The Peterson Health Technology Institute’s April 2026 report did not get the headline it deserved. PHTI convened senior leaders from health systems, health plans, and federal agencies, and the central finding was not a success story about administrative efficiency. It was a warning: administrative AI is reducing the cost of running individual transactions and simultaneously inflating total system costs. Both things are happening at once.
If your CFO is pointing to staff hours saved and calling it ROI, they are measuring the wrong variable.
The Problem: Your Dashboard Is Showing You the Wrong Number
The U.S. healthcare system wastes an estimated $350 billion annually on administrative overhead. Of that, $266 billion comes from administrative complexity alone [PHTI, April 2026]. Those numbers have driven every AI vendor pitch in the market over the past three years.
So health systems bought the tools. Prior authorization automation. AI-assisted coding paired with ambient scribes. Revenue cycle optimization platforms. The efficiency gains were real and immediate. Staff processed more PA submissions per shift. Coding captured more clinical complexity. Revenue per encounter climbed.
The ROI calculations looked clean at the departmental level: reduced cost per transaction, faster cycle times, fewer staff hours per claim. The metrics every CFO wanted to see on a vendor scorecard.
What was missing from those dashboards was what happens on the other side of the table.
The core problem with measuring administrative AI at the organizational level is that you are buying half an equation. You see your efficiency gains. You do not see the system-level activity those gains create. Because providers and payers are deploying AI against each other simultaneously, that system-level activity is not declining. It is multiplying.
The Insight: Competing Automation Makes Everyone Poorer
Here is the documented mechanism. When providers deployed PA submission bots, payers deployed their own AI to triage and deny submissions faster. The result is what PHTI workshop participants called “bot wars”: automated back-and-forth exchanges that increase the volume of communications per prior authorization without resolving the underlying clinical question [PHTI, April 2026].
Prior auth already costs providers $20 to $30 per submission cycle and health plans $40 to $50. AI made each individual submission cheaper to execute. But when both sides automate simultaneously, submission volume goes up, denials are generated faster, and appeals pile up in return. PHTI found no existing evidence that current deployments translate to lower average cost per claim when the AI tool’s own cost is factored in [PHTI, April 2026].
That is the core trap: the process metric (cost per submission) improved. The outcome metric (total spend per adjudicated claim) did not move.
The medical billing dynamic is the same problem from a different angle. Ambient scribes are now standard across large health systems. Every large system in PHTI’s research sample had adopted them. The primary documented result was not improved documentation accuracy or reduced clinician burden. It was increased billing intensity.
One system in the report saw a 5% increase in Level 5 encounters after deploying AI scribes, adding over $1,000 per provider per month in revenue [PHTI, April 2026]. Accurate? Possibly. Inflationary? Definitely. Health plans noticed. They responded with algorithmic downcoding and across-the-board reimbursement reductions. Systems that adopted ambient scribes early capture more. Payers cut broadly to compensate. Rural and community hospitals that have not adopted AI absorb those cuts without any offsetting revenue.
Part of the structural problem is regulatory. The CMS-0057-F rule mandated data standards for prior auth APIs but did not standardize the actual medical necessity criteria that vary across health plans [PHTI, April 2026]. You can have seamless data transmission and still face wildly different clinical thresholds depending on the insurer. Standardized pipes. Non-standardized content. The AI runs efficiently on those rails. The rails still go to different destinations.
“AI is reducing the cost for individual organizations to execute prior authorizations, but it has not reduced overall system-level costs.” [PHTI, April 2026]
Where Administrative AI Actually Works
The contrast case is pharmacy prior authorization, and it is instructive about what structural alignment looks like in practice.
In May 2026, Surescripts reported a significant expansion of its real-time PA automation platform: 68,000 prescribers across 42 health systems, 104 distinct medications, and a median approval time of 18 seconds for supported workflows [Surescripts, May 2026]. Early deployments showed an 88% reduction in appeals, a 68% reduction in denials from missing information, and a 34% automated approval rate for in-scope medications.
That is not a bot war outcome. It is a structurally different model.
Surescripts queries payer benefit data at the moment of prescribing and returns a determination in real time because the criteria are transparent, the data standards are agreed upon between sender and receiver, and both sides are working from the same clinical information before any claim is ever submitted. The provider is not submitting into a black box and waiting for a denial. The payer is not defending against a flood of automated requests. The AI is reducing submission volume, not multiplying it.
The difference between this model and a typical hospital PA automation tool is not technical sophistication. It is incentive structure. When both parties share data upfront and the AI surfaces the answer at the point of clinical decision, there is nothing to fight over. When the AI is deployed to win a submission game, both sides accelerate and costs compound.
Prior authorization automation that integrates payer benefit data at the prescribing moment contracts the problem. PA automation that makes your submissions faster and cheaper without changing the adversarial structure scales the volume of a bad system.
The Deployment Audit: Arms Race vs. Structural Alignment
Before renewing or expanding any administrative AI contract, map each tool against this framework:
| AI Use Case | What Most Orgs Measure | System-Level Reality | Risk Rating |
|---|---|---|---|
| PA submission bots (provider side) | Staff hours saved per submission | Bot wars; payer denial volume rises; total cost per resolved claim unchanged or increasing | High |
| Ambient scribes for billing | Revenue capture per encounter | Coding intensity up; payer responds with algorithmic downcoding; rural hospitals absorb cuts without offsetting revenue | High |
| Real-time PA at the prescribing moment | Approval speed; denial rate | 18-second approvals where payer data standards exist; reduces volume and friction simultaneously | Low |
| AI chart abstraction (internal only) | Time per abstraction | Efficient where it does not feed back into billing inflation; no adversarial counterpart in the loop | Moderate |
| AI PA triage on the payer side | Claim processing speed | Faster denials; providers escalate with submission bots; volume multiplies on both sides | High |
The key column is the third. If your AI is engaging with a counterpart that is also running AI, and neither side is sharing data before the transaction begins, you are in an arms race. The efficiency gains at the departmental level are real. The system-level cost trajectory is not improving.
The Bottom Line
The ROI numbers most health systems track are process metrics: staff hours saved, submission speed. The outcome metric is cost per claim fully adjudicated, inclusive of vendor contract cost, appeals handling, and payer reimbursement adjustments. Systems whose frameworks do not capture those figures are reading a vendor’s dashboard, and most of them are.
The dividing line on any PA automation tool is whether it reduces submission volume or just makes it cheaper to submit more. The Surescripts model reduces volume by surfacing benefit criteria at the point of prescribing. Most hospital PA bots reduce cost per submission by automating a process that still generates a fight. One contracts the problem; the other scales it.
The exposure runs past the billing department. When a system captures Level 5 encounters at higher rates from ambient scribes, payer retaliation through downcoding is algorithmic and non-targeted: it hits every provider in the plan’s network, including the rural and critical access hospitals that have not deployed AI and absorb the cuts without offsetting revenue. The systems whose pharmacy directors and CMOs sit out payer contract discussions, without data on how AI-driven coding intensity is landing across the broader provider mix, are the ones negotiating blind.