9 signals this week: 3 Quality, 1 Clinical trials, 5 Regulatory

No AI drug-safety tool has gone through ISTAND qualification before this LOI. If the model clears the remaining two steps, its output could stand alongside conventional preclinical data ahead of Phase 1 decisions, the qualification, not the algorithm, is what would make that usable.

Track LOI status through all three DDT qualification steps and document the model's context of use, its reference-drug training set, validation endpoints, and version history now, rather than assembling that record when FDA asks for it during qualification review.

Acting CBER Director Karim Mikhail framed the guidance as accelerating development "without compromising the rigorous scientific standards that patients and the public depend on." For rare-disease programs racing the clock, the guidance only saves time if a sponsor's applicability rationale survives review, a weak one costs more time than starting from scratch would have.

Build prior-knowledge files with source provenance, an explicit applicability rationale, CMC/nonclinical/clinical linkage, stated assumptions and limits, and a record of what FDA accepted or rejected, maintained as living evidence objects, not assembled after a submission is drafted.

Device firms often prepare for QMSR compliance in the abstract. CP 7382.850 replaces that with the actual selection logic investigators use to pick which QMS areas to dig into and how to connect a finding in one area to a clause in another.

Map each QMS area to its risk controls, process owner, records, CAPA history, supplier controls, and change-control trail, and keep evidence that top management, not just quality, is acting on the risk priorities CP 7382.850 tells investigators to look for.

If AI shapes which health problems get studied and how the underlying evidence gets synthesized, a governance gap at that stage produces policy that looks evidence-based while quietly inheriting whatever framing the model started with, and that error surfaces downstream, in decisions no one traces back to it.

Document AI-use cases at each policy-cycle stage, run algorithmic impact assessments and readiness reviews before deployment, and keep, in WHO's words, "human-in-the-loop decision gateways, and multidisciplinary oversight panels combining domain, methods and ethics expertise," with clear accountability for the final interpretation.

Mayo CEO Dr. Gianrico Farrugia described it as "building something new in healthcare" by pairing Mayo's data foundation with Microsoft's AI capabilities, a bet that institution-specific clinical data, not general model capability, is what makes a healthcare AI model trustworthy enough to deploy at the bedside.

Set explicit AI ownership rules, de-identification controls, validation protocols ahead of clinical use, ongoing monitoring plans, clinical oversight for model-supported decisions, and access controls for whoever eventually reaches the model through Azure Foundry.

A trial that validates an AI system once, at launch, says nothing about whether that system still performs the same way after it has retrained on six months of new data. The paper's point is structural: if the model keeps changing, the trial design has to keep watching, not just check in once.

Build lifecycle protocols now that separate delivery-monitoring from oversight-monitoring: version control and update-approval gates, defined performance-drift thresholds, and documentation of every change made after deployment, not just the validation record from before it.

Twenty-one inspections, zero critical deficiencies, and documentation still comes out as the top non-compliance category, source records especially, per TGA's own breakdown. Sound trial design doesn't protect against weak paperwork, and this report shows sponsors exactly where that gap sits.

Map trial records against TGA's finding categories, tighten source-data and documentation controls specifically, keep deviation rationale on file, evidence training and delegation logs, and close any CAPA with a traceable effectiveness check rather than a sign-off alone.

An agent that performs a task, rather than surfacing an insight for a person to act on, moves the accountability question. Sanofi Chief Digital Officer Emmanuel Frenehard framed the goal as letting teams "operate with greater speed, depth, and confidence", which only holds if the agents' outputs are checked, not just trusted.

Define each agent's role and permitted tasks, what data it can touch, who reviews its output before action, escalation points when it's wrong, an audit trail of what it did and why, and clear ownership of any decision made on an agent's recommendation.

A JSC happens before the clinical program locks in, which is the only point where a comparator choice or endpoint decision made for regulatory purposes can still be checked against what HTA bodies will actually want to see later.

Track JSC eligibility against Article 7(1) criteria, plan the briefing document and platform registration well ahead of the 1 July deadline, and keep regulatory, clinical, HTA, statistics, and market-access functions aligned on comparator logic before the consultation, not after.