A Second Layer of Safety: MDCalc and AMBOSS partner on CrossCover and Clinical AI

Clinical AI tools sound sure of themselves, even when they're wrong. In this episode, MDCalc co-founder and CEO Dr. Joe Habboushe talks with Dr. Madjid Salimi, co-founder and co-CEO of AMBOSS, about what it takes to build AI that clinicians can rely on. AMBOSS AI is one of the first models available in CrossCover, MDCalc's new independent brings evidence checks, privacy safeguards and MDCalc clinical tools to the AI physicians choose.

They start with a problem many clinicians don't know about: language models often get medical calculations wrong. Madjid explains why AMBOSS AI answers only from hand-selected, physician-reviewed sources, why it's built to flag conflicts and uncertainty instead of sounding certain, and why AMBOSS took a slower, more careful path to launch. That approach showed up in NOHARM, a benchmark from researchers at Stanford and Harvard, where AMBOSS AI had the lowest rate of potentially severe errors of any system tested, at 2.9%.

Joe then walks through how CrossCover works as an independent layer over that AI, from PHI Scrubbing and TrackerBlocking to CrossCheck and the MDCalc calculators shown when relevant. . The two also discuss why partner models in CrossCover never learn who is asking, the risk of hidden commercial influence in clinical AI, and why Madjid doesn't expect any single company to own clinical AI from end to end.

What You'll Learn

• Why language models struggle with clinical calculators, and why that matters for patient care

• How AMBOSS AI grounds its answers in physician-reviewed content and flags uncertainty instead of hiding it

• What the Stanford and Harvard NOHARM benchmark measured, and why even the best tools still made severe errors

• How CrossCover removes patient and clinician identifiers and checks answers against guidelines and the literature

• Why Madjid compares today's clinical AI to the early days of the internet, and what a more open system could look like

Links & Resources

• CrossCover: crosscover.ai

• AMBOSS: amboss.com

• NOHARM benchmark (arXiv preprint): https://arxiv.org/pdf/2512.01241

• MDCalc: mdcalc.com

• Follow MDCalc on LinkedIn: linkedin.com/company/mdcalc

• Follow MDCalc on X/Twitter: @MDCalc