Clinical AI's next job is watching itself

A supervisor for every call

Hippocratic AI released the AI Compliance Supervisor, a tool designed to monitor and evaluate call centre performance across human and AI agents. Unlike traditional quality programs that rely on sampling, the system reviews every call to detect compliance violations, clinical errors and resolution rates in real time. It evaluates calls for compliance, clinical errors, emotional intelligence, intent and resolution rates, and identifies regulatory and script violations in real time as well as medically consequential clinical errors for review. m:…6591

The system functions as a horizontal layer across human agents, Hippocratic agents and third-party AIs, and replaces quarterly audit cycles with same-day problem surfacing. The point is measurement, not just monitoring: a unified standard applied to human and AI interactions alike, so a health system can compare them on the same terms. For a sector where a missed clinical error is a patient safety event, the shift from sampling to full coverage is the difference between finding a problem and finding it in time. m:…6591

The patients the record missed

Viz.ai announced that two research abstracts evaluating its AI-powered Viz HCM algorithm have been accepted for presentation at the 2026 American Heart Association Scientific Sessions. The studies, conducted by Beth Israel Medical Center and the University of Colorado, test the algorithm's ability to identify suspected hypertrophic cardiomyopathy in patients who might otherwise remain undiagnosed. Viz HCM was the first FDA-cleared AI algorithm designed to help clinicians detect signs of HCM using a standard 12-lead ECG. m:…6844

The findings are striking. An estimated 85% of patients with hypertrophic cardiomyopathy may remain undiagnosed, underdiagnosed or misdiagnosed. Nearly 40% of patients flagged by Viz HCM for suspected HCM had no corresponding diagnosis code in their medical records. Patients flagged by the algorithm had a 2.68-fold higher hazard of adverse clinical events. The University of Colorado study analysed ECGs from more than 29,000 patients, and Viz HCM has been evaluated across 21 clinical abstracts and studies to date. Viz.ai's platform is deployed in 2,000 hospitals across the United States, and the company was the first awarded CMS reimbursement for AI. m:…6844

The cost of a bad data point

Saama's post prices a different kind of clinical failure: the data anomaly. Data anomalies are unusual or unexpected values, relationships or patterns — a Grade 4 lab abnormality with no corresponding adverse event, for example — and standard edit checks are not built to catch patterns that raise clinical questions. Phase III trials collect an average of 5.9 million data points per protocol, and data volume in clinical trials has been growing at 11% annually since 2020. Only 1.7% of data points were changed as a result of queries in a meta-analysis of 20 Phase III trials, which suggests most of the data was fine and the anomalies are the expensive part. m:…9307

The price tag is concrete. Standard edit checks cost $28 to $71 per query, systematic patterns can delay database lock by two to six weeks, and database lock delays cost $780,024 to $2.34 million. Each day of postponed approval costs an innovator roughly $500,000 in forgone prescription sales. Adequacy of records accounted for 23.8% to 31.9% of cited deficiencies in FDA inspections. The FDA adopted ICH E6(R3) in September 2025, shifting clinical practice from prescriptive monitoring to proportionate, risk-based quality management, and proactive detection catches anomalies at data entry rather than at database lock. m:…9307

Sovereign agents in the clinic

Corti's October wrapped adds the infrastructure angle. The company visited eight European cities in October 2026 to discuss sovereign AI, agentic systems and healthcare integration, and its CEO Lars Maaløe stated that sovereign AI is about independence and controlling technology, not just data hosting location. Corti Bridge allows developers to use open-weight Corti Models within coding environments like Claude Code, and the company plans to train and post-train trillion-parameter models in Copenhagen. m:…0636

The deployment detail matters most. Dedalus is building approximately 30 specialised AI agents on Corti's Agentic Framework for clinical EHR workflows, and Corti is integrating its medical AI models into Voicepoint's Xenon platform for the Swiss market, with enhanced support for Swiss German and Swiss French. Andreas Cleve, at Corti, warned that automation without control is a liability in healthcare. Thirty agents inside an electronic health record is exactly the kind of deployment where Hippocratic's supervisor and Saama's anomaly detection stop being optional. m:…0636

The pattern

The four stories share a premise: clinical AI is no longer being judged on whether it can do the task, but on whether it can be watched doing it. Hippocratic watches every call. Viz.ai's algorithm is watched by 21 studies and a reimbursement decision. Saama prices what happens when nobody watches the data. Corti builds agents for a setting where control is a liability issue. The next phase of health AI is not more capability. It is more accountability, and the companies that build the watching layer are the ones health systems will trust with the doing layer. m:…6591 m:…6844 m:…9307 m:…0636

For health system leaders, the question to ask of any clinical AI vendor is now the same in all four stories: show me the supervision. Who reviews every call, who catches the anomaly, who finds the undiagnosed patient, and who is accountable when the agent is wrong? The vendors with a good answer are the ones publishing it this week. m:…6591 m:…9307 m:…6844