W23: intelligence moved into controlled work.

FDA accepted a Letter of Intent for an AI-driven digital liver model into the ISTAND drug-development-tool qualification program. That is the first formal door opening, not the final regulatory acceptance of the model. The sponsor still has to submit a Qualification Plan and then a full qualification package before the tool can be used as a qualified DDT within a defined context of use.

The model is intended to predict drug-induced liver injury by comparing chemical structures of new drug candidates against historical reference drugs with known DILI risk. FDA frames it as a New Approach Methodology that could complement, not replace, other DILI risk-assessment methods as part of a weight-of-evidence approach.

The real signal is that AI models are beginning to look like regulated evidence objects. For QMS, regulatory intelligence, and AI governance teams, the issue is no longer simply whether a model performs well internally. The question is whether its training data, context of use, validation endpoints, limitations, version history, change control, and human review can be defended.

FDA’s draft guidance is about using what is already known without pretending every product starts from zero. For human gene therapy products, including genome editing in somatic cells, FDA says sponsors may be able to draw on publicly available information and established platform knowledge across CMC, nonclinical, and clinical areas.

The important boundary is that prior knowledge has to be justified. Sponsors must explain why the data they want to reuse are relevant to the specific product, mechanism, manufacturing approach, and development question. FDA also points sponsors toward early engagement, including INTERACT and pre-IND meetings, so the reuse strategy can be discussed before it becomes a submission problem.

For iFeed, this is a major evidence-operations signal. Acceleration is not coming from cutting corners; it is coming from better evidence architecture. Teams that can maintain platform knowledge, bridge rationales, traceability, and decision records will be better placed than teams that only assemble documents at submission time.

CP 7382.850 is the investigator-facing operating layer behind the QMSR transition. The document explains how FDA staff should conduct risk-based inspections of medical device manufacturers and evaluate whether the QMS provides reasonable assurance that devices will be safe and effective.

The program explicitly ties QMSR to ISO 13485:2016 and emphasizes risk management and risk-based decision-making inside the QMS. FDA describes the QMS as linked processes and expects top management to integrate QMS processes and embrace a culture of quality. Inspectors may review records to evaluate whether product risks are adequately controlled, and records are selected based on identified risks and investigator judgement.

For iFeed and QMS platform thinking, this is a foundational signal. Inspection readiness is not a binder; it is the ability to show how risk, design, supplier, production, change, complaint, CAPA, MDR, UDI, and records connect. A useful QMS cockpit should help teams navigate those links before an inspector asks for them.

WHO’s signal is that AI governance cannot stop at clinical deployment. AI can influence the evidence base before a policy decision is even made: which problems are defined, which options are compared, which outcomes are optimized, and which voices are visible in the data.

The paper names concrete risks: biased data can distort problem definition, over-optimization can narrow solution design, digital divides and cybersecurity can weaken implementation, and monitoring tools can slowly shift policy away from its original goals. WHO also highlights epistemic injustice, where quantifiable data can crowd out lived experience, local expertise, Indigenous knowledge, and community insight.

The practical governance move is clear: algorithmic impact assessments, technology readiness reviews, living evidence workflows with human verification, human-in-the-loop gates, and multidisciplinary oversight. iFeed reads this as a strong operating principle: AI may accelerate evidence work, but humans remain accountable for framing questions, judging quality, interpreting context, and making ethical trade-offs.

The Mayo-Microsoft signal is about healthcare AI moving from narrow tools to institutional models. Mayo describes a frontier model built specifically for healthcare, using Mayo’s clinical expertise, de-identified data foundation, and longitudinal insight together with Microsoft’s AI and cloud capabilities.

Two details matter. First, Mayo says the model will be owned by Mayo, which keeps trust, stewardship, and clinical governance close to the healthcare institution. Second, the model is initially deployed inside Mayo’s trusted clinical environment for continuous testing, refinement, and real-world improvement before broader access through Azure Foundry APIs.

For iFeed, this is the operating-infrastructure story in clinical AI. The hard work is not just model capability; it is clinical context, longitudinal data, governance, validation, deployment boundary, and accountability. Organizations watching this should ask whether their AI strategy is still point-solution procurement or whether they are building governed clinical intelligence infrastructure.

The Nature Medicine signal addresses a problem regulators, sponsors, and hospitals will face more often: many AI systems are not static interventions. They may be continuously monitored, updated, recalibrated, or adapted as data and workflows change.

That creates tension with traditional clinical-trial logic. A trial can evaluate a fixed intervention at a point in time, but an AI system that changes during or after deployment may need a design that explicitly separates two things: monitoring that is part of the AI intervention itself, and monitoring that is part of trial oversight.

For iFeed, this is a methodology signal. Clinical AI evidence cannot rely only on launch validation. Teams need pre-specified monitoring, update governance, performance drift checks, safety triggers, version documentation, and decision rules for when an update is allowed, paused, or re-evaluated.

TGA’s 2025 GCP Inspection Program report is a practical reminder that clinical-trial quality is inspectable evidence, not a policy aspiration. The report covers 21 inspections, including medicine, biological, and medical-device trials under Australia’s CTN and CTA schemes.

The headline is balanced: no critical deficiencies were identified, and TGA says all sites had resolved issues through CAPA by the time of publication. But the most important operational signal is where weaknesses appeared. Documentation had the highest level of non-compliance, source documentation had the most deficiencies within that category, and ALCOA-C data-integrity principles were central to the documentation findings.

For iFeed, this keeps W23 grounded. AI, adaptive trials, and evidence reuse only work if the basic inspection file is strong: consent content, protocol deviations, source records, training, contracts, investigational product controls, and CAPA. Quality teams should treat the report as a checklist for where trial operations become regulator-visible.

Owkin’s Sanofi collaboration is useful because it names the next step in pharma AI: embedded agents, not just analytics dashboards. Owkin says it will lead end-to-end development of novel AI-driven biopharma agents for Sanofi, deployed through K Pro, Owkin’s AI Scientist platform.

The source describes agents that work like intelligent assistants, autonomously performing complex tasks in drug research and development. K Pro is positioned as combining multimodal patient data with specialized biological and agentic AI systems to support stages from early discovery through clinical development, while also providing competitive intelligence.

For iFeed, the signal is workflow ownership. Once AI agents enter R&D workflows, governance cannot be limited to model selection. Teams need rules for which tasks agents can perform, how outputs are reviewed, what data they can access, how decisions are recorded, and how agent performance is monitored over time.

The EU HTA signal is not just an administrative window. It is a reminder that evidence planning now has to satisfy more than one audience. Developers can request Joint Scientific Consultations so they can discuss clinical-study planning before later Joint Clinical Assessment under the EU HTA Regulation.

The window opened on 3 June and runs to 1 July 2026. It applies to medicinal products and to selected higher-risk medical devices and diagnostics: Class IIb or III medical devices and Class D IVDs. The practical point is timing. If developers wait until evidence is already generated, it may be too late to align the study design with assessment expectations.

iFeed reads this as market access moving upstream into evidence architecture. Regulatory approval, HTA assessment, and reimbursement readiness cannot be treated as separate documents at the end. Teams need an integrated evidence plan that anticipates clinical endpoints, comparators, population, outcomes, and submission logistics early.