Why we chose audit-traceable over foundation model

Aidoc closed a $150M Goldman Sachs-led Series E earlier this year to build a foundation model for radiology reporting. The press cycle was respectful, the round was oversubscribed, and the strategic logic is clean: a single, large foundation model that handles every modality, every anatomy, every report type — sold to every hospital system in the country.

We considered building the same thing. We did not. This is the calculus for why we chose a different path — and why we believe the durable value in radiology AI for the next decade will be created at a layer most foundation-model companies are not building for.

The honest premise of the foundation-model bet

The bet is this: if one model can produce high-quality draft reports across all of radiology, the marginal cost of the next report is essentially zero. The economic flywheel is "scale the model, win the market, charge per report." This is the same playbook OpenAI used to win consumer AI and Microsoft used to win productivity software. It works when the underlying task is general-purpose and the training data is abundant.

Radiology reporting is not general-purpose. It is one of the most institutionally-specific, liability-sensitive, regulator-touched, contractually-bounded domains in all of clinical medicine. The reporting workflow at Hospital A is different from Hospital B in template structure, macro library, peer-review cadence, critical-results escalation, sign-off chain, and downstream integration with the EHR. Foundation models handle the variability of language well; they handle the variability of workflow poorly.

Three reasons we did not build a foundation model

1. The procurement audit trail is the product, not the draft

A foundation model that drafts a chest CT report with 95% accuracy is impressive. A foundation model that drafts a chest CT report with 95% accuracy and produces a per-field audit trail showing exactly which prior studies, which measurements, and which guideline references informed each sentence — that is what a hospital AI committee is willing to defend in front of a CMS audit, a malpractice deposition, and a value-based-care contract renegotiation.

The draft is the visible surface. The audit trail is what survives the next seven years of medical-legal review. We bet that buyers will pay more, and for longer, for the layer that survives that review than for the layer that produces the draft.

2. Liability flows downstream to whoever signs the report

In every clinical workflow we have studied, the radiologist is the author of record for the final report. The AI is an assistant — it produces suggestions, the radiologist edits and signs. When that workflow works, liability sits with the radiologist and the institution. When that workflow breaks — when the AI generates a confident but incorrect suggestion that the radiologist approves — liability flows to the institution, to the AI vendor's product liability insurance, and to the standard-of-care question of "what did the radiologist actually know and when."

A foundation model that generates a wrong report with high confidence and a generic provenance string is a liability multiplier. A system that produces a draft with a per-field audit trail — "this sentence is grounded in study #123 from 2024-03-12, with these specific measurements" — is a liability reducer. We chose to build the layer that reduces liability because we believe that is what will be procurement-grade in 2027 and beyond.

3. Regulatory pathways reward provenance, not capability

The FDA's Breakthrough Device Designation pathway, the CMS MCIT pathway, and the CPT Category III code submission process all reward evidence of traceability, not evidence of capability. A foundation model that produces 95% accuracy on a benchmark is not, on its own, sufficient for any of these pathways. A system that can demonstrate per-field provenance, per-report reproducibility, and per-decision auditability is sufficient for all three.

This is why we believe the regulatory ceiling will be set by provenance, not by capability. The next round of CMS reimbursement decisions will favor systems that can prove what they did, not just systems that did something well.

What we built instead

Our pipeline is a sequence of smaller models — each modality-specific, each anatomy-specific — wrapped in an audit-trail layer that captures exactly which prior studies informed each field of the draft. The architecture is deliberately not general-purpose. It is the opposite of the foundation-model bet: we are building smaller, more specific, more auditable systems, and we are wrapping them in a layer that produces the provenance record.

The output is not a single magical model that handles all of radiology. The output is a clinical-decisioning system that produces a draft with a verifiable, defensible audit trail. The draft is the user experience. The audit trail is the procurement-grade asset.

The hard part we are not hiding from

This is harder to build than a foundation model. Foundation models benefit from scale: more compute, more data, larger context windows. Audit-trail systems benefit from institutional specificity: every hospital has its own template library, its own macro shortcuts, its own prior-study conventions, its own peer-review workflow. The audit trail has to capture all of that or it is not actually defensible.

This means our go-to-market is slower. We do not sell "an AI that reads chest CTs." We sell "a reporting workflow that integrates with your template library, your prior-study archive, your peer-review process, and your escalation chain." That is a six-month enterprise sale, not a 30-day departmental one.

We are betting that the procurement decisions of 2027 will reward the slower sale. The CMS reimbursement pathways, the medical-legal pressure on documentation, and the institutional preference for systems that survive both — all of these favor depth over breadth, defensibility over capability, provenance over performance.

What this means for radiologists

If you are a radiologist evaluating AI for your practice, the question is not "which AI has the best accuracy benchmark." The question is "which AI produces a draft I can defend in front of a CMS audit, a malpractice deposition, and a value-based-care contract renegotiation." The first question favors a foundation model. The second question favors an audit-trail system.

If you are a practice administrator, the question is not "which AI is the cheapest per report." The question is "which AI vendor will still be in business in five years, with a defensible product, with an intact regulatory pathway, and with the documentation we need to survive an audit." That is a different procurement conversation than the one the foundation-model vendors are having with you.

The next decade

Foundation models will keep getting better. The benchmarks will keep improving. The press releases will keep coming. None of that is in dispute.

What is in dispute is whether the durable value in radiology AI will accrue to the company that produces the most accurate draft, or to the company that produces the most defensible audit trail. We chose the second. We are betting that the procurement-grade asset for the next decade is provenance, not performance.

We may be wrong. If we are, we will adjust. But for now, this is the path we are building — and the path we believe will produce the most defensible, most procurement-ready, most clinically-trusted radiology AI system of the next decade.

— Aldo Ruffolo, DO, MBA
Founder, ApertureAI
Connect on LinkedIn