AI Drug Diversion Detection Failure at Erlanger Hospital: What Healthcare Organizations Can Learn

A recent case at Erlanger Medical Center in Chattanooga, Tennessee, has drawn attention to a growing challenge in healthcare technology: what happens when an AI monitoring system fails to detect the problem it was deployed to catch.

According to a Tennessee Board of Nursing consent order and reporting from Local 3 News, a certified registered nurse anesthetist (CRNA) diverted waste fentanyl for personal use over a period of several months in 2025. The hospital's AI-powered drug diversion detection software did not flag the activity. The case offers a well-documented example of the risks that come with AI monitoring deployments and the safeguards that healthcare organizations and other regulated industries can put in place.

What Happened at Erlanger Hospital

The timeline, as documented in the consent order:

The nurse began diverting waste fentanyl (i.e., unused medication remaining after a prescribed dose is administered) around March 2025, initially once or twice a week. By June 2025, the use had become daily.

The diversion was not detected by software. Instead, in late June 2025, coworkers reported visible signs of impairment, including slurred speech and difficulty staying alert during a shift at the surgery center. A urine drug screen the following day returned positive for fentanyl, and the nurse's employer, an anesthesia services contractor, terminated his employment.

A subsequent audit of dispensing records from March to June 2025 identified five instances of missing waste fentanyl that the AI monitoring system had not flagged. A manual review of anesthesia charts by the chief CRNA found additional inconsistencies between drug dispensing and waste documentation that the software also missed.

The consent order notes a key detail: at the time, the AI system was in its initial learning phase.

Why AI Monitoring Systems Miss Events During Learning Phases

Many AI-based anomaly detection tools, including drug diversion monitoring, fraud detection, and clinical surveillance systems, require a calibration period. During this phase, the system builds a statistical baseline of "normal" behavior for a facility, department, or individual clinician before it can reliably identify deviations.

This creates a structural risk that is easy to overlook during procurement and rollout:

• Coverage gap. During calibration, detection capability is reduced or absent, even though the system is technically live.

• Baseline contamination. If anomalous behavior is present during the learning phase, the system may incorporate it into its definition of normal, making future detection harder.

• Misplaced confidence. Staff and leadership may assume monitoring is fully operational from day one, and manual controls may be scaled back prematurely.

None of these risks means AI monitoring is ineffective. Mature, well-calibrated systems can review volumes of transaction data that manual audits cannot match. The risk is specific to the transition period — and to deployments where no one has defined what happens during it.

Compensating Controls During AI Deployment

Industry practice in healthcare compliance and AI governance points to several measures that reduce risk while an AI monitoring system ramps up:

• Maintain parallel manual controls. Manual audits, reconciliation of dispensing and waste records, and supervisor chart reviews should continue at full strength — or increase — until the AI system's detection performance has been validated. In the Erlanger case, it was a manual chart review that ultimately surfaced the discrepancies.

• Define and document the learning phase. Deployment plans should state how long calibration is expected to take, what detection capability exists during that window, and who is accountable for coverage in the interim.

• Validate with seeded tests. Detection systems fail silently: the absence of alerts can mean either "nothing to detect" or "detection isn't working." Periodic seeded anomalies or red-team-style tests distinguish between the two.

• Keep humans in the loop. The software vendor in this case stated its commitment to supporting human-in-the-loop decision making. That framing reflects broader industry consensus: AI monitoring tools are most reliable as one layer in a defense-in-depth model, alongside human observation and procedural controls — not as a replacement for them.

• Monitor the monitor. Ongoing performance metrics — alert rates, false-negative reviews, comparison against audit findings — should be part of the operating model, not just the pilot.

What This Means Beyond Healthcare

While this case involves controlled substances, the underlying pattern applies to any organization deploying AI for oversight tasks: financial anomaly detection, security monitoring, quality assurance, and compliance surveillance all share the same failure mode. An AI system that has been deployed is not the same as an AI system that has been validated, and the gap between the two is where undetected incidents accumulate.

Regulatory scrutiny of AI in high-stakes settings is also increasing. Documentation of deployment decisions, calibration periods, and compensating controls is becoming part of what auditors, boards, and regulators expect to see when an AI-assisted process fails.

Key Takeaways

• An AI drug diversion detection system at Erlanger Hospital did not flag five instances of missing waste fentanyl while in its initial learning phase; the diversion was identified through human observation and manual review.

• AI monitoring systems commonly have reduced detection capability during calibration, creating a coverage gap that requires compensating manual controls.

• Silent failure is the defining risk of detection systems: organizations should validate performance through seeded testing and parallel audits rather than assuming no alerts means no incidents.

• Effective AI oversight combines automated detection, human review, and documented governance — a defense-in-depth approach rather than reliance on any single layer.

  
  

Organizations deploying AI monitoring tools can reduce these risks with a structured approach to validation and governance. Alignmt AI works with teams to assess AI system readiness, define compensating controls during deployment phases, and build ongoing assurance processes.