Leyoxa Care Ledger is a product direction pending a bounded acquisition-readiness sample. It is not represented here as a deployed diagnostic system.
The useful job is classification with provenance
Free-text notes contain declines, watches, deferrals, and referrals that structured reports cannot reliably return. AI can help map that language into a constrained schema when each field remains tied to the exact source text.
Reconciliation requires more than a note summary
The narrative must be compared with recorded evidence, treatment-plan history, procedure completion, time, and benefit context. The output is not “this patient has a condition.” It is “evidence documented; diagnosis not recorded; dentist should re-evaluate.”
The controls belong in the product
- read-only PMS access;
- no generated diagnosis or treatment recommendation;
- source link on every returned field;
- dentist review before any clinical interpretation;
- no PMS write in the current product;
- elective outreach always requires explicit human approval.
Evaluate disagreement, not fluency
A polished summary is not evidence of accuracy. A useful proof set should measure whether reviewers can trace each item to its source, how often the extraction is wrong, which note styles fail, and whether the reconciliation survives dentist review.
AI is valuable here because it can reduce the cost of reading years of narrative. The practice still owns meaning, judgment, and action.