Trust AI with more.

More decisions. More volume. More autonomy.

AI can dramatically increase the speed and capacity of prior authorization. But deploying it at scale creates a new problem: proving that the system is reliable, compliant, and behaving as intended.

Ascerta evaluates the entire deployed pipeline to establish where it is reliable, where it fails, and the conditions under which it can safely operate. Continuous monitoring keeps the system within that operating envelope, while every decision is captured in a verifiable record built for ongoing compliance, audits, appeals, and regulatory review.

More of utilization management can be automated with the assurance, evidence, and regulatory readiness required at scale.

Product one

Ascerta Trace

Audit-ready at all times. Every determination is observable, reconstructable, and already assembled — so nothing has to be scattered together after the request arrives.

Plans are accountable for the decisions their models shape. CMS-0057-F, the Interoperability and Prior Authorization rule; CMS-4201-F, the Medicare Advantage utilization-management rule; state AI-in-UM statutes; and NAIC model guidance all point the same direction: a plan must be able to show, after the fact, how a determination was made, who was accountable for it, and that the system behind it behaves the way the plan says it does. Screenshots and spreadsheets do not survive that question. Trace produces the evidence as the work happens, so the complete record for any determination is available the moment it is asked for — not assembled weeks later out of whatever survived.

What it captures

Enough to rebuild the determination from nothing, years later, without calling the person who made it.

  • The input surface: request, exact prompt and instructions, the coverage policy hash-pinned to the version in force, retrieval and tool results.
  • The pipeline, recorded stage by stage: extraction, summarization, codification, and the determination, each with its own captured inputs and outputs.
  • Human oversight, measured on real reviews: on every non-approval, the licensed reviewer's identity, their credential and specialty against the clinical issue, the routing path, their reasoning, time with the record, and whether the determination changed.
  • Turnaround clocks, denial reason codes, and appeal and overturn outcomes.

Integrity is critical

Every audit-trail vendor will tell you their ledger is tamper-evident. Almost all of them verify the chain against itself, which proves only that the chain was not touched. The readable record lives in a different store, and nothing checks it.

Ascerta verifies both. Each stored artifact is re-fingerprinted and compared against the fingerprint sealed at determination time. A record edited afterward fails that comparison even though the chain still reports clean. The difference shows up on the day it matters:

Compliance Automation

  • Ascerta collects evidence automatically and delivers full prebuilt packs on demand: appeal packs, CMS program audit packs, DOI and DMHC exam packs, ODAG-shaped universes, conformance reports and more.

Meet Theo, the built-in agent

Ask Theo in plain language. It keeps threads and institutional memory and works from a registered tool set: summarize a case, verify the ledger, assemble a regulator pack, generate a defense pack for a named matter, respond to a records request, draft a threshold-change attestation, add a probe, recall and record institutional memory, and list the rules in force.

Your public trust profiles

Built to the Ascerta Provenance Standard v1.0, so the record can be shown to outsiders without exposing PHI.

  • An optional public trust profile: a live page showing system health, certification state, and compliance posture.
  • Scoped portals for regulators, oversight bodies, and outside counsel, with expiring access, and each disclosure of the record written as its own ledger event (APS §9.5).
  • An evidence manifest on the decision path: every document, its digest at decision time, when it arrived, and whether the model and the reviewer actually saw it (APS §3.1).
  • Enforced pipeline continuity — each stage's input digest matched against the prior stage's output, so a chain with a hole cannot pass (APS §4.1).
Product two

Ascerta Assure

Evaluate

An independent third-party evaluation and certification of the deployed system, built on Ascerta's peer-reviewed research into how these systems fail in deployment, the published literature on model behavior, and established evaluation science.

The biggest exposure isn't a single critical failure or attack, it's an extractor that misses the one line in the chart that changes the answer. A summary that drops a disqualifying fact before the reviewer sees it. A codifier that maps to the wrong criteria. A model optimizing for something other than the clinical question, triggered by a variable nobody would flag as a risk.

Nine capability lines across the pipeline the plan actually runs: decision integrity, medical correctness, extraction fidelity, summarization fidelity, codification integrity, evidence resolution, robustness under adversarial input, proxy inducibility, and confidence integrity.

What comes back is a complete behavioral map of the deployment. Where it is strong, the plan can widen automation and run it unattended with a documented basis for doing so. Where it is weak, the specific failure conditions are named, remediations mapped, and monitoring watches for them in production.

Monitor

A certified system is a living system. Certification describes a deployment as it behaved on the day it was examined. But the model is updated behind an endpoint you do not control. Coverage policy is revised. The retrieval corpus grows. Case mix shifts with a new line of business or a new state. Any one of these changes behavior without anyone having changed anything on purpose.

These failures are also silent. There is no exception, no error rate, no alert. The system returns a confident, well-formed determination that is simply not the one it would have returned last quarter. Without monitoring, the first signal is an appeal, an overturn, or a pattern somebody notices months later.

Monitoring is what keeps a certificate a statement about the present. Certification decays for three reasons: the model is swapped, the policy is revised, or the input distribution shifts. Monitoring catches all three — and it catches the individual case before it becomes an appeal.

  • Flag → confirm → escalate. A suspect determination is flagged in either direction — a denial that should not have happened, or an approval that should not have been granted. Agents re-derive the case against the policy in force. Escalation happens only when the re-derivation actually flipped the determination, and arrives with that evidence attached.
  • A model's opinion can flag a case. It can never certify one. Escalation requires a matched-pair re-run that actually flipped the determination — an AI finding, however confident, cannot by itself mark a decision as wrong. The rule is enforced in code and pinned by a test: a maximally-confident judge on an unproven case still yields no escalation.
  • The same mechanism finds wrongful approvals, which is recovered revenue rather than avoided cost. It is also the under-watched half: 80 to 90 percent of approvals never see a human at all.
  • The mechanisms are cohort drift, live envelope comparison against the certified intervals, matched-pair re-derivation, the integrity judge, and calibration tracking.
  • Drift and calibration are trended as their own metrics rather than buried in an overall score.

How the pieces fit

Trace answers what happened on this decision. Assure answers can this system be trusted to make decisions like it at all. Each decision record in Trace carries the assurance state that was in force when it was made, so an auditor reading a single case sees the certificate lines, the conditions attached to them, and whether any monitor had flagged the behavior that day.

Step 1
Connect

Ascerta sits alongside the UM stack and captures the decision record in flight. No change to how clinicians work.

Step 2
Certify

The examination runs against pre-registered thresholds and issues a per-line certificate for the deployed configuration.

Step 3
Stay current

Monitors run continuously, on event triggers, and on a periodic cadence, with recertification triggers defined per line.