ASCERTA

Clinical reasoning,
at infinite scale.

Ascerta is the trust infrastructure for clinical AI: the governance that proves
when an AI system can be trusted, and the clinical models that power it.

Try the live demo at ascertahealth.com/demoAccess code 0313
What Ascerta is

Healthcare is moving to clinical AI.
Ascerta is the layer that makes it trustworthy.

Ascerta is governance infrastructure for healthcare AI, powered by our own clinical models and the most extensive research program on clinical AI behavior.

Govern

Governance infrastructure.

Model-agnostic evaluation, live monitoring and analysis of every decision in real time, and a signed provenance record behind every decision, delivering continuous, on-demand compliance.

Clinical models

Trained for the work.

Our own models are competitive with frontier models and excel in clinical extraction, codification, criteria reasoning and prior-authorization optimization. Deployed in VPC, on-premise, or air-gapped.

We evaluate and govern every determination to ensure it is
GroundedCorrectCompleteCalibratedFairFaithful

Point of entry: prior authorization, on both sides.  →  Trajectory: every consequential decision in healthcare.

The problem

A new class of invisible failure.

01

Software failed loudly.

Errors were visible and deterministic.
Find the fault and fix it.

02

Early AI was inaccurate.

Less capable and often wrong, but visibly so. It could be measured and corrected.

03

Frontier AI fails in ways that look like success.

It can game the objective, satisfy the rule while defeating its intent, and produce a convincing explanation that was not what drove the decision. At the frontier, models can circumvent constraints, evade monitors, and behave differently when they know they are being evaluated.

Even the frontier labs cannot reliably reconstruct why a model made a specific decision. When the reasoning behind an output cannot be inspected, trust must come from independently testing behavior and preserving exactly what happened.
Why now

The slowest industry to adopt new technology
is now adopting AI faster than anything before it.

Healthcare moves deliberately, as it should when the stakes are lives. But it never built the trust and governance layer other industries did alongside their last technology wave, so AI is now arriving into an empty middle: capable models, no infrastructure to trust them.

81%
38% in 2023
of US physicians use AI in practice, up from 38% in 2023
AMA, 2026
100M
Payer and provider workflows
US prior authorization decisions a year, now AI-mediated on both sides
US annual volume
$90B+
FY2025
improper payments across CMS programs, FY2025
CMS

Capability has outrun governance.

71% of organizations are not expanding AI at pace, even where it has delivered measurable value, and leaders reject autonomous, black-box AI for high-stakes clinical decisions (Carta Healthcare, 2026). Public trust in AI-assisted care fell from 52% to 42% in two years (Ohio State, 2026).

AI has the ability to radically transform healthcare once it can be governed at scale. Ascerta is the trust layer that makes that transformation possible.

The platform

The single trust platform for AI across healthcare.
Starting with prior authorization, on both sides.

One hundred million US decisions a year, and the smallest surface we will touch. One decision, two sides, AI on both, and each side pays when it is wrong: the payer in exposure and improper payment, the provider in lost revenue, the patient in the care they’re owed. One governance engine spans every surface, so an organization governs its AI once, across every workflow.

Now

Coverage & payment

  • Prior authorizationMVP
  • Utilization management
  • Payment integrity
  • Risk adjustment
Expansion

Clinical operations

  • Evidence synthesis
  • Chart review
  • Care management
  • Documentation
Expansion

Clinical reasoning

  • Detection
  • Monitoring
  • Decision support
  • Diagnosis
Expansion

Research & discovery

  • Cohort discovery
  • Evidence review
  • Trial workflows
  • Safety signals

Prior authorization is in MVP. Utilization management, payment integrity and risk adjustment follow on the same engine, then clinical operations, clinical reasoning, and research.

A robust vendor market exists on both the payer and provider sides, but almost every solution targets one side rather than bridging the two (Health Affairs, January 2026). One engine spans both, with a hard information wall between payer and provider deployments.

Why prior authorization, now

Prior authorization is where AI accountability arrives first.

More electronically visible, more regulated and more litigated than any other decision in healthcare. Plans must now show how an AI-influenced determination was made.

01

The mandate

CMS-0057 sets decision windows of 72 hours and 7 days, in force since January. Since March 31, 2026, plans must publicly report approval rates, denial rates and approval-after-appeal rates. Electronic exchange follows January 1, 2027.

02

The oversight

Federal CMS initiatives on AI in coverage review, 250 AI bills across 47 states, the NAIC model bulletin adopted by state insurance regulators, and state utilization-review statutes. Any AI in a PA workflow must be transparent, auditable and fully documented.

03

The liability

Record settlements, including $556M in a single case, surging multi-front litigation, and a stated federal principle that AI does not shift liability away from the plan or the provider.

04

The gap

94% of payers are deploying AI, fewer than half at scale. Even a single fact dropped upstream corrupts a deterministic decision downstream. Four in five appealed denials are overturned. Overturn rates are now public and comparable across plans.

Decision windows shrinking. Volume rising. The results are soon going public.
For the payer

Make every determination financially sound — and ready to prove.

Ascerta gives the plan two operating systems: governance for better decisions and stronger economics; compliance that continuously assembles the proof behind every one.

01 / Governance

Protect revenue at the point of decision.

Every determination is correct, complete, fair, policy-grounded and provable.

Higher automationVerify more decisions and safely raise auto-adjudication.
Fewer improper paymentsCatch wrongful approvals before money leaves the plan.
Lower appeal costGenerate the evidence pack automatically and cut manual review time.
Live determination monitor 12,481 scored today
AUTH-2026-0165Wrongful denialHigh
AUTH-2026-0192Wrongful approvalHigh
Matched-pair evidence isolatedOpen investigation →
02 / Compliance

Continuous compliance, not audit-season archaeology.

Every decision reconstructable, observable and backed by its complete record.

Automatic evidence + appeal packsFull observability + signed provenanceCMS-0057 + FHIR Da Vinci alignedAPS-conformant decision recordsCompliance documentation on demandTrust portal + scoped audit access
CMS program audit · Q3READY
SCOPE1,284 determinationsCardiology · 07/01—09/30
RECORDS100% reconstructableEvidence, rationale, policy + action
Decision universe.csvAPS records.pdfMetrics workbook.xlsx
Access expires Oct 31Open scoped portal →

See the live payer systemascertahealth.com/demo · access code 0313

Enterprise platform priced on covered lives and decision volume.
For the provider

Prior auth submissions are written for human reviewers.
The reader is increasingly an AI model.

A submission can be clinically complete and still be denied, because the deciding fact never survived the payer’s pipeline.
We rebuild the case against how that pipeline actually reads it.

Why we know where to intervene

Our research characterizes how these systems fail, and how payer models get it wrong. Those failures are specific, repeatable and engineerable — we close every one of them before the packet is sent.

Why the standards do not close it

CMS-0057 and Da Vinci standardize the transaction. The proof still arrives as unstructured notes and faxed records, so standardizing the envelope is not a solve.

Specialty models

One trained per job: extraction, structuring, coding, adversarial review and more. Narrow and tuned, they often outperform frontier models on these specialty tasks — and deploy in your VPC or on premises.

Precomputation

Duration across notes, sequence and causality are resolved in advance, so nothing rests on the payer's model making the connection.

Placement

Position, structure and phrasing decide whether a fact reaches the rule at all. Decisive evidence is placed against the measured behavior of the systems reading it.

Crosswalk

Every fact is bound to the criterion it satisfies, as a DTR QuestionnaireResponse and highlighted supporting clinical information on the PAS claim.

Clinical citation

Every fact cites the document, page and sentence it came from. Nothing has to be taken on trust or searched for.

Adversarial review

A second model attacks the submission before it goes out and finds what a reviewer would use against it. Holes are filled in advance, not during a months-long appeal.

Intelligence Center

Every case makes the next one stronger. We build the provider’s own self-improving intelligence dataset.

Captured by configuration, not by payer. Which patient profile, treatment, indication and stage, alongside the documentation sent and the policy version in force. A plan name alone is too coarse to submit against.
Turnaround becomes predictable. Time to decision by payer and configuration, so authorization stops being an unknown in scheduling and revenue projections.
Appeals analyzed for what worked. Which argument overturned it, which evidence closed it, which framing failed. The overturn rate is the scoreboard; the reason is what improves the next one.
Built only from your own cases. The plans you bill, the treatments you deliver, the patients you see. Not a general model of how payers act, which is not one you can submit against.
Separate by design.

Every provider’s research and intelligence is theirs alone, walled off from other providers and from our payer-side work.

Live demo — access code 0313ascertahealth.com/demo
Ascerta research

Published research on AI behavior inside healthcare.

3M+
AI clinical decisions evaluated, matched-pair and causally structured
300K+
released publicly, among the largest studies of its kind
Mechanistic
interpretability that reaches inside the model, not only its outputs

Our own clinical models, trained and tuned on this corpus, are competitive with frontier models and excel in clinical extraction, codification, criteria reasoning, and prior-authorization optimization. Deployed in VPC, on-premise or air-gapped.

The standard

APS, a proposed provenance standard for AI coverage decisions bound to CMS-0057 and Da Vinci FHIR, is published for comment. Engagement with NAIC, state insurance regulators and HL7 Da Vinci is next.

The independence

Self-attestation is not enough. Governance and trust require an independent evaluation layer: sealed tests, immutable evidence, and standards applied equally to first-party and third-party systems.

The landscape

Governance stops at the model registry.
The determination is where it has to reach.

Every regulated industry that automated consequential decisions arrived at the same requirement: the builder cannot verify its own system. Healthcare is building that layer now around inventory, documentation and drift, one layer above the decision itself. None of the categories below has the behavioral science, the research corpus, or the specialist clinical models to catch or prevent a wrongful determination in real time, across every workflow and every model a plan runs.

The EHR platformsEpic, Oracle Health
Embedding AI natively at 40%+ hospital market share, and attesting to their own. Not governance specialists: no behavioral-science research program, no evaluation corpus, no reach beyond their own workflow and models. A platform cannot independently verify its own AI, and governance must span every vendor a plan encounters.
AI governance platformsCensinet, Solytics, Arthur, Arize, Ferrum, Skan
A real and fast-growing category: model inventory, vendor risk, documentation, bias monitoring, drift detection and audit logging. At most, these systems tell you something changed. They do not identify which determination failed, where it went wrong, what caused it or how to resolve it.
The deciderseviCore, Cohere, Availity, Myndshft, XSOLIS
Authorization and UM vendors that make the determination, then attest to their own work. A builder grading itself is not governance.
Financial auditorsCotiviti, Optum, Zelis, Machinify
Payment integrity: retrospective and claims-focused. Built to recover dollars after the fact, not to prove a decision was made correctly.
The compliance recordSkan, Inovaare
Process logging, access control and policy attestation. Silent on the behavior of the model itself, which is where the failure lives.
The lab benchmarksGeneral-purpose AI evaluation
Capability on static tests. Not behavior inside a live clinical pipeline under real policy, pressure and adversarial input.
Model inventory, drift detection and bias monitoring are table stakes, and we do them. But they measure the model from the outside, and this failure produces no drift signal — a dropped fact and a gamed scoring rule both look like a healthy, confident output. Ascerta evaluates the determination itself.
What exists today

Two working products in MVP, and the assets beneath them.

Live in MVP

Payer MVP

  • Model-agnostic evaluation, real-time monitoring and analysis of every determination, and re-evaluation on every change.
  • The full compliance stack: CMS-0057 API, record capture, denial reasons and metrics, and an audit and appeal portal.
Live in MVP

Provider MVP

  • A full intelligence platform, from EHR and record intake through clinical extraction and submission optimization to a ready-to-submit case.
  • First-pass approval and gold-card optimization, with an automatic appeal packet when a determination is flagged.
Internal assets

Dataset and benchmark

3M+ evaluated AI clinical decisions, 300K+ released publicly, and a public clinical-AI reliability benchmark.

Trained clinical models

Tuned on the corpus for extraction, codification and criteria reasoning, and deployable in VPC, on-premise or air-gapped.

APS

A proposed provenance standard for AI coverage decisions, published for comment.

Ongoing research

Behavioral evaluation and mechanistic interpretability of clinical AI, continuing.

Founder

Built the research, the standard,the products, and the models.

Anthony Cruz, founder of Ascerta

Anthony Cruz

Founder

A decade building software, scaling startups, and several years of independent AI research. In 3 months, alone: designed and ran a three-million-decision evaluation program on clinical AI behavior, authored the APS provenance standard, trained the clinical models, and built both products to working MVP. Along the way, he spoke with dozens of people across healthcare to ground the work in firsthand operating insight, some of whom will be the first hires.

The researchDesigned and ran the research program on clinical AI behavior: millions of evaluated decisions, novel findings, and original mechanistic interpretability research.
The standardAuthored APS, a proposed provenance standard for AI coverage decisions, bound to CMS-0057 and Da Vinci FHIR.
The models and the productsTrained and tuned the clinical models on the corpus, and built the governance infrastructure and provider platform end to end.
ASCERTA

Reasoning is no longer scarce. Trust is.

Ascerta is the trust infrastructure for clinical AI: the governance that proves when a system can be trusted, and the clinical models that power it. It is what enables intelligence to reach the decisions that matter most.

Clinical reasoning, at infinite scale.
ascertahealth.comContact@AscertaHealth.comAnthony@AscertaResearch.com
Try the live demo at ascertahealth.com/demoAccess code 0313
ASCERTA  /  SEPTEMBER 2026 01 / 13