The end of sample-based audits: what reviewing every claim means for your notes

The traditional audit is a sampling exercise. Auditors pull a statistically valid sample of claims, request the records, find the error rate and extrapolate it to the whole population. That approach still exists. Since late 2025, though, payers have added a second layer: software screens every claim before payment, and people review the claims it flags. The two layers expose different weaknesses, and agencies need to prepare for both.

How sampling works, and what it finds#

Recent HHS Office of Inspector General audits of Medicaid autism services show the sampling model clearly. In a report issued in February 2026 (A-09-24-02004), OIG reviewed Colorado's fee-for-service payments for applied behavior analysis in 2022 and 2023. All 100 sampled enrollee-months included at least one improper or potentially improper claim line. From that sample, OIG estimated at least $77.8 million in improper payments. A January 2026 audit of Maine estimated $45.6 million in improper payments for autism-related services. The leading error there was a missing comprehensive assessment, or one without a required staff or parent signature.

Most of these findings are documentation failures, not evidence that care was never delivered. Under sampling, they become an extrapolated repayment demand, usually years after the services were provided.

Screening every claim: Minnesota#

Minnesota made the clearest move to prepayment review. Governor Walz announced the change on October 29, 2025, and it began in late December 2025. According to the Minnesota Department of Human Services FAQ:

• Every fee-for-service claim for 14 high-risk services is held and run through analytics by Optum, a third-party vendor, which shares findings with DHS. The behavioral health services on the list include adult rehabilitative mental health services (ARMHS), assertive community treatment, intensive residential treatment services, recovery peer support, EIDBI and housing stabilization services.
• Unflagged claims are released "in the following warrant cycle (two weeks later)." Flagged claims receive a documentation request in the provider's MN–ITS mailbox.
• DHS says it is "unable to expedite the pre-payment review process," and that enrolled providers may not stop serving members while claims are under review.
• Managed care organizations are excluded because they "conduct their own pre-payment review."

In February 2026, DHS said it would spend the next nine months building a more automated prepayment system that uses advanced data analytics and possibly AI to check billing against policy. A provider enrollment freeze on 13 high-risk service categories took effect on January 27, 2026, for at least six months.

Medicare and federal Medicaid oversight#

Federal programs are moving in the same direction on several fronts:

Fraud Defense Operations Center (Medicare)

CMS launched it in March 2025. AI builds a heat map of frequently billed items and services, and machine learning flags documentation patterns within those areas for human review. CMS said in September 2026 that it had stopped nearly $2.5 billion in potentially fraudulent Medicare payments and taken action against just under 800 providers. CMS's chief operating officer named incorrect or missing documentation as one of the biggest drivers of improper payments.

Medicaid Fraud War Room

CMS launched it in April 2026 with HHS-OIG. It uses analytics to find high-risk providers and helps states put payment suspensions in place faster (CMS fact sheet, June 2026).

WISeR (Medicare)

This model brings prior authorization and prepayment review, using AI plus human clinical review, to selected Part B services in six states from January 1, 2026 through 2031. A clinician must review every denial. The initial service list does not include behavioral health. It is still significant because it shows CMS is willing to put technology-assisted review in front of payment in traditional Medicare.

HHS AERO

Launched May 21, 2026, according to trade and legal press. It reportedly uses AI to re-analyze at least five years of Single Audit data for recipients of federal funds, including Medicaid. We have not found a primary HHS document describing it, so treat the details as provisional.

What changes for documentation#

The third row matters most. Under sampling, a cloned note is a problem only if it lands in the sample. When every claim is screened, cloning is easy to detect: identical text across clients, the same start time and duration every visit, or a plan of care that never changes. A note that is good on its own can still look bad next to the clinician's other 400 notes.

A readiness checklist#

• Every note ties to a current plan goal, and plans are reviewed on schedule.
• Start and stop times are recorded, not just durations, and no clinician's sessions overlap.
• Templates prompt for what is specific to the session, not boilerplate that fills itself in.
• Assessments and signatures that a service depends on, such as the comprehensive assessment in the Maine audit, are checked before the claim goes out.
• Someone monitors the payer portal or mailbox for records requests daily, and you can produce a complete record within days.
• You run the same pattern checks on your own claims every month before a payer does.

How Margent helps#

When payers check every claim, providers need to check every note before it becomes a claim. Note checks cover 100% of notes at signing for plan linkage, required elements, time and overlap conflicts, and similarity to the clinician's recent notes, and offers one-click fixes. Letters & appeals assembles records-request packets, with each claim linked to its source document, so answering a request doesn't mean a week of pulling charts.