By Andreas Scherer, PhD, Founder and CEO, Salto Partners
Most debate about AI in clinical genomics asks whether a model can make the call. That framing leads nowhere for a regulated laboratory, because a black box that produces an answer is exactly what it cannot trust. The more useful question is how AI can join the workflow without giving up the grounding and human control that make interpretation defensible.
The core argument
- Start where the value is clear. Large language models add something new with unstructured evidence: summarizing the case for a variant, extracting findings from the literature with links back to the source, and translating reports into local languages.
- Grounding beats model size. AI should reason over curated, versioned evidence, not the open-ended contents of its training data. Every statement must trace to its source so a result can be reproduced and reviewed.
- The non-negotiables. Validation under the quality system, auditability, change control for every model version, continuous performance monitoring, and a human who owns every decision.
- The economics shift. Sequencing scaled but interpretation did not, because it scales with headcount. Grounded AI loosens that link and makes continuous reanalysis practical as evidence changes.
What it means for leaders
Regulators are converging on the same question. FDA thinking on AI-enabled devices, CAP and CLIA traceability, IVDR and the EU AI Act all ask: can you show what the system did, why, on what evidence and under whose authority? The governance a good platform imposes and the governance regulation requires turn out to be the same engineering.
For executives and investors, that makes AI in diagnostics a strategy, execution and compliance question at once. The winners will not be the companies with the flashiest model. They will be the ones whose platforms can hold AI accountable.
Read the full perspective: AI as Infrastructure: Reshaping Clinical Variant Interpretation, published in GenomeWeb.
Working through one of these questions in your own organization? Let’s talk.