The fastest way to make a dental AI product unsafe is to remove the clinician from the loop. The second fastest is to put the clinician in the loop only at the very end — after the AI has already produced a confident-looking diagnosis or plan.
We have spent the last year building Faurenis with a different premise: AI is a documentation and continuity tool, not a diagnostic appliance. The clinician remains the decision-maker on every clinical question. The product is designed so that the AI's role ends at "draft" — not "result".
Three concrete design choices follow from that premise.
First, every AI-generated paragraph is visibly badged in the chart. If you cannot tell at a glance which sentence was drafted and which one you wrote, you cannot review carefully. Faurenis never silently merges AI text with clinician text.
Second, claims are grounded. Every clinical statement Faurenis drafts points back to a source — a transcript line, a chart entry, an imaging reading. If a statement does not have a source, Faurenis marks it as low confidence and routes it to the validation queue rather than committing it.
Third, confidence is honest. Modern language models can generate plausible-sounding nonsense with high internal probability. We do not surface internal probability as confidence. We combine it with grounding ("is every claim sourced?") and clinical specificity heuristics ("did the model just hedge in three sentences?") and only then assign the score.
None of this is novel science. What is hard is keeping it consistent as the product grows. The temptation to add a "Faurenis thinks this is most likely X" badge to a radiograph is real. We have not added it, and we will not add it, because the moment a confidence-coloured label sits next to a clinical claim, the clinician's role shifts from "decision-maker" to "label reviewer". That is the wrong gradient.
The frame we hold internally: dental AI should make the clinician faster and more thorough, not more passive. Every product decision we make is tested against that frame.