Small wins, in order: each one proves the idea and pays for the next. Two tracks on one clock. The regulated track is the moat and carries a hard decision point; the service track runs on its own economics and never waits for it.
We agree with FDA on how the agent will be judged, and how much evidence that takes, before collecting any of it. Every booked patient can join the study from the first call.
Rules lock · versioned with the protocolEach patient's four-week recovery becomes evidence. Doctors double-check the agent's work, especially the calls where it said all clear.
The target comes from FDA's bar, roughly 1,000 to 4,000 patients. Fix it, slow it, or stop it; answered below.
How often the agent caught real problems, misses included. Published, not marketed.
Frozen rules · the eval measures this versionB+L's first software submission, with a pre-agreed plan for future updates built in, so later improvements don't restart the process. The filing, not the clearance, is the deadline.
Buys the chain after it: each next clearance gets cheaper.
Rule changes · ride the PCCP from hereThe first cheaper clearance the update plan was built for.
Cheaper again, on the same precedent.
After the clearance, on its own evidence.
The agent calls patients during recovery. About 1 in 6 answers raises a flag. A nurse reads every one, and anything urgent reaches a person in under a minute. This is the pressure test: we learn where the agent breaks in a setting that needs no FDA approval, before anything is filed.
Rules v1 · clinician sign-off before the agent dialsThe bookings engine. Non-clinical, no clearance needed. Revenue from day one, and the patient's record opens at the first call.
Most patients come back for the second eye, and the agent remembers everything from the first. The record compounds.
the calls keep running →
Cataract is one surgery per eye; dry-eye therapy is every month, for years. These calls help patients stay on their drops, launched on purpose the same quarter cheap generics arrive. No clearance needed.
Today the agent only spots trouble and hands it to a person. With clearance, it may assess symptoms itself. Same calls, bigger job. The service earned the evidence; the approval hands it back.
Data from the surgical instruments themselves: eye measurements, the lens implanted, what happened in the OR. The camera stays undated until a date would be honest.
The real target is events, not enrollment. The score rests on urgent cases caught, and those are rare, about 4 in 100 patients. FDA's agreed bar sets how many the study needs: a modest bar might mean 40 cases and 1,000 patients; a strict one, 150 cases and 4,000. The number is fixed when FDA agrees to the scoring rules in Q4 2026, and the gate checks pace against it. The date itself sits near LUMIFY's expected loss of exclusivity: a reference, not a certainty, but the program should have its answer before budgets tighten.
A miss forces a diagnosis. Consent and completion high but volume short: the engine works, sign more practices, slip a quarter and pay the price below. Patients or practices rejecting the calls: the thesis is wrong, stop the track. But stopping means keeping a phone service anyone can copy while abandoning the moat, so it requires a named alternative with a better revenue case. Without one, a slower moat beats none.
One caveat: only finished, clinician-reviewed cases count toward the target, and they trail enrollment by months. Pace has to clear with room.
Data collection takes two quarters. The rest goes to agreeing the scoring rules with FDA first, so the evidence counts, and to clinicians rereading the calls where the agent said everything was fine, the slow part by design. Add the paperwork of B+L's first software filing, and nine months is the aggressive schedule, not the padded one.
Two quarters. A one-quarter enrollment slip drops a quarter of finished patients from the score, moves the filing back a quarter, and moves FDA's answer a full review cycle behind it.
A planner, and it is not the AI. It reads her record and the clock and decides the next call, or no call: never before 8am or after 8pm, never more than two a day, and it steps back once her drops have gone well for three days. After any urgent flag, the planner makes no calls for that eye until a nurse has closed it. A worry that isn't urgent means a nurse rings her back, in place of that day's check-in call. Two days of no answer go to a coordinator, not another call. Every call it holds back is logged with the reason, and the practice owns the numbers.
In a simulated patient's full journey, both eyes, it placed 31 calls and held back 50, with no breach of its own rules. The numbers are placeholders until a practice signs them off. This is the layer vendors now sell as an "orchestrator"; B+L writing its rules is what makes B+L the demanding customer rather than the buyer.
Each call has an empathy map: who answers, what the weeks before did to her, and the four to six people who pull away from the typical caller: the proxy daughter, the hard of hearing, the one let down before. Every expectation on the map is scored against the call as built. 40 are met, 23 are partly met, 8 are gaps. The gaps are the backlog. The feelings are inferred, not from interviews; five real calls at a practice would move rows.
The clinician-written rules decide when the agent flags and when it hands the call to a person. The vendor sells pre-surgery education, not triage, so the rules are built here and signed off before the agent dials. They lock when FDA agrees to the study; changing them midway restarts the evidence. The score measures the locked version, and after approval, changes follow the update plan inside the filing, which is why that plan is there.