PPO vs Fee-for-Service: The Participation Mix That Decides What Clinical AI Can Recover

A general practice billing $1.4 million in annual production with roughly 80 percent of its patients on PPO plans does not collect $1.4 million. After contractual adjustments running approximately 25 to 35 percent across the major carriers, it collects closer to $1.05 million, and every new dollar of production it adds arrives pre-discounted at that same rate.

None of that is news to anyone who has read a month-end adjustment report. What is new is that the same arithmetic now sets the ceiling on what a clinical AI subscription can return, and nearly every vendor ROI model handed to a practice ignores it.

The pitch is denominated in diagnosed production: the detection model finds more, the perio model stages more, the practice diagnoses more. Your fee schedule decides what fraction of that survives to the deposit.

Why Diagnosed Production Is The Wrong Denominator

Vendors quote lift against diagnosis because diagnosis is the only variable their model touches. A radiograph model that surfaces interproximal lesions earlier is measured, honestly enough, in additional D2391 and D2392 findings per hundred bitewing series.

However, the chain from that finding to a deposit has four more links, and each one discounts the last. The lesion has to be presented, the case has to be accepted, the procedure has to be delivered, and the claim has to be paid at your contracted rate rather than your posted fee.

Your case acceptance rate governs link two, and your participation mix governs link four. As a result, a practice that improves diagnosis by 8 percent while collecting 74 cents on the scheduled dollar has improved collections by under 6 percent, before it pays for a single crown coping.

This is the same failure mode that produces optimistic dental AI ROI projections generally. The model is not wrong about the clinical finding; it is wrong about the denominator.

What Your Participation Mix Actually Is

Participation mix is the share of production, measured in dollars, flowing through each contracted fee schedule, weighted by that schedule's own write-off against your UCR fees. It is not the count of plans you are credentialed with, and it is not the percentage of patients carrying a card.

Pulling it accurately takes about an hour in Dentrix, Eaglesoft, or Open Dental and one adjustment report. Here is what you need on the table before any AI evaluation is meaningful:

• Production by carrier. Trailing twelve months of gross production split by primary carrier, in dollars rather than patient counts. Two hundred Delta patients coming in for prophies matter far less than forty MetLife patients in active restorative.
• Write-off by carrier. Contractual adjustments as a percentage of full-fee production for each individual plan. Ranges of approximately 18 to 45 percent are normal, and the spread inside a single office is usually wider than owners expect.
• Write-off by code family. The same carrier rarely discounts D0274 bitewings, D2392 posterior composites, and D4341 scaling and root planing at anything like the same rate.
• Out-of-network production. Dollars billed at full fee, whether the patient is uninsured or the practice is out of network and the patient assigns benefits.

Those four inputs produce a weighted realization rate, which is the only figure that belongs in an AI business case. Practices already tracking fee schedule leakage will have most of it assembled, since it is the same data pull with a different summary column.

The Same Acceptance Lift, Modeled Twice

Take two general practices with identical clinical profiles, identical staffing, and identical diagnosed treatment of approximately $900,000 a year at full fee. Both accept 62 percent of what they diagnose, both add the same clinical AI stack at $1,850 a month, and both realize a four-point acceptance lift from it.

The only difference between them is participation. Practice A runs 80 percent PPO and Practice B runs 30 percent PPO, both carry an average PPO write-off of 32 percent, and both incur variable delivery cost — lab, materials, and incremental chair time — of roughly 22 percent of the full fee.

Line item	Practice A — 80% PPO	Practice B — 30% PPO
Annual diagnosed treatment (full fee)	$900,000	$900,000
Case acceptance, before and after	62% to 66%	62% to 66%
Incremental accepted treatment	$36,000	$36,000
Weighted realization rate	74.4%	90.4%
Incremental collected	$26,784	$32,544
Variable delivery cost (22% of full fee)	$7,920	$7,920
Contribution	$18,864	$24,624
Clinical AI subscription	$22,200	$22,200
Net, year one	-$3,336	+$2,424
Break-even acceptance lift required	4.7 points	3.6 points

Same model, same clinicians, same clinical lift, and the sign flips. In fact, the entire $5,760 swing is nothing more than the 16-point realization gap applied to $36,000 of incremental accepted treatment.

Note that the break-even row is the one to write down. Practice A has to find nearly a point and a half more acceptance than Practice B to reach the same zero, which is a meaningful difference in how hard the model has to work.

Why The Write-Off Lands Entirely On Margin

The reason that gap is so punishing is that a contractual adjustment discounts revenue and nothing else. Your lab bills the same $165 for the coping whether the claim pays at your $1,300 posted fee or at a contracted $884.

Impression material, burs, sterilization cycles, assistant time, and the operatory hour do not renegotiate with Cigna. Accordingly, every point of write-off lands on contribution margin, which is how a 32 percent discount on revenue becomes something closer to a 55 percent discount on the profit of that procedure.

That also means incremental production is not automatically good production. If the model surfaces treatment that consumes your last unclaimed operatory hours at your worst-realizing plan, you have converted hygiene production per hour into busywork with a lab bill attached.

Keep in mind that realization is a separate question from collection speed. If you are already carrying elevated A/R days, the incremental dollars land later than any twelve-month model implies, and the subscription bills monthly regardless.

Which Codes The Model Surfaces Changes The Math

Weighted realization at the practice level is a starting approximation. It gets considerably sharper once you ask which codes a given model actually produces, because those are the schedule lines you will be billing against.

The mapping runs roughly like this:

• Caries detection models. AI caries detection mostly surfaces D2391 through D2394 posterior composites, which are the codes most frequently subject to an alternate benefit downgrade to the amalgam fee.
• Periodontal staging models. AI periodontal screening drives D4341 and D4342, where carriers apply quadrant limits, multi-year frequency caps, and pocket-depth documentation requirements before they pay anything at all.
• Radiograph review models. AI radiograph analysis surfaces periapical findings that route into endodontics and crown-and-bridge, generally the steepest discounts in a PPO schedule and the largest patient portion.
• Implant and airway models. These frequently land on codes a dental plan excludes outright, which makes the resulting treatment close to pure fee-for-service revenue no matter what your participation mix looks like.

Note that the last row inverts the entire argument. A heavily PPO practice can still see strong returns from a model whose findings route to non-covered treatment, because the write-off simply never applies to that production.

The Counterweight: Fee-For-Service Acceptance Starts Lower

Reading only the realization column would suggest that fee-for-service practices should buy every clinical AI product on the market. That conclusion does not survive contact with the acceptance data.

When a patient absorbs the entire fee rather than a 20 or 50 percent coinsurance, presented treatment converts at a materially lower rate. It is common to see baseline acceptance eight to twelve points lower in a predominantly fee-for-service practice on clinically identical diagnoses.

Rerun Practice B with a 54 percent baseline and the same relative improvement, and the four-point lift becomes roughly 3.5 points. That yields about $31,500 in incremental accepted treatment and approximately $21,546 in contribution, which lands about $654 short of the same $22,200 subscription.

The fee-for-service advantage does not disappear, but it is smaller than the realization gap alone implies. Both effects belong in the model, or you are running a one-variable argument against a two-variable problem.

How Do You Price This Before You Sign?

The honest answer is that you measure it in your own chairs instead of accepting a projection built on someone else's patient panel. Shadow mode is how every other production system gets validated, and there is no reason clinical AI should be exempt from it.

A defensible ninety-day evaluation looks like this:

1. Run the model with output hidden. Findings write to an audit trail and clinicians never see them during the window, so treatment planning stays unchanged and your baseline stays clean.
2. Count what it surfaced that you did not. Reconcile model findings against what was actually diagnosed in the same period, code by code rather than in aggregate.
3. Price the delta through your contracted schedules. Not your UCR fees. Apply each patient's actual plan, including alternate benefit downgrades and frequency limitations.
4. Apply your real acceptance rate by code family. Crown acceptance and composite acceptance are not the same number, and blending them will hide the result you are trying to measure.
5. Subtract variable delivery cost and the subscription. Then compare what remains against the break-even lift you calculated from your participation mix.

Pin the model version for the full window, since a vendor-side update mid-evaluation quietly invalidates the comparison. Moreover, a clinical AI evaluation suite running against a held-out set of your own radiographs is worth standing up before the window opens rather than after it closes.

Be aware that eligibility accuracy contaminates this measurement too. If the front office is quoting benefits from stale data, the acceptance rate you record is measuring your verification process as much as the model, which is a good argument for settling AI insurance verification before you evaluate anything clinical.

The BAA Comes Before The Business Case

None of this arithmetic matters if the vendor cannot execute a business associate agreement. Any system ingesting radiographs, periodontal charting, or narrative notes is handling protected health information and is a business associate under HIPAA, with no exception for the fact that inference happens on someone else's GPUs.

Confirm the BAA, the subprocessor list, PHI retention and deletion terms, whether your images are used for model training, and where inference physically runs. Practices deploying on Amazon Bedrock inside their own AWS account with CloudTrail logging and KMS-managed keys have a materially easier audit story than those posting images to a vendor endpoint under terms of service alone, and our note on HIPAA-compliant clinical AI in dental practices walks through the specific controls.

Settle that question first. A model with excellent sensitivity and no executed BAA is not the cheaper option; it is the unbuyable one.

What To Compute Before The Renewal Date

The number that decides this purchase is not the model's sensitivity or its published accuracy against a public dataset. It is your weighted realization rate, your break-even acceptance lift, and whether the codes the model produces are the codes your carriers happen to pay closest to full fee.

Practices at 70 percent PPO and above should expect to need a larger clinical lift than the vendor case study describes, and should weight models whose findings route toward non-covered or lightly discounted treatment. Practices under 40 percent PPO have more headroom on realization, but should verify their baseline acceptance against their active patient count before assuming the advantage holds.

If you are scoping a clinical AI deployment and want a second set of eyes on the economics, the NexV team builds and operates HIPAA-grade clinical AI in production dental environments every week. Reach out for a working session: we will pull your participation mix into a realization model, compute your break-even acceptance lift by code family, and leave you with a ninety-day shadow-mode evaluation plan you can run before you sign anything.