The Wrong Frame

Medical loss ratio is the metric health plan leaders are measured against, and the instinct when a metric is under pressure is to manage the thing the metric measures. For MLR, that means managing medical cost — tightening prior authorisation criteria, negotiating provider rates, implementing step therapy protocols, reviewing high-cost outlier cases. These are recognisable tools of health plan management. They work at the margin. And they share a structural limitation that is rarely made explicit: they are all interventions that happen after the patient is already in an expensive care pathway.

Prior authorisation happens after a provider has decided a patient needs a service and written a request. Concurrent review happens after the patient has been admitted. Post-acute management happens after the discharge. The care decision, in each case, has already been made. The plan is managing access to a pathway that is already in motion.

This is not a criticism of utilisation management as a discipline. It is an observation about where in the clinical timeline these tools operate — and therefore what they can and cannot achieve. A plan that invests exclusively in the authorisation and review layer is investing in interventions that occur at a point where the cost trajectory has largely been established.

The health plans that are materially moving their MLR with AI have recognised something different. High-cost events are not random. A hospitalisation does not appear from nowhere. A readmission, a preventable emergency visit, a chronic disease crisis — each of these is preceded by detectable signals: a change in medication adherence, a missed appointment, a laboratory value trending in the wrong direction, a pattern of increasing utilisation over months. The signal precedes the event. The intervention window exists. And most plans are not acting within it.

The Intervention Window

The concept worth making precise is the intervention window — the period between the moment a signal becomes detectable and the moment the high-cost event occurs. The width of this window determines what intervention is possible. A window of 48 hours allows crisis triage. A window of 30 days allows care coordination. A window of 90 days allows the kind of proactive clinical engagement that changes the trajectory.

The AI investment question that matters most in healthcare insurance is not “does this model detect the right thing?” It is “how far upstream does this model move the point of action?”

Claims-based risk stratification — the dominant approach in most health plans — identifies high-risk members by analysing their historical utilisation patterns. A member who has been hospitalised twice in the last twelve months, who has a chronic condition with documented gaps in care, who has used the emergency department as a primary care setting — these are correctly identified as high-cost risks. They are also risks that have already expressed themselves. The signal is not preceding the event. The signal is the event. The intervention window for these members, at the point of claims-based identification, has closed or is closing.

The plans that are compressing the intervention window are using different signals. Pharmacy fill patterns that indicate medication adherence changes weeks before a clinical deterioration. Laboratory values that are trending toward problematic thresholds rather than crossing them. Social determinants signals — housing instability, food insecurity, transportation barriers — that predict care access failures before they occur. Remote monitoring data from members with congestive heart failure, diabetes, or COPD that generates alerts days before a hospitalisation that claims data would only record after the fact.

The difference in intervention window between claims-based and forward-looking signal models is not marginal. It is often the difference between acting before a hospitalisation and acting to manage the readmission that follows one.

Where Prior Authorisation Fits

The prior authorisation debate deserves specific attention because it absorbs a disproportionate share of plan resources — clinical, operational, and reputational — relative to its MLR impact.

Prior authorisation has genuine utility as a tool for ensuring clinical appropriateness. The problem is not the tool. It is where in the clinical timeline it operates. Prior authorisation intercepts a care decision that has already been made by a clinician and a patient. The interaction at that point is adversarial by design: the plan is evaluating a decision that someone else has made, at the moment when reversing it generates the most friction. The clinical benefit of denials that are subsequently overturned on appeal is zero. The administrative cost of processing those appeals — on both sides — is material. The member and provider satisfaction cost is significant and measurable in disenrollment and network attrition.

This does not mean prior authorisation should be abandoned. It means the investment case for AI in prior authorisation should be evaluated clearly: it is an investment in managing a care pathway that is already in motion, at the point of highest friction and lowest clinical leverage. The same capital invested upstream — in identifying the members at highest risk of needing high-cost services before the service request is made, and engaging them through care management before the provider visit that generates the request — creates more MLR impact, with less friction and better clinical outcomes.

The plans that have reduced prior authorisation volume through upstream care management have not done so by relaxing criteria. They have done so by reaching members before the care decision is made, and influencing the trajectory at the point where clinical leverage is highest.

Risk Adjustment as a Timing Problem

The timing argument extends beyond medical management into risk adjustment — a domain where the financial stakes are high and the timing dimension is consistently underappreciated.

In Medicare Advantage, each 0.01 RAF score point is worth approximately $80–100 per member per year. HCC coding gaps — conditions that are clinically present and documentable but not submitted — represent both a revenue integrity risk and, more importantly, a care management failure. A member whose chronic kidney disease is not coded is not only generating a revenue gap. They are likely not enrolled in the chronic disease management programme that their diagnosis would qualify them for. The coding gap and the care gap are the same gap.

Retrospective chart review — the dominant approach to HCC accuracy in most plans — identifies coding gaps after the diagnosis year has closed. The revenue impact is recovered through the submission cycle. The care management opportunity is not. By the time a retrospective review identifies that a member’s condition was undercoded last year, the intervention window for that condition in that year has long since closed.

Prospective risk stratification — using current pharmacy, laboratory, and clinical signals to identify members whose likely diagnosis burden is not yet reflected in their submitted codes — serves both objectives simultaneously. It identifies the coding gaps before the submission year closes. And it identifies the members who need care management engagement now, rather than the members who needed it last year.

The Recommendation

Before the next AI investment decision is made, evaluate every candidate initiative against a single question: how far upstream does this model move the point of intervention?

Apply this question to the existing portfolio as well. Map each deployed AI system to its position in the clinical timeline — how many days, weeks, or months before a high-cost event does it generate an actionable signal? Calculate the proportion of total AI investment that is operating in the pre-event layer versus the active-care-pathway layer. In most health plans, this exercise will reveal a significant concentration downstream.

The rebalancing that follows is not about abandoning claims management, utilisation review, or authorisation processes. Those tools serve real purposes. It is about recognising that the plans with the strongest MLR trajectories are not the ones that manage care most aggressively once it has started. They are the ones that reach members before the care decision is made — because that is where the clinical leverage is highest and the cost of action is lowest.

MLR is a timing problem. The AI investments that solve it are the ones that move the intervention window upstream. Every other investment is managing the cost of events that better upstream AI would have prevented.