Moving the collection point: at-home diagnostics and what a result is supposed to trigger

The bet

Healthcare has moved almost everything out of the building. The consult happens on video. The follow-up happens by message. The prescription arrives by mail.

The blood draw never left.

That is the gap, and it is a large one. Roughly 70% of healthcare decisions originate in a diagnostic result, and the diagnostic is now the step most likely to force a person into a car. When people don’t go, the decision doesn’t get made. A third of patients delay or avoid a blood draw outright.

Our bet — and it is a bet, stated plainly — is that reimbursement parity for remote care is inevitable rather than speculative. Payors have already conceded the consult. Once they concede the collection, the economics invert, and the organisations that win are the ones whose infrastructure was already built for it.

If healthcare were invented today, it would be organised around asynchronous communication.

Why this is harder than shipping a kit

The consumer at-home testing market is crowded with companies that mail a box. The box is the easy part. Three things are hard, and they are the product.

The collection has to be right

An invalid specimen closes nothing — it costs the kit, the shipping, the lab slot and the patient’s willingness to try again. Roughly 80% of patients find self-collection acceptable in principle; far fewer are confident they did it correctly.

So we treat the collection itself as the product. Video-recorded self-collection with asynchronous expert review. Live remote proctoring where the session is a state machine with typed interruptions that genuinely block advancement rather than warn. Automatic replacement-kit reshipment when an observer marks a failure. And a post-capture grader that segments the recorded video into procedure phases and returns per-phase pass/fail with machine-readable failure reasons, so the human review is reserved for the cases that need it.

Training that grader required labelled video of procedures performed correctly and incorrectly in every plausible way, which does not exist as a dataset. So we built a generator for it — a permutation engine across procedure phase, correctness category and error type, with Latin-hypercube sampling of the nuisance variables that wreck real-world models: camera placement, lighting, room, operator handedness, demographics. Deterministically seeded, so any bucket regenerates identically.

The result has to go somewhere

A diagnostic that ends in a PDF is a dead end. The care pathway is: qualifying result → telehealth encounter routed by the patient’s geography and the provider’s licensure → prescription → and then the part everyone skips, a mandatory confirmatory retest before the episode can close. Test of cure as a gate, not a suggestion.

That closed loop is the difference between a testing company and a healthcare company.

It has to talk to systems built in 1989

Every serious counterpart in this industry speaks HL7 v2 — ORM in, ORU out — and increasingly FHIR R4, and some of them still want a fax. Interoperability is not a feature you add later; it is the cost of entry, and it is where most consumer-health platforms discover they have built something that cannot be sold to an institution.

On AI, and where we refuse to put it

We use AI heavily. We also draw a hard line, and I want it stated precisely because the industry is about to have a painful few years over this:

No AI diagnosis. No AI clinical recommendation. Clinical routing, ordering rules and result handling are explicit, auditable, if-this-then-that logic that a compliance officer can read line by line and a regulator can audit.

AI does the work where being wrong is recoverable: generating patient education, watching a collection and flagging a probable error, translating a message, drafting a result report from a structured template, choosing which outreach to send next. Every one of those has a human or a deterministic rule between it and a clinical decision.

The engineering organisation runs AI-first — senior humans as subject-matter leads and reviewers, with code, tests, documentation and translation produced by AI systems under direction. The practical consequence is that adding a country or a new rule is a configuration change rather than a multi-month engineering project. The philosophical consequence is that I now have three years of direct evidence about what that operating model does and does not do, which is most of what clients want to talk to me about.

What I take from it into client work

This is the current, live version of all three service lines at once — a real platform under real compliance, sold to real institutional buyers, with a marketing motion and an analytics layer attached. When a healthcare or regulated-industry client asks whether we have done this, the answer is that we are doing it today, not that we did it once.

It is also the reason I am careful about AI promises. It is very easy to demo. It is hard to ship under HIPAA, and harder to defend afterwards.