Actuarial AI is trying to pull value-based contracts out of the black box
Physician groups in value-based contracts often operate for months without knowing whether they are generating shared savings or drifting toward a repayment obligation. In Medicare’s Shared Savings Program, that delay is especially stark: the 476 accountable care organizations active in 2025 will not learn until November whether they earned savings or owe losses for the year. CMS has also delayed 2025 financial reconciliation until after it finalizes the 2027 Medicare Physician Fee Schedule, a move tied to a proposed retroactive guardrail for the spending trend used to update ACO benchmarks.
The basic problem is not just administrative lag. Contract performance is buried across claims feeds, attribution files, benchmark logic, carve-outs, and utilization patterns that do not arrive in a single, tidy stream. By the time a group can assemble a credible picture, the performance year may already be closed, and in some cases settlement disputes stretch back years. One example in the conversation was stark: organizations were still trying to determine what was owed from 2022 while the calendar had already moved into 2026.
Arbital Health is building software around that gap. Its model combines contract ingestion, actuarial methods, and an AI assistant designed to answer plain-language questions, run multistep analyses, and store results in shared workspaces. The company says its platform has completed more than 1,500 contract reconciliations. The goal is not simply to speed up settlement. It is to make contract performance legible while the year is still unfolding.
Why payers and providers end up arguing over the same contract
Disputes in reconciliation usually begin with a simple fact: the two sides are not always looking at the same version of reality. Data may live in different systems, arrive at different times, or be interpreted through different operational lenses. Even when both parties have access to the same records, the challenge shifts to contract language itself, which can be difficult to map cleanly onto the underlying utilization data.
That combination creates a familiar settlement problem in risk arrangements. One side may believe the population was managed within expected cost limits, while the other side reaches a different conclusion because the calculations depend on assumptions, timing, and contract-specific definitions. In that setting, a universal source of truth becomes more than an operational convenience. It becomes the only way to know whether the benchmark, attribution, and settlement logic are being applied consistently.
A shared analytic framework also has a bedside consequence, even if the work happens in spreadsheets and software rather than exam rooms. If a group cannot see rising utilization until long after the fact, it cannot redirect care management, medication adherence support, or post-discharge outreach while those interventions still have a chance to change the trajectory.
What actuarial AI is supposed to do before reconciliation arrives
Jai described actuarial AI as a way to democratize access to actuarial expertise. For trained actuaries, the software can function as an augmentation layer, adding benchmarks, data-querying tools, and faster analytic iteration. For provider leaders without actuarial training, it translates contract minutiae into operational terms: where spending is moving, which populations are driving the trend, and what can be done before settlement.
The underlying workflow is bluntly practical. Arbital ingests contractual information, then reconciles it with the relevant data fields available from a payer or provider. From there, the platform layers in calculations such as incurred but not reported claims estimates, benchmarking forecasts, and propensity matching to approximate what utilization should look like in a comparable population. Those inputs feed forecasts of medical loss ratio trajectory and create a running assessment of whether the contract is on track.
The point is not to replace clinical judgment or operational leadership. It is to reduce the delay between signal and action. In a value-based environment, a forecast that arrives after the performance year closes is little help. A forecast that identifies rising cost pressure in real time can influence which care managers are deployed, which patients are prioritized, and which interventions are worth the limited staff time available.
The operational value of seeing rising risk while there is still time to act
Finite resources drive many of the software use cases Jai described. Primary care groups may have medication adherence teams, patient engagement staff, or post-hospital discharge programs, but none of those levers are unlimited. The analytic question, then, is not simply whether utilization is increasing. It is which subgroup is responsible, what type of event is driving the increase, and where intervention will do the most good.
A patient population with congestive heart failure in a defined geography is a useful example. Recurrent hospitalizations in that cohort can rapidly alter the financial profile of a contract, especially if discharge follow-up is inconsistent or medication regimens are not stabilized quickly. A platform that identifies that trend early can point care managers toward the patients most likely to benefit from outreach, follow-up visits, and medication review.
That kind of targeting matters because broad intervention can dilute the effect of scarce resources. If every patient is treated as equally urgent, the group may exhaust its operational capacity without changing the claims trend. If the highest-risk patients are identified early enough, the response can be sharper and more efficient.
CMS policy changes are raising the stakes for midyear visibility
The timing issue is not abstract. CMS estimated in January 2026 that 14.3 million Medicare beneficiaries were receiving care through ACOs, up from 13.7 million a year earlier, and the Shared Savings Program expanded to 511 ACOs. The CMS Innovation Center is also replacing the ACO REACH Model with the Long-term Enhanced ACO Design, or LEAD, Model on Jan. 1, 2027. As more organizations move into risk-bearing arrangements, the need for reliable midyear performance intelligence grows sharper.
A contract that remains opaque until reconciliation can still function as a financial instrument, but it does little to support day-to-day management. By contrast, a system that shows utilization trends, expected claims lag, and contract performance before the year ends gives provider groups a chance to adjust care delivery while there is still time to affect both patient outcomes and settlement results.
The remaining difficulty is not whether the tools can calculate. It is whether the underlying data streams are clean enough, the contract terms are standardized enough, and the organizations using them are ready to act on what the numbers reveal. Settlement will still happen months after the performance year closes. The clinical and financial question is whether the warning arrives soon enough to matter.