Which Prior Auth Work Can Come In-House? The 4-Column Payer Grid for RCM Leaders

Which Prior Auth Work Can Come In-House? The 4-Column Payer Grid for RCM Leaders

Which prior auth work can come in-house depends on the payer, not the function. Score payers on contract, rules, volume, and AI readiness to see which you own.

Published

Last Modified

Topic

AI Vendor Selection

Author

Jill Davis, Content Writer

TLDR: Deciding which prior auth work can come in-house is a payer-by-payer decision, not a vendor-wide one. Score every payer on four columns (contract constraints, provider-submission rules, volume, and AI readiness) and bring in-house first the payers whose rules are stable and whose denials get overturned most often, not the payers with the most volume. Keep with the vendor whatever your own data cannot yet support.

Best For: VPs and Directors of Revenue Cycle or Payer Operations at US provider, lab, and diagnostics enterprises (1,000 to 15,000 employees) who have an outsourcing contract coming up for review and need a defensible sequence for what returns in-house.

A payer-by-payer in-house decision is a sourcing method that scores each payer's prior authorization work separately, instead of treating prior authorization as one function that is either outsourced or not. The method exists because payers do not behave alike: KFF's 2024 Medicare Advantage data shows denial rates ranging from 4.2% at one insurer to 12.8% at another, and appeal overturn rates from 51% to 95.5%. Work that is cheap to run in-house for one payer can be a loss for the next. The four-column grid below makes the call one payer at a time, with a scoring rule and a sequencing rule, and it is blunt about what AI cannot do for any payer until the data and the clinical documentation are in place.

Why which prior auth work can come in-house depends on the payer, not the function

Which prior auth work can come in-house depends on the payer because the cost of a prior authorization is set by that payer's contract terms, submission rules, request volume, and how predictable its decisions are. Outsourcing contracts hide this variation behind a single per-case or per-FTE price. Once the work is scored per payer, some of it is obviously cheaper to own and some is obviously not.

The variation is large and it is public. In KFF's analysis of 2024 Medicare Advantage data, insurers made 52.8 million prior authorization determinations, denied 7.7% of them, and overturned 80.7% of the denials that were appealed. Behind those averages, one insurer received 0.6 requests per enrollee and another 3.0; one denied 4.2% of requests and another 12.8%. The AMA's 2025 physician survey found the same spread from the provider side: 75% of physicians rated one national insurer's prior authorization burden as high or extremely high, against 56% for the lowest-rated plan on the list.

The per-case price flattens the payer differences

A vendor's per-case fee treats a routine imaging authorization with a stable payer and a specialty drug authorization with a payer that changes its criteria quarterly as the same unit of work. They are not. The 2025 CAQH Index put fully electronic prior authorization adoption at 40% of transactions, up from 31% two years earlier, which means a majority of the work still runs through portals, fax, or phone, and the mix differs by payer. An MGMA poll from March 2026 found 61% of practices had staff logging into seven or more payer portals every week. The cost of owning that work depends entirely on which seven.

What "come in-house" actually means for one payer

Bringing a payer's prior authorization in-house means your own staff, on your own data, submit the request, track the decision, and manage the appeal for that payer's members. It can be partial. A common pattern is to keep the vendor on the initial submission for a payer with a complicated portal while taking back the appeals, because appeals are where clinical judgment and institutional memory about that payer's reviewers matter most. The earlier post on how the outsourcing math changed with AI made the case for why the question is open at all. This post is about how to answer it one payer at a time.

Why the vendor-wide decision is the wrong unit

Most provider organizations already run a mixed model without calling it that. The 2026 Guidehouse and HFMA revenue cycle survey found 69% of providers outsource all or part of the revenue cycle, and among those, only 39% outsource denials management while 67% outsource accounts receivable follow-up. The function is already split by task. Splitting it by payer is the next cut, and it is the one that exposes where the money and the risk actually sit.

The 4-Column Payer Grid

The 4-Column Payer Grid is a decision table with one row per payer (or per payer and service line) and four scored columns: contract constraints, provider-submission rules, volume, and AI readiness. Each column is scored from 0 to 3 using data the RCM team already holds, and the row total decides whether that payer's prior authorization work comes in-house now, later, or stays with the vendor.

The grid works because each column answers a different blocking question. Contract constraints ask whether you are allowed to take the work. Provider-submission rules ask whether the payer will accept it from you. Volume asks whether it is worth staffing. AI readiness asks whether your data can support any automation for that payer. A payer can score high on three and still be a bad candidate if the fourth is a zero.

Column

Question it answers

Score 0

Score 3

Data source

Contract constraints

Can we take this work without breaching the vendor or payer contract?

Vendor contract has an exclusivity clause or minimum volume covering this payer; payer contract requires a delegated entity

No exclusivity, no minimum, notice period under 90 days; payer contract silent on who submits

Vendor MSA, payer participation agreement

Provider-submission rules

Will the payer accept and process requests from our staff as efficiently as from the vendor?

Payer requires a specific clearinghouse or portal credential tied to the vendor; no electronic path

Payer supports the electronic prior authorization standard and accepts provider submissions directly; turnaround data available

Payer portal documentation, CMS-0057-F metrics, 278 transaction logs

Volume

Is there enough work to staff and keep competent?

Under 50 requests a month; too few for a trained reviewer to stay current

Over 500 a month, or a concentration in two or three service lines with repeatable criteria

Internal authorization log, denial master table

AI readiness

Can our data support automation for this payer?

No validated denial history; clinical documentation not retrievable at request time; criteria change without notice

Three or more years of validated denials and appeals, documents retrievable by order, published and stable medical policy

RCM data foundation tables, payer policy archive

Column 1: contract constraints

Contract constraints score the legal and commercial room you have to move the work. Read the vendor agreement for exclusivity, volume minimums, and termination notice, and read each payer agreement for any clause on who may submit on the provider's behalf. The CMS-0057-F final rule does not change these contracts, but from January 1, 2026 it requires Medicare Advantage, Medicaid, and CHIP plans to decide standard requests within seven calendar days and expedited ones within 72 hours, and from January 1, 2027 to expose a Prior Authorization API. A payer that is building that API is a payer whose submission path will soon be open to your own systems.

Column 2: provider-submission rules

Provider-submission rules score how a payer treats requests that come from you rather than from a vendor with a dedicated channel. Some payers run the same queue regardless of sender. Others give delegated vendors a faster lane or a dedicated reviewer. CMS-0057-F also required impacted payers to publish prior authorization metrics on their websites by March 31, 2026, which gives you turnaround and approval rates by payer for the first time without relying on vendor reports. An MGMA Stat poll from September 2026 found 44% of medical group leaders said prior authorization was slower than the year before and only 7% said faster, so assume the published numbers will be the best case and measure your own.

Column 3: volume

Volume scores whether there is enough work for an in-house reviewer to stay competent on that payer's criteria. The AMA survey found physicians complete an average of 40 prior authorizations a week and that 40% of practices employ staff who do nothing else. At enterprise scale the question is not whether you have the staff, it is whether a given payer's volume justifies a reviewer who knows that payer's medical policy cold. Premier's 2025 claims study found 30.5% of Medicare Advantage claims required prior authorization in 2023, so for most provider enterprises the Medicare Advantage payers will clear the volume bar first. A low-volume commercial payer with unusual rules usually should not.

Column 4: AI readiness, scored per payer

AI readiness scores whether your own data for that payer can support automation at all. The honest answer for most organizations is "for some payers, not others." Automation of documentation matching or appeal drafting needs a validated history of what that payer approved, denied, and overturned, and it needs the clinical documents to be retrievable at the moment the request is built. The RCM data foundation post describes the master tables that make this possible. If the denial table for a payer is incomplete or the vendor holds it, that payer scores a zero in this column no matter how attractive the volume looks. Experian Health's 2025 State of Claims survey found missing or inaccurate data was the top cause of denials at 50% of respondents, ahead of authorizations at 35%. Automating against bad data produces faster bad submissions.

Payer-by-payer grid vs. an all-or-nothing outsourcing decision

A payer-by-payer grid differs from an all-or-nothing outsourcing decision in what it optimizes and what it risks. The all-or-nothing decision optimizes contract simplicity and risks cash flow, because every payer transitions on the same date whether or not the team is ready. The grid optimizes the order of transition and risks complexity, because the organization runs a mixed model for a period. For an enterprise with a dozen material payers, the grid is the safer bet.

Dimension

All-or-nothing decision

Payer-by-payer grid

Unit of decision

The whole prior authorization function

One payer, or one payer and service line

Evidence required

Aggregate vendor cost vs. estimated in-house cost

Per-payer contract, submission, volume, and data readiness scores

Transition risk

Concentrated on one cutover date

Spread across a sequence; each step is reversible

Cash-flow exposure

High during cutover; lag on every payer at once

Limited to the payer in transition

Where AI fits

Assumed to apply uniformly

Scored per payer; applied only where data supports it

Vendor relationship

Ends or continues wholesale

Narrows to the payers that score low

Board narrative

"We brought prior auth in-house"

"We own the payers where we win; the vendor keeps the rest for now"

When the all-or-nothing decision is still right

An all-or-nothing transition makes sense when the vendor contract has a single termination date with no partial-exit clause, when payer volume is concentrated in one or two plans, or when the vendor's performance has deteriorated to the point where the cost of keeping any of it exceeds the cost of a hard cutover. It also makes sense for a small organization that cannot run two models at once. For an enterprise RCM function with a mixed payer book, those conditions are rare.

What the grid gives the CFO that the aggregate number does not

The grid gives the CFO a sequence with a reason attached to every step, and a measure that can be checked after each one. The Guidehouse and HFMA survey found the share of providers reporting final denial rates above 5% nearly doubled to 20% between its last two editions, and 88% of executives put prior authorization delays among their top three payer concerns. A finance leader looking at those numbers wants to know which payers drive them. The grid is the answer, and the RCM leader who built it from the organization's own data is the one who can defend it.

How to score which prior auth work can come in-house

Scoring which prior auth work can come in-house is a four-step exercise: build the payer rows from your authorization and denial logs, score each column from 0 to 3 using the data sources in the grid, apply a gating rule that disqualifies any payer with a zero in contract constraints or AI readiness, and rank the remaining payers by total score with overturn rate as the tiebreaker. The output is a sequenced list, not a single yes or no.

Step 1: build the rows from your own data

Start from the authorization log and the denial master table, not from the vendor's report. Each row is a payer, split by service line where the criteria differ materially (imaging, infusion, genetic testing, and surgery usually need their own rows). Record twelve months of request volume, denial count, appeal count, and overturn count per row. If any of those counts comes only from the vendor, flag the row; you will need the vendor to hand over the data before you can score it, and that handover is itself a contract question for column one.

Step 2: score the four columns and apply the gating rule

Score each column from 0 to 3 using the definitions in the grid. Then apply the gating rule: any payer with a 0 in contract constraints or a 0 in AI readiness is out of this cycle, regardless of its other scores. The first gate is legal. The second is practical. Taking back a payer whose data you do not hold means rebuilding the history from scratch while the clock on the appeal deadlines runs. The OIG's 2022 audit found that 13% of Medicare Advantage prior authorization denials met Medicare coverage rules and should have been approved. Catching those requires the record of what that payer has approved before. Without it, there is nothing to match against.

Step 3: rank by total score, then by overturn rate

Rank the surviving payers by total score. Where two payers tie, the tiebreaker is the appeal overturn rate, highest first. This is the contrarian part of the method and it deserves a defense. The instinct is to start with the highest-volume payer because that is where the fee is largest. The better first candidate is the payer whose denials get overturned most often, because a high overturn rate means the payer's decisions are predictable and its errors are correctable by someone who knows the file. KFF's data shows overturn rates in Medicare Advantage appeals ranging from 51% to 95.5% by insurer. The 95.5% payer is a payer whose prior authorization work rewards institutional memory. That is the work to own. A payer that denies 12.8% of requests and overturns few of them on appeal is a payer whose decisions are driven by criteria you cannot see, and that work is better left with a vendor until the published metrics make the criteria legible.

For a provider or diagnostics enterprise where a single payer accounts for more than a quarter of authorized revenue and that payer's medical policy changes more than twice a year, a generic in-house transition plan is not enough; the answer is a dedicated payer program that does three things: maintains a versioned archive of that payer's published policy and the organization's own approval and denial history against each version, assigns a named reviewer who owns the appeal relationship with that payer's medical directors, and measures the overturn rate and days-to-decision for that payer separately from the rest of the book, so the decision to keep it in-house can be reversed on evidence rather than on frustration.

Step 4: sequence the transitions and set the exit test

Bring back one payer at a time, in rank order, and set the test before the transition starts. The test is two numbers compared against the vendor's last twelve months for that payer: days from order to decision, and overturn rate on appeal. If after one full quarter in-house both numbers are at or better than the vendor's, the next payer in rank order starts. If either is worse, stop and find out why before moving. McKinsey's analysis of prior authorization automation found that electronic submission completed more than 60% of requests in under two hours against none by phone or fax, so a payer whose electronic path you can use directly should show improvement on the first number quickly. If it does not, the problem is usually documentation retrieval, which is a column four issue the score should have caught.

What AI cannot do for prior authorization without clean data and clinical documentation

AI cannot decide which prior auth work can come in-house, and it cannot make a payer's work automatable when the data behind it is missing. AI can match an order to published criteria, flag a missing document, prioritize a worklist, and draft an appeal from a validated history. Each of those needs the policy, the document, and the history to exist in a form the organization controls. None of them fixes a chart whose supporting note was never written.

The data AI needs per payer, and who holds it

For each payer in the grid, the AI readiness column is really asking four questions: do you hold the approval and denial history, do you hold the appeal outcomes, can the clinical documents be retrieved by order number at the moment the request is built, and is the payer's medical policy archived with version dates? Where the vendor holds any of these, the AI readiness score is capped at 1 until the data is returned and validated. The 3-column map of what AI can and cannot fix in denials applies here: the right-hand column, the things AI cannot do, is mostly a data column.

The evidence that automation on bad inputs makes things worse

Two findings should temper any plan that leads with AI. The Experian 2025 survey found only 14% of providers were using AI in claims management, and 90% of denials were still reworked with human review before resubmission. A 2026 report from the Peterson Health Technology Institute found no evidence that administrative AI lowered average cost per claim once the cost of the tools was included. Neither finding says AI does not work. Both say that the gain depends on what it is pointed at, and a payer with a validated history and retrievable documents is a very different target from one without.

The clinical documentation problem is not an AI problem

The AMA survey found that only 24% of physicians believe medical necessity denials are consistently reviewed by qualified clinicians, and 32% report requests are often or always denied. A good share of those denials come back to documentation that was insufficient at the time of request, and no amount of automation downstream creates a clinical note that was never written. If a payer's denials cluster on documentation reasons, the fix is upstream at the point of order, with the clinician, before the request is built. That is a workflow redesign problem. It belongs in the sequence before the AI does.

Common objections from RCM leaders, and what to say to them

RCM leaders raise three objections to the payer-by-payer method, and each has a direct answer. The first is that a mixed model is too complex to run. The second is that the vendor will not cooperate with a partial exit. The third is that the overturn-rate tiebreaker rewards the wrong behavior. All three are fair. None of them holds up once the organization's own numbers are on the table.

"We cannot run two models at once"

Most enterprise RCM functions already run several. The Guidehouse and HFMA data shows that providers outsource some functions and keep others at rates that vary by task, and the Kaufman Hall outlook reported by HFMA found revenue cycle outsourcing rose from 32% of health systems in 2023 to 58% in 2025, which is a lot of organizations adding a vendor to part of a function they already ran. The grid does not create the mixed model. It makes the existing one explicit and gives it an order.

"The vendor will not hand over the data or accept a partial exit"

That is a contract question, and it is exactly what column one scores. If the vendor agreement does not provide for data return on partial termination, the payer rows that depend on vendor-held data score a zero in AI readiness and the organization learns, in advance, that the first negotiation is about data, not about scope. That is useful information to have before a renewal conversation rather than after.

"Starting with high-overturn payers just means we win the easy ones"

Yes. That is the point. The first in-house payer has to prove the model works to a CFO who was told, by someone, that automation would solve prior authorization. Winning on a payer whose decisions are predictable builds the evidence and the reviewer competence that the harder payers will need. An enterprise that starts with its most opaque payer and loses the first quarter does not get a second one. The 4-question test for evaluating an AI denials vendor has the same logic in the other direction: ask the vendor what happens to the request when the documentation is missing, because the payers where that happens most are the ones you should not be taking back yet.

This analysis was developed using methodologies and operating experience from Assembly.

Frequently Asked Questions

Which prior auth work can come in-house first?

The first prior authorization work to bring in-house is the payer whose rules are most stable and whose denials are overturned most often on appeal, not the payer with the most volume. KFF's 2024 data shows Medicare Advantage overturn rates from 51% to 95.5% by insurer. High overturn means predictable decisions, which reward a reviewer who knows the file.

What is the 4-Column Payer Grid?

The 4-Column Payer Grid is a decision table with one row per payer and four scored columns: contract constraints, provider-submission rules, volume, and AI readiness. Each column scores 0 to 3 from data the team already holds. A zero in contract constraints or AI readiness removes the payer; the rest rank by total score, overturn rate as tiebreaker.

Which prior authorization work should stay outsourced?

Prior authorization work should stay outsourced when the payer's criteria are opaque, the volume is too low to keep a reviewer competent, or the organization does not yet hold that payer's validated denial history. A payer that denies 12.8% of requests and overturns few appeals, as one insurer did in KFF's 2024 analysis, belongs with the vendor for now.

How do contract constraints affect the in-house decision?

Contract constraints decide whether an organization is allowed to take back a payer's work at all. Vendor agreements with exclusivity clauses, volume minimums, or no partial-termination provision block the move, and payer agreements can require a delegated submitter. In the grid, a zero on this column is a hard gate that no other score can override.

What volume justifies bringing a payer's prior auth in-house?

A payer justifies in-house prior authorization when it generates enough requests for a trained reviewer to stay current on its medical policy, typically several hundred a month or a concentration in two or three service lines. Premier's 2025 study found 30.5% of Medicare Advantage claims required prior authorization in 2023, so those payers usually clear the bar first.

How is AI readiness scored per payer?

AI readiness is scored per payer on whether the organization holds that payer's validated approval, denial, and appeal history, can retrieve clinical documents by order at request time, and has the payer's medical policy archived with version dates. Where the vendor holds any of that data, the score is capped until it is returned.

How does CMS-0057-F change which prior auth work can come in-house?

CMS-0057-F makes more payers' work eligible for in-house handling by opening electronic submission paths and publishing payer metrics. From January 1, 2026 impacted payers must decide standard requests within seven days and expedited ones within 72 hours, and from 2027 expose a Prior Authorization API, per the CMS fact sheet. Public metrics let you score submission rules from payer data.

What is the difference between a payer-by-payer grid and an all-or-nothing outsourcing decision?

A payer-by-payer grid sequences the transition one payer at a time, while an all-or-nothing decision moves the whole prior authorization function on one date. The grid spreads cash-flow risk, makes each step reversible, and applies AI only where a payer's data supports it. The all-or-nothing choice is simpler to contract but concentrates every transition risk at a single cutover.

Why rank payers by overturn rate instead of volume?

Overturn rate is a better tiebreaker than volume because it measures how predictable and correctable a payer's decisions are. A payer that overturns most denials on appeal is one whose errors a knowledgeable in-house reviewer can fix. High volume with an opaque payer produces more work without more wins, and the first in-house payer has to prove the model.

What can AI not do for prior authorization?

AI cannot create missing clinical documentation, cannot read criteria a payer has not published, and cannot learn from a denial history the organization does not hold. The AMA's 2025 survey found 32% of physicians say requests are often or always denied; many trace to documentation gaps at the point of order, a workflow problem upstream of any automation.

How many payers should transition in-house at once?

One payer should transition at a time, in rank order, with an exit test set before the move begins. The test compares days from order to decision and appeal overturn rate against the vendor's last twelve months for that payer. If both hold or improve after a full quarter, the next payer starts. If either is worse, the sequence pauses.

What data is needed to build the payer grid?

Building the payer grid needs twelve months of request volume, denials, appeals, and overturns per payer and service line, drawn from the organization's own authorization log and denial master table. It also needs the vendor agreement, each payer participation agreement, and the payer's published medical policy. The RCM data foundation describes the master tables that hold most of this.

Why do provider-submission rules matter for the decision?

Provider-submission rules matter because some payers process requests from a delegated vendor faster than requests from the provider's own staff. A payer that accepts direct electronic provider submissions scores high; one that ties credentials to a vendor's channel scores low. The 2025 CAQH Index found only 40% of prior authorizations were fully electronic, so the path varies by payer.

How should a lab or diagnostics company apply the grid differently?

A lab or diagnostics enterprise should add service line rows for genetic and molecular testing, because payer criteria there change more often than for imaging or infusion. These organizations also tend to start service before authorization clears, so days from order to decision should be tracked separately for tests already performed. Retrievable ordering documentation is usually the main blocker.

What is the biggest mistake in bringing prior authorization in-house?

The biggest mistake is taking back a payer's work before holding that payer's data. The OIG's 2022 audit found 13% of Medicare Advantage prior authorization denials met Medicare coverage rules. Catching those errors requires the history of what the payer approved before. Starting without it means losing the first quarter while rebuilding records.

How often should the payer grid be rescored?

The payer grid should be rescored every quarter and whenever a payer publishes new metrics, changes a medical policy, or a vendor contract hits a notice window. An MGMA poll from September 2026 found 44% of medical groups said prior authorization got slower in a year, so submission-rule scores move. Rescoring keeps the sequence tied to current conditions.

Your AI Transformation Partner.

Your AI Transformation Partner.

ยฉ 2026 Assembly, Inc.