The connectivity layer payer AI is missing

Health plans have spent a decade building real data infrastructure. The investment was sound. It lets a plan store data, exchange it between systems and build models on top. What it doesn't do is tell a plan when two records belong to the same person.

That connection is identity, and it is the layer most payer AI is missing. A plan can run any model, but if it can’t tell whether the member in the enrollment system is the member in the care management system, the results can't be trusted. Gartner has estimated that poor data quality costs organizations an average of $12.9 million a year. For a health plan, it shows up as duplicate records and the hours spent reconciling them by hand.

Getting identity right comes down to three capabilities.

Recognizing the same person over time

A member enrolls, drops coverage, and returns two years later, often under a new member ID. This happens across every line of business, from Medicaid redeterminations to the annual churn in commercial and Medicare Advantage plans. Recognizing that returning member as the same person, with their history intact, is the first capability.

Everything else sits on top of identity. Every model a plan runs, from risk stratification to fraud detection, inherits the quality of the data beneath it. When records are fragmented or duplicated, errors compound across every downstream decision. Seeing the member whole is what lets a plan document conditions completely, across providers and pharmacies and back through prior enrollment and find the members who would benefit from care management.

Once identity is resolved, the plan can include richer inputs. Social determinants and behavioral health data sharpen models. All of it rests on a foundation that ties every record to the right person.

Seeing who and what is connected to them

The second capability is seeing the people and organizations around the member. The caregiver who manages the appointments. The household the member belongs to. The provider, who is never just a name but a person at a location, under a tax ID, in a set of contracts that differ by payer.

Linking members who share a household is hard to build, but financial systems have figured it out. If you apply for a mortgage, the lender may tell you someone in your household has already applied for a loan. Healthcare hasn't built similar capabilities.

The payoff is that members stop doing the work the plan was supposed to do. Today a member repeats their medical history at every new visit and carries results from one office to the next. The question a plan never wants to hear is, "You don't have information about me?"

Connection also shapes the moments that decide how a member feels about the plan. A prior authorization is decided on the provider's actual network status and the member’s own history, so care isn't delayed by a wrong denial. Or the member isn't charged out-of-network rates for a doctor the directory listed as in-network.

The member never sees a data problem. They see a plan that delayed their care or sent a bill for a visit it promised was covered. Trust, once lost, is hard to win back.

Standing behind every answer

The third is that every output should trace back to the data that produced it, the transformations applied along the way, and the rules that shaped it, with someone accountable for the result. You can’t reconstruct traceability once a model is in production. It must be built into the workflow from the start, through deterministic lineage and human review.

The slowest part of deploying AI inside a health plan is getting the compliance and clinical reviewers to trust the output enough to sign off. A plan should be able to say who owns the data, who is accountable when it is wrong, and what a model is being optimized for.

In June 2024, the largest Medicare Advantage organizations overturned 95% of appealed prior authorization denials for skilled nursing facility admission, a rate federal reviewers said raises concerns about the initial denials themselves. A reversal rate that high means the plan could not stand behind its decisions. A single contractor handled half of all SNF requests in that review, which makes it a provenance problem as much as an accuracy one.

A plan that skips this and goes straight to deployment simply gets confident wrong answers at scale.

Start where the data enters

Validate and standardize identity in real time where data enters the plan, at onboarding and at live call intake, rather than starting with the historical member database. That shifts the cleansing work out of a central IT budget and into daily operations, where it starts paying off sooner.

The full effort is a multiyear project that takes real capital, and leaders serve their organizations better by saying so, including how much time it takes and what it costs in budget and staffing, rather than by promising a foundation on a schedule no one can meet.

The industry has shifted from careful stewardship of data toward something closer to careful freedom. The plans that move thoughtfully but decisively will define what the health plan of the future looks like.

Get the identity layer right, and every system in the health plan can answer the one question everything else depends on. Who is this person, and who are they connected to?

Thomasina Anane is the associate vice president, enterprise analytics at the Alliance of Community Health Plans; Vinay Kulkarni is the chief information officer of SCAN Health Plan; and Martin Hougaard is general manager, product marketing, at Verato.