In THIS ISSUE Signal of the Week,
Novo Nordisk rents an agent layer from AWS.
Framework of the Week, the Agentic Medical Affairs Map
5 Things Worth Reading, deals, regulation, and the clinical AI market
Tool of the Week, Vera HealthReading, time: ~8 min read
🎯 Signal of the Week
Novo Nordisk did not buy AI this week. It rented the layer that runs it.
Welcome to Issue 03,
On 10 August, Novo Nordisk and Amazon Web Services announced a partnership that makes AWS Novo's preferred cloud provider and AI partner, and stands up a hub inside Novo's existing London facility to work on accelerating new medicines. The wire read it the way wires read these things: Novo doubles down on AI, another cloud deal, faster drug discovery. All three are accurate. All three miss the sentence that does the real work.
Read the release for the verbs, not the adjectives. Novo gets Amazon Bio Discovery, which serves scientists AI models trained on biological datasets, and Amazon Bedrock to build AI applications that identify targets, design therapies, and pull clinical, genomic, and imaging data together. Then comes the tell: Novo will deploy Amazon Bedrock AgentCore, described by AWS as managed infrastructure to help developers deploy and scale AI agents, to make its operations more efficient. AWS framed the announcement inside its newly launched Forward Deployed Engineering organization, where AWS engineers embed with a customer to co-develop and design agentic AI across the business. So Novo did not buy seats. It rented an agent operating layer and the people who wire it in.
Why this is the consequential read for Medical Affairs, not just R and D: the productivity that Novo is already claiming did not land in the lab. AWS and Novo say the collaboration has already reduced clinical documentation time and made 25,000 employees more productive. Documentation, evidence handling, and cross-functional data are Medical Affairs terrain. When a top-tier pharma answers the build-versus-buy question with rent the agent layer, keep the science, it sets the template your own vendor conversations will be measured against. The quiet part is that the agent, not the chatbot, is now the unit pharma is contracting for.
Figure 1. The deal as reported versus the architecture it commits to. Facts from the 10 August announcement as reported by Fierce Biotech
One more thing worth filing. This is an extension, not a first date: Novo already runs Amazon Pharmacy, Amazon Ads, and Amazon One Medical across its business, works with OpenAI from discovery to commercial operations, and has signed up as a customer of Denmark's first supercomputer. AWS, for its part, has struck similar AI pacts with Flagship Pioneering and with Johnson and Johnson. The pattern is not one pharma picking one vendor. It is the largest players standardizing on hyperscaler agent platforms plus embedded engineers as the default way AI enters the building. If you lead Medical Affairs, the strategic question is no longer whether to use AI. It is who owns the agent layer your team will be asked to work inside, and on whose terms.
🧠 Framework of the Week
The Agentic Medical Affairs Map
Grounded in Tika's 2026 analysis of agentic AI in Medical Affairs
Start with the distinction that most content blurs. Generative AI responds. You prompt it to summarize a paper, draft an inquiry response, or suggest talking points, and it returns an output. It is reactive by design. Agentic AI is structurally different: autonomous, goal-oriented systems that monitor data streams, reason through multi-step tasks, and recommend or execute next actions with minimal prompting. The MAPS Insights Forum framed it as technology that autonomously sets goals and executes tasks while adapting to a changing environment. In plain terms, generative AI is the tool you open when you need something; agentic AI is the colleague already working when you arrive.
Why now, and why Medical Affairs. McKinsey's review of more than 270 life sciences workflows found that 75 to 85% contain tasks that agents could enhance or automate, potentially freeing 25 to 40% of organizational capacity. At the same time the ground is shifting under the field: pricing pressure, the Veeva and Salesforce CRM split forcing platform re-evaluation, and rising scientific complexity. The window is real, and so is the failure mode. Gartner estimates that more than 40% of agentic AI initiatives will be cancelled by 2027 if they are not anchored in clear business value. The lesson is not adopt faster. It is choose where to start with intent.
The map: one MSL day, three moments. The clearest way to see agentic AI in Medical Affairs is not a capability list, it is a working day. Before the meeting, an agent has scanned overnight literature across the therapeutic area, flagged the two publications relevant to today's conversations, and assembled a pre-call brief from engagement history and the KOL's research interests. During the interaction, speech-to-text is only the surface; the agent classifies insights in real time, tags a formulary comment as market access, and flags a possible off-label question for review. After the call, it routes each insight to the right owner, market access, commercial intelligence, or pharmacovigilance, updates the CRM, and refreshes the KOL profile. The point is the loop, not any single feature.
Figure 2. The agentic loop across a single field day. Structure and examples from Tika Mobile.
The CRM is the foundation nobody connects to AI. You can deploy the most capable agents in the field and still get nothing, because agents need inputs, and enterprise CRM adoption fails between 50 and 63% of the time. The barriers are specific: data entry that feels punitive, desktop-first designs that break on mobile between meetings, and integrations that force four apps to log one interaction. The flywheel only turns one way: higher adoption yields richer data, which makes agents more accurate, which makes the CRM worth using. Starve the loop and the agent produces generic output that confirms the field's suspicion that the platform is not worth it.
Compliance is the architecture, not a review layer. Privacy and compliance are the number one barrier to adoption, cited by 69% of Medical Affairs leaders in a recent survey, and that caution is earned given off-label, MLR, and adverse-event rules. The better framing is not AI versus compliance but where the checks live. In a sound design, the agent cross-references a draft against current labeling and safety data before a human sees it, flags potential off-label language during the interaction, and logs every decision for audit. Done this way, agentic systems can improve consistency versus manual judgment: ACMA data suggest standardized, AI-supported workflows lift compliance by 30% and effectiveness by up to 35%.
HOW TO START: THREE PHASES
Phase 1, pick where it hurts and make data agent-ready. Take the 2 to 3 workflows with the highest admin burden and lowest strategic value, usually pre-call planning, insight capture, and literature monitoring, and set consistent data tagging. Phase 2, crawl then walk (3 to 6 months). Ship one discrete capability, earn confidence, then connect it to the next. Phase 3, scale with governance and measurement. Build auditability from day one, set thresholds for human oversight, and align Medical Affairs, compliance, IT, and data science. The prize is concrete: MSLs currently spend 40 to 80% of their time traveling, early adopters report 20 to 35% less admin, and McKinsey estimates agents could add 3.4 to 5.4 points of pharma EBITDA over 3 to 5 years.
Figure 3. The magnitudes behind the map. McKinsey, Gartner, and survey figures as compiled by Tika Mobile.
Generative AI is the tool you open when you need something. Agentic AI is the colleague already working when you arrive.
📚 5 things worth to read
Deals, regulators, and the clinical AI market
Five paywall-free reads, each summarized so you get the value without the click
01
Pharma bets big on AI platforms, and the deals are about infrastructure, not moleculesThis piece is the frame that makes the Novo and AWS deal legible: 2026 opened with a run of AI platform deals that buy models and infrastructure rather than single assets. Chai Discovery partnered with Eli Lilly to design novel biologics across multiple targets and to build a Lilly-exclusive model trained on Lilly's proprietary data; Chai had just closed a 130 million dollar Series B at a 1.3 billion dollar valuation. Noetik signed a five-year licensing deal with GSK for its non-small-cell lung and colorectal cancer foundation models, with a 50 million dollar upfront payment and a subscription framework that its CEO calls one of the first true foundation-model licensing deals in biotech. Boltz, launched with 28 million dollars, locked in a multi-year Pfizer collaboration, and Isomorphic Labs added Johnson and Johnson to deals with Lilly and Novartis. The common thread, in the words of one investor quoted, is that value now sits in the systems that make discovery faster and more reliable, not in a single better model. For Medical Affairs, the takeaway is directional: your R and D colleagues are contracting for capability layers, and the evidence base you communicate will increasingly be generated by them. The detail to remember: models are commoditizing, so pharma is paying for proprietary data and workflow, not novelty.
02
Regulatory trends in AI-enabled software as a medical device
If you want a sober map of how regulators are actually clearing AI in medicine, this is it, and it is fresh. More than half of all FDA authorizations for AI-enabled software as a medical device have occurred since 2022, and 76% sit in radiology, which tells you where the evidence and the scrutiny are concentrated today. The authors organize the field into five building blocks that climb from single-data-type interpretation (imaging, then in-vitro diagnostics) toward multimodal diagnosis and, eventually, personalized treatment planning, borrowing radiology's triage, detection, and diagnosis vocabulary. They note the EU MDR route has produced widespread CE marking of imaging tools, with examples such as Canon's CT VScore Plus and Siemens MAGNETOM systems, signaling integrated AI in European practice. Crucially, they place treatment-directing tools like DreaMed Advisor Pro and the Tempus and Geneseeq sequencing analyses at the current frontier, and forecast the first true multimodal clearances only around 2030 to 2032. For Medical Affairs, this is a reality check on how far, and how slowly, autonomous decision support is being allowed near patients. The line to keep: computation is no longer the bottleneck; diverse, compliant data and rigorous clinical validation are.
03
The FDA's first draft guidance on AI in drug development
A week into the year, the FDA issued its first draft guidance on using AI to develop drugs and biologics, and this explainer is the cleanest short read on what it means. The regulator is not policing day-to-day AI use; it is setting expectations for when AI helps decide whether a product is safe and effective. The proposed approach is a risk-based credibility framework: sponsors describe how they will use AI (the context of use), assess the risk that use carries, and gather evidence that the model's outputs are reliable for that purpose. The piece makes the boundary vivid with a worked example, using AI to speed literature review or standardize notes with humans deciding next steps sits outside the guidance, while using the same model to select which safety signals go into a submission falls squarely inside it. The draft was open for comment until 7 April, and it followed years of exponential growth in AI-containing submissions. For Medical Affairs and evidence teams, the durable lesson is that credibility now starts with the source data and the documented context of use, not the model. The detail to keep: efficiency uses are low-stakes, decision-influencing uses are where the regulator wants proof.
04
From support to strategy: will 2026 be the year of the AI-augmented Medical Affairs team?
The Journal of mHealth, Natalie Yeadon (Impetus Digital), 19 February 2026
This is the strategic companion to this week's framework, written from inside the function. Yeadon argues Medical Affairs has already earned its seat, so 2026 is about scaling, not proving value, and AI is becoming the operating system that makes scaling possible. She sketches the MSL 2.0: static CRM notes and quarterly check-ins give way to real-time analysis of prescription flows, publications, and engagement signals that individualize KOL strategy, while AI serves pre-approved, KOL-specific materials so conversations get sharper rather than more generic. She is direct about the MLR bottleneck, positioning LLMs trained on verified internal data as a drafting desk that can cut first-pass response letters and slide decks from days to hours, provided skilled writers still own scientific nuance and final validation. The compliance caution is explicit: tracing an AI-generated insight back to its source data will be non-negotiable, and success depends more on culture and governance than on tools. Her closing line lands: AI will not replace Medical Affairs professionals, but those who use it well may replace those who do not. The detail to keep: the shift is from retrospective dashboards to a continuously closing insight-to-action loop.
05
The clinical AI landscape in 2026, and why geography decides what you can use
Before you standardize your field team on a clinical AI tool, read this taxonomy, because availability and guideline alignment vary sharply by region. Tytler splits the market into general-purpose models (ChatGPT, Claude, Gemini, Perplexity), purpose-built clinician tools (OpenEvidence, iatroX, Medwise, Glass), AI scribes, and enterprise systems, and stresses that none of the general models carry regulatory status or reliably cite local guidelines. The eye-opener for European readers is concrete: OpenEvidence, valued at 12 billion dollars after a 250 million dollar Series D and handling roughly 15 million consultations a month, is largely funded by pharmaceutical advertising and remains US-centric in guideline coverage. Meanwhile a cited Nature Medicine study found 52% undertriage of emergencies in the consumer ChatGPT product, a reminder that fluent is not the same as safe. Tytler's thesis is that as model capability converges, the real differentiator becomes does it know my medicine, meaning local guidelines, plus workflow integration inside the record. For Medical Affairs, that reframes tool selection as a governance and geography question, not a capability race. The detail to keep: US-trained models can quietly cite US guidelines for non-US clinical questions.
🔧 TOOL OF THE WEEK
Vera Health, the free clinical answer engine that still works in Europe
A practical pick for MSLs and Medical Information teams, chosen with the last read in mind
Last week's reads make the point that most of the well-known clinical AI is US-first, ad-funded, or geo-blocked. Vera Health is the counter-example worth putting in front of a European field team. It is an AI-powered clinical decision-support engine that synthesizes more than 60 million peer-reviewed papers, guidelines, and care pathways into cited, evidence-graded answers, paired with more than 900 clinical calculators and curated medical news. It was built by AI researchers from MIT with clinicians from institutions including Mayo Clinic and Yale, is validated in emergency medicine through a partnership with the American College of Emergency Physicians, and reports more than 300,000 users. It is free for licensed clinicians and students, and, unlike the ad-and-pharma-funded option that pulled out of the EU and UK in April 2026, it is available globally.
WHAT IT IS
A cited, evidence-graded clinical answer engine over 60 million-plus articles and guidelines, plus 900-plus calculators and curated news, offered free to verified clinicians and students worldwide.
HOW TO USE IT FOR MEDICAL AFFAIRS AND MSLS
Pull a fast, cited evidence brief before a KOL meeting; sanity-check a clinical question against graded sources; scan curated news to stay current in your therapeutic area; and lean on multilingual support (English, French, Spanish, Italian, German, Japanese, and more) for cross-border field teams. Use it as a research and preparation aid, never as a source of external or promotional copy.
WHERE TO ACCESS IT
verahealth.ai on the web, plus iOS and Android apps. Free for licensed healthcare professionals and medical students, with no US-only restriction and no paid tier gating core use.
COMPLIANCE NOTE
Vera Health is decision-support that augments, not replaces, clinical and scientific judgment, and it states as much. The headline accuracy figures (for example 97.5% on USMLE and 84.9% on NEJM-AI) are vendor-reported and should be read that way. It is HIPAA and GDPR compliant, but treat it as an internal research aid: do not use its output to generate MLR-bound or promotional materials, keep a human in the loop on anything that leaves your team, and do not paste patient-identifiable or unpublished company data into any external tool. As always, verify every citation against the primary source before you rely on it.
That is Issue 01. If a colleague forwarded this to you, you can get it yourself every Monday at newsletter.ichealth-ai.com/subscribe.
One email a week. No sponsorship. See you next Monday.
iCHealth Pathway is a weekly research note by Dr. Issam Chebouti on AI transformation for Pharma, Medical Affairs, and Healthcare leaders.
iCHealth-AI, ichealth-ai.com, [email protected],
https://www.linkedin.com/in/dr-issam-chebouti-26bb30149/
Issue 03, Monday, August/17th/ 2026.
Parts of this issue were produced with AI assistance and checked before publication.