Utilization review triage

The clock is the constraint, and it is tightening

Utilization management has always been a queueing problem dressed as a clinical one. A request arrives, someone has to decide, and there is a deadline. What has changed is that the deadlines are shortening across several lines of business, adverse determinations increasingly have to carry a specific reason rather than a generic citation, decision statistics are becoming publicly reportable, and standardized programmatic interfaces for submitting and answering requests are being required on a defined timetable. Confirm the exact dates and scope that apply to your plan and your states — they differ, and they have moved — but the direction is not ambiguous and it is not going to reverse.

The operational consequence is arithmetic. If the clock shortens and volume holds, either reviewer capacity goes up, or the fraction of cases that reach a reviewer goes down. Hiring nurse reviewers at the rate the calendar implies is not realistic for most plans. So the fraction has to fall, which means triage.

That is a solvable engineering problem with one hard boundary, and the boundary is where most programs get into trouble.

The asymmetry rule

Automation may clear a case toward approval. Automation may not produce an adverse determination. That is the whole rule and it is worth stating as plainly as possible, because every serious problem we have seen in this space is a violation of it.

The reasons are practical rather than philosophical. In most regimes an adverse determination on medical necessity has to be made by a qualified clinical reviewer, frequently a physician, and the requirements are specific about who and about what the notice must contain. A denial that a person did not actually make is not merely a policy failure; it is a determination the plan may be unable to defend at appeal, at audit, or in litigation, and the pattern is legible in the data long after the fact.

An approval is different in kind. Approving something that would have been approved anyway costs the plan the amount of the service and creates no adverse action against a member. It is a business decision with a bounded downside that can be measured, sampled and tuned. The asymmetry between those two error costs is enormous, and it should drive the entire architecture.

The corollary matters as much as the rule. A model output that is technically a routing decision but functionally a denial — a low-confidence score that sends a case to a reviewer with a pre-written adverse rationale in the box, a queue whose default action on timeout is an adverse outcome, a pend for information that is never actually resolvable — is the same failure wearing different clothes. If the practical effect of the automation is that a member does not get the service, it is an adverse determination however it is labelled internally.

Where the capacity actually comes from

Three sources, in descending order of return.

Stop reviewing things you approve at a very high rate. Pull two years of decisions and compute the approval rate by procedure code, by place of service, and by requesting provider. Some cells approve at rates so close to universal that the review is a formality costing real clinical hours. Those cells are candidates for removal from the review list entirely, or for automatic approval on submission. This is the largest single lever in most programs and it requires no machine learning at all — only a query and the willingness to act on it. The cost is genuinely bounded, and you can compute it from the same two years of data before you decide.

Exempt providers who are consistently right. Programs that release well-performing providers from review for specific services are established in several states and increasingly common by contract. The engineering is a rolling window, a minimum volume threshold, an approval rate threshold, and a review path that can put a provider back if performance changes. The important design detail is that the criteria and the current status must be visible to the provider, because the point is to change their experience, not just your workload.

Automatically confirm the cases where the criteria are plainly met. Most plans license structured criteria sets, and much of what a nurse reviewer does on a straightforward case is check whether specific facts are present. When those facts are available in structured form — a diagnosis, a duration, a prior therapy, a lab value — the check is deterministic and a rules engine does it correctly and repeatably. Route to approval when the criteria are met. Route to a human when they are not met, when they cannot be evaluated, or when anything is ambiguous. Never route to denial.

Note what that last design does to the shape of the work. The reviewer stops seeing the easy approvals and sees a queue that is denser in judgment, which is better clinical work and also harder work. Plan for it, because reviewer burnout in a queue of nothing but hard cases is real and it shows up as inconsistency.

The documentation arrives as a fax, and that is the real problem

Everything above assumes the facts are available in structured form. Frequently they are not. Clinical documentation supporting a request arrives as attached documents, often scanned, often faxed, often a forty-page chart export in which four sentences matter.

This is ordinary document extraction work, and it is where a large share of the achievable time savings sits — not in deciding the case but in finding the four sentences. The design that works has three properties.

It cites. Every extracted fact points to the page and the passage it came from, and the reviewer sees the passage, not a summary of it. A reviewer who has to go hunting to verify a claim has been given more work rather than less.

It reports what is missing. The most valuable output is often not a fact but the absence of one. If the criteria require a documented trial of a prior therapy and no such documentation is present, saying so immediately — and asking for it immediately — is worth more than any other automation in the workflow, because the clock is running and the information request is the long pole.

It never writes the rationale. Generated text that reads like a clinical justification is the single most dangerous artifact you can put in a review workflow. It is persuasive, it is fast, and it invites a reviewer under time pressure to accept reasoning they did not construct. Extract facts and cite them. Let the human write the determination.

Route by what is missing, not by a score

Most triage designs rank cases by a predicted probability and hand a reviewer the ranked list. That is the wrong shape for this work. A probability tells the reviewer nothing about what to do next.

Route by the reason the case is here. A case where every criterion is met and one document is unreadable goes to a document queue, not a clinical one. A case missing a required piece of documentation goes to an outbound information request within the hour, because that is the step with the longest external latency. A case where the criteria are met but a benefit limit is in question goes to benefits, not to a nurse. A case where the clinical picture is genuinely unusual goes to the physician reviewer with the extraction already done. And a case that is nearly out of time goes to the front of the queue regardless of anything else.

Each of those is a different queue with a different owner and a different service level. Building them separately makes the operation legible to the person running it, which is worth more than an elegant unified score. When a supervisor can see that eighty cases are waiting on outside documentation and eleven are waiting on a specialist, they can act. A single ranked list gives them nothing to act on.

One more routing rule that pays for itself: identify the requests where the member is already receiving the service, or where a delay carries a clinical consequence, and give them their own path. Those cases are where a slow queue turns into a real harm and a serious complaint, and they are usually identifiable from the request itself.

The metrics that tell you the truth

Utilization management reporting is full of numbers that look like performance and are not.

Report the turnaround distribution, not the average. An average within the standard is compatible with a meaningful number of cases breaching it, and the breaches are what create complaints, regulatory attention and harm. Report the ninety-fifth and ninety-ninth percentile and the count of breaches, by request type and by urgency.

Start the clock when the request arrived, not when someone opened it. This sounds obvious. It is the most common measurement defect in the field, and it flatters the number by exactly the amount of the intake backlog — which is the part that is usually broken.

Watch the overturn rate on appeal. This is the honest quality metric for a review program and it is uncomfortable, which is why it is under-reported. A high overturn rate means the original determinations were wrong at a measurable rate, and no amount of throughput improvement makes that acceptable. Break it down by reviewer, by criteria set and by service type; the pattern is usually concentrated rather than diffuse.

Track pends for information as its own outcome. A request that pends and is never resolved is functionally a denial, and it does not appear in the denial statistics. Measure how many pends close as approvals, how many close as denials, and how many simply expire. That third number is the one to look at first.

Measure reviewer agreement. Give the same anonymized cases to multiple reviewers periodically and measure how often they reach the same determination. Low agreement is a criteria clarity problem, not a staffing problem, and it is the root cause behind a lot of appeal overturns.

What to do about the interfaces

Standardized programmatic submission and response for authorization requests is coming on a schedule for several categories of payer, and it changes the operating picture in a way worth planning for rather than reacting to. Requests arrive structured. Status becomes queryable. Response times become observable to the other side. Confirm which requirements apply to which of your lines of business and by when — the scope has exclusions and the dates have moved.

What the shift means practically is that intake stops being the place where information is lost. A structured submission carries the codes, the dates and often the supporting documentation in a form your rules can evaluate on arrival. That makes the deterministic-approval path much more effective than it is in a fax-based world, and it makes your response latency visible in a way that will drive behaviour.

It also means that a plan whose internal state is not clean will have that fact exposed. If your system cannot say reliably what stage a request is at, an interface that answers status queries will report that confusion to providers continuously. Fix the internal state model first. It is the least glamorous item on any roadmap and it is the one that determines how the rest goes.

Where you do not need us

If your overturn rate on appeal is high, that is not an automation problem and buying software will not touch it. It means determinations are being made inconsistently or against criteria that are not being applied uniformly, and the fix is criteria clarity, reviewer calibration and a second-look process before adverse notices go out. That work is done by clinical leadership with a spreadsheet and a monthly meeting.

And if you have never computed approval rates by procedure code, do that first, in-house, this month. It takes a query. It will tell you what fraction of your review capacity is being spent on decisions that were never in doubt, and that number will shape every subsequent conversation about tooling — including whether you need any.

What we see go wrong

• Automation on the denial side, in any form, including a default adverse outcome on timeout
• Generated clinical rationales presented to a reviewer working against a clock
• Turnaround measured from case open rather than from receipt
• Averages reported where the tail is what creates the exposure
• Pends that expire, counted as neither approval nor denial
• A single ranked queue instead of separate queues with separate owners
• Rules that drift from the licensed criteria they were built to implement
• No reviewer agreement measurement, so inconsistency is invisible until appeal

What good looks like

• Written policy that automation may clear toward approval and never toward denial
• Approval rates computed by code, place of service and provider, and acted on
• Deterministic criteria checks traceable to the licensed criteria version
• Extraction that cites the page and shows the passage
• Missing-documentation detection that triggers a request within the hour
• Separate queues for documents, information requests, benefits and clinical judgment
• A fast path for cases where delay carries clinical consequence
• Turnaround reported as a distribution, timed from receipt
• Overturn rate broken out by reviewer, criteria set and service type
• Periodic blinded reviewer-agreement measurement

Bottom line

Utilization review is being asked to go faster, explain itself better, and publish its numbers. The way to meet that is to stop spending clinical judgment on cases that never needed it, to get the supporting facts out of the attachments and in front of the reviewer with citations, and to ask for missing documentation on the first day instead of the fifth. Automate toward approval, never toward denial, and route by what is missing rather than by a score. Then measure the tail of the clock and the overturn rate, because those two numbers are what an outsider will eventually look at.

Frequently asked questions