Patient
1 / 10
Plays on its own · about 40 seconds
From the moment a visit is booked to a clean insurance claim. Tavik checks in with the patient, listens in the room, writes the note, checks the claim against the CMS policy that governs each code, and drafts the appeal if the insurer still says no.
Built for orthopaedic and musculoskeletal practices billing Medicare — the coverage policy, denial causes and code rules shipped here are theirs.
A simulated visit with a made-up patient and insurer. The film has no sound; what Tavik is doing is written along the bottom.
Not fraud. Not bad medicine. A missing sentence in a note that was signed weeks ago.
The sentence that would have prevented it takes the doctor about eight words to say. The only problem is that the moment to say it has already passed.
Denial rate, top-cause and provider figures: Experian Health, State of Claims 2025 (survey of 250 healthcare professionals, June–July 2025; denial rate is 2024 data). Rework cost: MGMA, varying with claim complexity.
One knee, from the patient's phone to the appeal — and to the check that stops it happening twice. Run it yourself →
Ambient speech with no speaker labels and no codes spoken. A model reads the encounter into facts — which joint, which side, whose treatment, done today or booked — and every fact is dropped unless it can be quoted from the transcript.
No model names a code. Deterministic rules map the facts to CPT and ICD, then pull that code's coverage criteria from the CMS article or NCD that governs it — quoted, cited and checked against the source.
A gap is a question, not a verdict. It appears when the visit reaches the thing that risks the denial and disappears the moment the doctor says the missing sentence — mid-dictation, while it is still cheap to answer.
Only deterministic data — NCCI, MUE, the ICD tabular, the LCD — can block a claim. A model's opinion is labelled as one and never blocks. Asked to appeal on a thin record, it withholds the letter rather than inventing support for it.
These are inputs — what the engine reads. What it scores on them is below.
Most billing AI is sold on a first-pass rate nobody can reproduce. Here is ours, and what it does and does not establish.
That the deterministic layers behave as specified: NCCI bundling both ways (catches what it should, stays quiet on what it shouldn't), MUE caps, ICD tabular legality and specificity, modifier selection, ordered-versus-performed billing, prior-auth negation, and whether the appeal generator is honest about a weak record. Every category exists because something once got it wrong — the suite is a scar record, not a showcase.
That it will lower your denial rate. These are our cases, and a benchmark you write yourself can only prove the system does what you already thought to check. Establishing the rest takes your claims — which is what a pilot is for.
Nothing here is a mock-up. Every panel on this site is output from the engine the product runs, and every page will run it again on a visit you dictate yourself. Start wherever you like — none of it needs an account.
Every case shown on this site is synthetic. No patient data appears anywhere on these pages — please do not paste real patient information into them.
We are taking a small number of orthopaedic practices as pilots. You send us denials you have already received; we show you which ones the engine would have flagged before submission, and which it would have missed. No integration to start, and nothing to install.
One reply from a person, not a sequence. Don't send us any patient data — we'll tell you how to de-identify what you send.