Phase I — Foundation

Years 1–5. The phase where individual capabilities are demonstrated independently and the integration patterns that will matter later are established. Mostly low-risk, high-value work that builds infrastructure for everything subsequent.

Years 1–5

Risk profile Low

Decision gate Architecture viability

Strategic objective

Phase I is about getting the field ready to attempt AIHS. No single advance in this phase produces an AIHS capability. What this phase produces is the substrate — datasets, infrastructure, validated component technologies, regulatory pathways, and standards — without which the harder phases cannot begin.

The most important deliverable of Phase I is not a technology. It is the demonstration that the three buckets can be developed in tightly-coupled rather than siloed fashion, with cross-bucket integration patterns visible early. A field that exits Phase I with three parallel research communities that don't talk to each other will struggle in Phase II.

Bucket A · Diagnostic priorities

Three Bucket A advances see meaningful Phase I investment. A3 (continuous temporal sampling) is closest to deployment — the existing biosensor industry is mature and Phase I work concentrates on extending sensor breadth, improving biocompatibility for chronic implantation, and developing standardised data formats. A4 (integrated state representation) piggybacks on commercial foundation-model development; the Phase I institutional work is providing standardised medical training corpora and benchmarks rather than building models from scratch. A1 (in vivo single-cell readout) remains at proof-of-concept in animal models — Phase I produces the first chronic in vivo cellular readouts, in mice and non-human primates, demonstrating one or more candidate approaches.

A2 (whole-body molecular mapping) and A5 (distributed pathogen and damage detection) are not Phase I priorities — they depend on capabilities not yet mature and would represent premature investment.

Bucket B · Interpretation priorities

B1 (complete variant effect prediction) is a Phase I priority for two reasons. First, the data infrastructure (functional genomics screens, perturbation atlases) requires sustained build-up that should start now. Second, the foundation models being trained for variant interpretation set the engineering precedent for later Bucket B work; even the early ones — AlphaMissense-class systems — produce clinical value.

B5 (continuous learning across patients) is the most tractable B advance and a Phase I deliverable. The technical pieces exist; what's needed is regulatory clarity, institutional coordination, and workflow integration. By end-of-phase, federated medical AI should be in routine production use at multiple academic medical centres.

B3 (causal disease modelling) is the central scientific challenge but cannot be solved in Phase I. The Phase I work is scaling perturbation biology and laying the data foundation. Whether B3 will yield to current approaches becomes clearer by end of Phase II.

Mechanistic modelling efforts on tractable organ systems — heart, liver, immune compartments — should advance under the umbrella of B4 (digital twins) preparation, even though full digital twins are not a Phase I deliverable.

Bucket C · Therapeutic priorities

C1 (multi-target coordinated intervention) work begins with programmable cell therapies featuring increasingly sophisticated logic-gate behaviour. The Phase I deliverable is logic-gated cell therapies in clinical trial for at least one indication, demonstrating that the platform principle works in patients.

C2 (universal tissue access) investment concentrates on the blood-brain barrier — focused ultrasound techniques moving through clinical trial, antibody-shuttle technologies maturing, AAV-based CNS delivery expanding in indication. The Phase I goal is robust, repeatable CNS access for at least gene-therapy payloads.

C5 (bounded reversible agents) is a Phase I priority in parallel with C1. Every therapeutic agent being developed should be developed with explicit reversibility from the start; retrofitting kill switches into mature agents is harder than designing them in.

C6 (multi-layered safety architecture) is mostly standards and validation work at this stage — FDA guidance documents, professional society standards, validation methodologies. Boring infrastructure work that pays off enormously in later phases when capability outruns governance.

Integration milestones

The cross-bucket work in Phase I focuses on infrastructure that makes integration possible later, not on demonstrating integration directly.

Limited closed-loop pilots for specific indications. Continuous glucose with insulin delivery is the existing template. Phase I expands the pattern to one or two additional indications — perhaps automated anaesthesia depth control, perhaps closed-loop dialysis. Establishes that closed-loop autonomous medical action can pass regulatory review when scope is bounded.
Cross-bucket research consortia. At least two funded consortia bringing Bucket A and Bucket B work together (diagnostic + interpretation) and at least one bringing Bucket B and Bucket C together (interpretation + delivery). Funding agencies should require cross-bucket collaboration as a condition of major grants.
Standards for autonomous medical AI safety. Published professional society standards (FDA guidance, equivalent international) for autonomous decision systems in medicine. Sets the architectural ground rules before capability becomes available to deploy without them.
Reference test bed for closed-loop autonomous systems. A shared evaluation environment where new autonomous medical systems can be benchmarked against a standardised set of clinical scenarios. Modelled on benchmark infrastructure in AI research. Stretch goal — likely to slip into Phase II.

Funding allocation profile

Rough proportions for how an integrated AIHS programme should distribute funding in Phase I. Numbers are illustrative — the point is the shape, not specific percentages.

Phase I · Approximate allocation

A · 28%

B · 25%

C · 22%

Safety · 15%

Integ · 10%

Diagnostics

Interpretation

Therapeutics

Safety architecture

Integration

The notably high safety-architecture allocation (15%) reflects the recommendation that safety be a first-class research investment from the start. The historical default is 1–5%; the AIHS architecture argues for substantially more.

Dominant risks during Phase I

The Phase I risk profile is genuinely low for most of the work. The infrastructure being built is well-understood, the proof-of-concept demonstrations target known-tractable problems, and the regulatory work has clear precedents. The two non-trivial risks:

Siloed development. The strongest research groups in each bucket have their own funding ecosystems, communities, and incentive structures. Without deliberate funding pressure toward cross-bucket coupling, Phase I will produce three excellent siloed fields that don't know how to integrate. This risk is best managed by funders requiring cross-bucket collaboration in major grants.

Underinvestment in safety standards. Phase I is the time to build safety architecture norms when no one is yet shipping autonomous medical systems at AIHS scale. Once such systems exist, the leverage funders have to shape standards drops dramatically. Underinvesting in safety architecture now produces a much harder problem later.

End-of-Phase I Gate

Architecture viability assessment

Are the three buckets developing in genuinely coupled fashion, with integration patterns visible early — or in parallel silos with weak coupling?

This is the operational question Phase I exists to answer. If integration patterns are emerging and cross-bucket teams are productive, Phase II can scale toward capability. If the buckets have become parallel silos, Phase II priorities should pivot heavily toward integration infrastructure before any further capability investment.

If integration is healthy

Proceed to Phase II as planned. Scale capability work within each bucket. Begin Bucket B central-problem (B3) push.

If integration is weak

Reweight Phase II toward integration infrastructure. Delay capability scale-up by 2–3 years. Consider funder-level intervention on grant structure.

Research Questions

Open questions from the field

Detail-level questions a researcher or programme planner landing on this page would likely ask. Expand each to see the answer.

What concretely has to exist at the end of Phase I for Phase II to be able to start?

Phase I succeeds not by producing an AIHS capability but by producing the substrate that makes the harder phases attemptable. Concretely, four things must exist at exit.

Validated component technologies. Each bucket needs at least its low-risk advances demonstrated at proof-of-concept: continuous sensing (A3), integrated state representation (A4), continuous-learning infrastructure (B5), and the first elements of bounded, safe therapeutic agents (C5, C6). These are the advances with no fundamental scientific blockers, so they are the reasonable Phase I targets.

Datasets and reference maps at the scale the interpretation bucket will need — the perturbation atlases, cell atlases, and longitudinal multi-omic cohorts that later causal-modelling work trains against.

Regulatory pathways — not approvals, but the beginnings of the frameworks for validating adaptive, multi-component systems, since the body of this phase stresses that regulatory science is itself a deliverable.

Demonstrated cross-bucket coupling. The single most important Phase I deliverable is evidence that the three buckets can be developed in a tightly-coupled rather than siloed way. A field that exits Phase I with three parallel but disconnected research streams has, in the terms this study uses, failed Phase I even if every individual technology advanced — because the integration that defines AIHS would have been deferred to a point where it is far harder to retrofit.

Why is cross-bucket coupling the make-or-break property of Phase I rather than any specific technology?

Because the failure mode it guards against is the most likely one, and the most expensive to reverse. The natural institutional tendency is for diagnostics, interpretation, and therapeutics to develop in separate communities, on separate grants, with separate conferences — each making real progress, none building toward integration.

If that happens through Phase I, then Phase II arrives with three mature-but-incompatible streams that were never designed to interoperate: sensors whose outputs no interpretation layer was built to consume, models trained on data that no diagnostic actually produces, therapies with no closed-loop control interface. Integrating them after the fact means re-engineering each stream against the others — far costlier than building the interfaces in from the start.

This is why the study treats coupling as a Phase I property to be actively demonstrated, not assumed. The A4 integrated-representation work and the integration agenda are the concrete vehicles: an early integrated demonstration, even a crude one, proves the interfaces exist and forces the three communities to design against shared standards. The cost of establishing coupling early is modest; the cost of retrofitting it late can sink the programme.

What would indicate Phase I is failing, even if individual technologies are progressing?

The warning signs are organisational and structural rather than technical, which is what makes them easy to miss while celebrating component milestones.

Siloing is the primary one: if diagnostic, interpretation, and therapeutic efforts have no shared data standards, no joint demonstrations, and no common integration roadmap by mid-Phase I, the coupling deliverable is slipping regardless of how impressive each silo's individual results are.

Regulatory-science neglect is the second: if validation frameworks for adaptive, multi-component systems are treated as a later-phase problem rather than active Phase I work, the programme is accumulating a debt that will halt deployment later — a technically successful system that no framework can approve.

Safety-as-afterthought is the third: if the safety architecture (C6) and bounded-agent (C5) work is deferred in favour of capability demonstrations, the programme is heading toward the failure mode the whole study is organised to prevent. Phase I is where safety must be established as a first-class constraint, because retrofitting it onto mature capability is exactly what does not work. A Phase I that produces dazzling component demos but none of these structural foundations has failed at its actual job.

Referenced from across the study

Topic Roadmap — → Full Phase I detail with funding allocation, risks, and decision gate Phase II — Capability building (Years 5–15) The second phase scales the individual capabilities and begins serious cross-bucket integration.

Continue

Next phase

Phase II — Capability Building

Years 5–15. Where individual capabilities scale and the central scientific challenge (B3) gets its first real test.

Back

Roadmap overview

Return to the four-phase summary view.

Sideways

Decision gates

The three explicit review moments across the full roadmap.