different source systems each insisted they held the canonical version of the same fact about the same member, and a business decision had to be made about which one the reports would treat as truth.
That data-truth arbitration is the prerequisite for every AI use case a payor is being pitched today. A model that recommends auto-adjudication on a borderline claim is making a decision against a member’s eligibility record, a provider’s contract terms, a benefit plan’s coverage rules, and a clinical coding interpretation. If any of those four upstream sources is in conflict with the others, the model’s recommendation is built on sand, and the appeals team finds out later. Payors that already did the data-truth work have a runway. Payors that have not done it do not have a runway, no matter what the model vendor says.
The five AI use cases payors actually fund
The vendor universe is enormous. The actual budget concentration at large payors clusters around five use cases, in roughly this order.
Claims auto-adjudication on the clean-claim band is the workhorse. Most large payors already auto-adjudicate a meaningful share of inbound volume through traditional rules engines. The AI lift is not in replacing that path. The lift is in pushing the auto-adjudication boundary outward into the band of claims that historically pended for human review, by using a model trained on prior adjudication outcomes to recommend a confident path-forward decision. The economics are real because a human reviewer touching a claim costs meaningfully more than a model scoring it. The economics collapse if the model’s recommendations have to be human-reviewed anyway because nobody trusts them yet, which is the failure mode of most early deployments.
Fraud, waste, and abuse detection on high-risk claims is the second. A small share of claim volume drives a disproportionate share of recovery dollars, and a model that scores that share for human investigation has a clear payback story. The catch is that the model is only as valuable as the special investigations unit it feeds. Scoring without investigator headcount to pursue the flags is academic. The payors who get real returns on fraud AI staffed the investigative team first, then added the scoring layer.
Provider network scoring is the third. A model scoring providers on quality, cost efficiency, member satisfaction, and access patterns gives the network team a defensible basis for contract negotiation and tier placement. The economics are not in cost reduction. They are in network composition decisions that shift utilization toward higher-performing providers, which over a multi-year horizon moves the medical loss ratio in a direction the CFO notices.
Prior authorization triage is the fourth. The opportunity is the queue. Clinical reviewers are expensive and in short supply. A model that routes a meaningful share of requests to a clear-approve or clear-deny path, leaving the genuinely ambiguous cases for clinical review, reduces queue time and reviewer load. The regulatory exposure is real, because a denied prior auth is a member-facing decision with appeal rights. Serious deployments keep a human in the loop on every denial and use the model only for routing, not final adjudication.
Member service chatbots are the fifth and most over-pitched. The economics work when the bot deflects high-volume, low-complexity calls about benefits, claims status, and ID cards. They break when the bot tries to handle coverage decisions, appeals, or anything with a regulatory dimension, because the cost of a wrong answer is enormous. Successful deployments draw the line between transactional service and decision service early and hold it.
Where payor AI economics stay stubbornly expensive
There are entire categories of payor work where AI has been pitched for years and the economics still do not clear. High-complexity adjudication is the canonical case. The claims that historically required a human reviewer required them because the decision involved clinical judgment, multi-policy benefit coordination, or fact patterns the rules engine could not encode. A model can assist on the periphery, but the decision itself will keep involving a human reviewer, because the regulatory and contractual environment requires a defensible human decision-maker on file. The cost per decision in this band is bounded by the cost of a qualified human, not an AI curve.
Appeals workflow is the second stubborn category. An appeal is a contested decision by definition. The economics are dominated by the clinical and legal review required to defend or overturn the original decision. A model can organize the case file and surface precedent. It cannot make the appeal decision, and the cost structure tracks skilled review labor, not inference.
Provider contract negotiation is the third. Negotiations are relationship-driven, multi-party, and shaped by local-market competitive dynamics no model has visibility into. AI can prepare the data and arm the negotiator. It cannot conduct the negotiation, and the negotiator is the cost driver.
Regulatory reporting is the fourth. The reports are not the cost. The cost is the compilation, reconciliation, and certification work required to produce reports the payor can stand behind in front of state insurance commissioners, CMS, and external auditors. A model can accelerate compilation. The certification, where a named human signs off on submission accuracy, remains a human accountability function.
The data-foundation work that has to come first
Nothing in the use-case list above lands if the data foundation is not in place. A payor putting AI in the claims path needs three things to be true.
The claims data has to be integrated across source systems. At most large payors this means the EDI history, the provider master, the member eligibility history, and the prior authorization decisions all queryable together, with consistent identifiers. The WellPoint reporting system I led had to solve a version of this before any reporting could happen. Modern AI deployments have to solve a harder version, because the model needs not just the current state but the historical state at the time each prior decision was made.
The historical adjudication outcomes have to be retained with enough fidelity to train against, including the overturn-on-appeal cases, because those teach the model where the front-line decision got it wrong. Many payors retain adjudication outcomes but lose the appeal-overturn linkage in the warehouse, which biases the training data toward the original decision. A model trained on that data learns to reproduce the original errors.
The data has to be governed in a way that supports periodic retraining. A model deployed in a claims path drifts as the population, the provider network, and the benefit configurations change. Retraining requires a defined access path to fresh data, a defined review path for the new model version, and a defined deployment path that does not require a full litigation event each time. Payors without this scaffolding find that their first model deployment is also their last.
HIPAA and the regulatory dimension shape the architecture
HIPAA does not change which use cases are economically viable. HIPAA changes the deployment topology, and the topology changes the unit cost of inference. A model running in a generic commercial API tenant is not deployable against protected health information without a business associate agreement, network isolation, access logging, and a defined data-handling path. The cheapest API call on a pricing page is not the cheapest production call once those controls are priced in. Most serious payor deployments end up on a private cloud tenant or a regional data residency commitment, with infrastructure overhead the model card never mentions.
The state insurance commissioner dimension is the other architectural shaper. Any AI system that touches a coverage decision, a claim denial, or a prior auth denial has to produce an audit trail the commissioner’s office can inspect, reproducing for any individual decision the model version, input features, recommendation, human override (if any), and final outcome. Building that auditability after the fact is hugely expensive. Building it in from the start separates a deployment that survives its first regulatory inquiry from one that gets paused.
ERISA adds a fiduciary layer on the self-funded side. The Affordable Care Act adds member-appeal-rights dimensions affecting how denials have to be communicated, including the underlying basis for the decision. A model in any of those paths has to produce an explanation that is legally and clinically defensible. Successful claims AI deployments treat explainability as a feature requirement from day one, not as a bolt-on.
What this means for a payor IT leader sizing the investment
The honest sizing of payor AI economics requires three numbers most pitches skip. The first is the cost of the data-foundation work the use case actually depends on, including integration, historization, and governance scaffolding. The second is the all-in cost of inference inside the HIPAA-compliant deployment topology, not the public API rate. The third is the cost of the human review capacity that has to remain in place because the model is decision-supporting, not decision-making, in every category that matters. Adding those three numbers honestly to the projected savings produces a payback timeline that is meaningfully longer than the vendor case, and meaningfully shorter than the do-nothing case.
The WellPoint reporting work I led decades ago did not have AI in it because the technology of the era did not support it. What that work did have was a serious discipline about data lineage, source-of-truth arbitration, and decision defensibility. Those are the disciplines that decide whether a modern payor AI investment earns its return. The technology has changed. The disciplines have not.
