Software generation is collapsing toward zero. That makes the front end disposable — and makes everything it used to hide suddenly load-bearing.
I've spent the last two years building on that assumption: one engine, many doors. A durable core, and interfaces generated per audience, thrown away when the audience changes. This page is the argument, the evidence, and the honest limit.
A junior building beside me has the same models I do. The edge that survives is narrower and harder to copy: twenty years inside clinical and regulatory reality is what makes an AI's output correct enough to ship — the thing neither an AI-native junior with the same tools nor a senior strategist without them produces. In one working session this month I turned a pharmaceutical launch's consumer platform and its full go-to-market into a single sourced page. Speed is the AI. Correctness is the judgment. It has to be both, or it isn't deployable. Recent, live, each read from the primary source:
I read the CY2027 payment rule from the primary text and shipped the surgeon-facing tool the same day — including the correction that the headline everyone was repeating does not actually apply to orthopedics.
surgeonvalue.com/cy2027 →A model writes the persuasive version happily — the claim that reads well and fails the rule. Judgment is what flags it before it ships: the efficacy language 21 CFR 312.7 forbids, the "makes it eligible" the IRS named in its 2024 alert. That catch is the twenty years, not the tool.
Where accountability lives →Not forty projects — one engine wearing forty doors, each generated for its audience and discarded when the audience changes. The durable company is the core underneath; the interfaces are the cheap part.
The systems underneath →Not "AI is powerful." Something narrower and more useful: the specific cost of producing a working interface fell off a cliff, while the cost of producing trust did not move at all.
Independent benchmarking now puts several labs within a few points of each other, including open-weight challengers arriving within days of the leaders. Capability is spreading sideways.
Independent model benchmarks →When any competent model can produce the plausible first draft — of code, of a summary, of a clinical note — the draft stops being the product. I've written this argument out in full elsewhere.
The commoditization argument →If a surface can be regenerated in an afternoon for whoever is standing in front of you, owning the surface is not a moat. Owning what it reads from is.
Three systems I've architected →The pattern I build on. A single core that holds the state and the rules, and a set of thin, disposable surfaces — each one shaped for exactly one audience, none of them precious.
The doors are generated. The engine is not. When an audience changes, I rebuild the door, not the company.
The surgeon door. Revenue integrity and coding intelligence for orthopedic practices — the engine reads the same clinical state, presented as the economics a surgeon actually decides on.
surgeonvalue.com →The family door. Home-based care where the same underlying record becomes a care plan a family can read, and structured clinical data a health system can accept.
co-op.care →The accountability layer. Where the question of who signs — and what that signature is worth once the draft is free — gets answered explicitly.
harnesshealth.ai →The reviewer door. Physician review as a first-class step, built so the human judgment in the loop is visible, logged, and defensible rather than assumed.
clinicalswipe.com →The memory layer. Context that persists across sessions and tools, so the engine accumulates rather than restarting — the substrate the other doors read from.
chanio.com →When CMS proposed the CY2027 payment rules, I read the primary text and shipped the honest read — including the part that says the headline does not apply to orthopedics.
See the surface →This is the whole point, and it's where most of the current advice stops short. If everyone can generate any interface, then interfaces are worth nothing — and the moat moves to whatever generation cannot produce. In clinical work there are exactly four of those.
Not a licensed snapshot. Longitudinal state you captured and nobody can re-derive — how function changes, whether someone adhered, how recovery bent before the clinical metric moved. Models are trained on what is abundant. Faithful state over time is the scarce thing.
How I architect the chain →A licensed clinician who reviews the output and carries the consequence. A model cannot hold a license, cannot be sued, and cannot be disciplined. That asymmetry is not a temporary gap in the technology — it is the structure of the profession.
Review as a first-class step →Regulatory standing, which is now being written down explicitly. CMS has proposed that remote monitoring be payable only when the clinical staff are direct employees of the billing practice — accountability turned into a condition of payment.
Read from the primary rule →When any model can generate any draft, the scarce skill becomes knowing which model to use, which data path to trust, and when a $0.87 specialist beats a $15 frontier. Generation is cheap; selection is expensive.
The memory layer that learns →The draft is becoming free. The signature is not. Every dollar in clinical care still flows through someone willing to put their name on a decision — and that is the one part of the stack that gets more valuable as generation gets cheaper, not less.
I'd rather tell you where this argument stops than oversell it, because the oversell is everywhere right now and it is not useful to anyone hiring.
Acemoglu's estimate puts total factor productivity gains from AI at no more than roughly 0.66% over ten years, with a more conservative case under 0.53%. Individual leverage is real and large. The aggregate effect is modest — because diffusion, not capability, is the binding constraint.
NBER Working Paper 32487 →The scarce skill is knowing which five percent must not be generated, and then getting the result adopted inside an organization built to resist it. Clinical adoption, regulatory standing, and reimbursement are where these projects actually die. That is the work I do.
How I engage →Stated plainly: a person who can generate any interface and cannot get a clinician to change one habit has produced nothing. I've spent twenty years on the second problem — in orthopedic technology and health-system partnerships, not in practice — and the first one is now the easy half.
Three ways this is useful to a company, in the order people usually need it.
I grew BrainLAB's orthopedic vertical to $250M across two continents and hold five patents in image-guided navigation. I know how clinical technology gets adopted here, and where non-U.S. companies reliably misjudge the entry.
The career arc →An MD who reads your model output, designs the evaluation, and finds the seam where a human has to be accountable — sitting inside the engineering loop rather than reviewing it from a committee.
Systems I've architected →The attestation, the audit trail, and the reimbursement path — designed together, because in clinical AI they are the same problem. This is the layer that decides whether a product is deployable or a demo.
Start a conversation →