AI research, explained in one morning email
Most of what shapes AI next year is sitting in a research paper today. Keeping up with them is a full-time job, and the abstracts don't help.
TLDR AI is the newsletter that does the translating. Each morning you get the handful of papers worth knowing about, curated by Anthropic and ex-Google engineers, with a plain-English summary of what they found and why it matters.
You come away smarter about AI research in the time it takes to finish your coffee.
Free, and delivered to 1.1M+ inboxes every morning.
|
AI Just Showed Up on the Earnings Call Pharma CFOs put AI savings in their numbers this week. Boston Scientific put a number on a cyberattack. Your CFO noticed both. Life Sciences CIO Weekly • Coverage: October 26 to November 1, 2026 Q3 earnings calls carried a new kind of number this week. Finance chiefs started putting AI and cyber on the income statement.
When the CFO starts counting something, the CIO gets asked to prove it. This week’s notes are about being ready for that question. 🔍 The Quick Read
Reading something a colleague should see? Forward the whole edition or copy the section that fits their world. If this reached you from someone else, subscribe free here to get every edition, plus access to the back archive. 📌 This Week’s LeadNovartis and IQVIA put AI savings on the earnings call ↗AI is now showing up in what companies pay outside firms, and finance is the one counting it. Novartis reported third-quarter results on October 27 with a core operating margin of 41.6% (Novartis). On the call, CFO Mukul Mehta said selling, general and administrative costs fell 7% while sales rose 4%. He said Novartis spends “upwards of $11 billion every year” with third parties. Agency, travel and consulting spend has been “significantly impacted by use of technology, use of AI,” and he wants Novartis’s fair share of those supplier gains. CEO Vas Narasimhan told CNBC that AI had tangibly improved the P&L. The same day, IQVIA reported revenue of $4.43 billion, up 8%, and said it has deployed 415 AI agents (IQVIA). CEO Ari Bousbib said agents now handle “automated tasks that we used to standardize and perform in offshore centers,” including software testing, finance and HR work (transcript). AI in IQVIA’s own processes helped lift its operating margin. For a client’s Phase III program across nearly 50 countries, AI site modeling cut nine months. What to watch: IT services and outsourcing partners are getting cheaper to run. Expect finance and procurement to ask what share of that shows up in your next renewal. 🤖 AI & DataAnthropic reports Claude cut batch-record review time 89% at Enveda ↗AI document review is moving into GMP manufacturing records, with Quality still making the call. On October 27, Anthropic published results from six drug-development partners and disclosed that it had acquired Mithril Medicine, a one-person rare-disease company (Anthropic). The most relevant result for IT and quality leaders came from Enveda. Its Claude-based tool reviewed 1,061 pages of batch records from a Phase 2 campaign. It caught all 11 issues that human reviewers found, plus 7 more that needed action. Review time fell from 25.9 hours to 2.8 (Enveda). The tool has read-only access, and Quality keeps the batch-release decision. At Bristol Myers Squibb, Claude helps assess early-stage clinical trial plans, and BMS plans to expand that work in 2027. The other results are preclinical and self-reported. (Disclosure: this newsletter’s research uses Claude.) What to watch: Whether Enveda or others publish validation documentation for AI-assisted batch-record review. Waters builds AI peak integration into Empower as a paid tier ↗AI is arriving inside a validated QC system of record through a vendor subscription. On October 28, Waters launched AI-Powered Peak Integration for Empower, the chromatography data system it says reaches about 500,000 users (Waters). In a 2025 pilot with a large pharma company, manual interventions across 400 chromatograms fell from 90 to 4. The models are version-controlled and “no-drift,” so they stay fixed until a lab approves a new one. Early access starts in December, with validated deployment planned for mid-2027 (SelectScience). What to watch: Change control, validation and licensing talks with QC before mid-2027, and similar paid AI tiers from other lab software vendors. Dassault closes ArisGlobal and puts Medidata, BIOVIA and ArisGlobal under one leader ↗One vendor now spans lab, clinical, safety and regulatory systems at many large sponsors. On October 28, Dassault Systèmes completed its $1.8 billion acquisition of ArisGlobal, with up to $200 million more tied to AI revenue milestones (HPCwire). It is combining BIOVIA, Medidata and ArisGlobal into one AI platform. Dr. Ashwini Zenooz becomes CEO of Life Sciences and Healthcare over all three brands, reporting to CEO Pascal Daloz. ArisGlobal serves more than 200 clients, including half of the top 50 biopharma companies. What to watch: Bundling and pricing at your next Medidata or ArisGlobal renewal, and whether validated versions stay stable through integration. Pfizer deploys Edison Scientific’s AI scientist across its R&D pipeline ↗An outside AI agent gets continuing access to a big pharma company’s historical R&D data. On October 29, Edison Scientific said it will deploy its Kosmos “AI scientist” across Pfizer’s R&D pipeline, including oncology, internal medicine and immunology (Edison). Kosmos combines public data with Pfizer’s proprietary data to inform research decisions. Edison’s own scientists and engineers will work inside Pfizer research teams. Terms were not disclosed (Business Wire). What to watch: Who owns models trained on sponsor data when the contract ends, and how agent-generated hypotheses get logged. AstraZeneca tests AI on preclinical toxicologic pathology with Tremont AI ↗Foundation models are reaching GLP safety data and some of R&D’s heaviest image files. On October 27, Tremont AI announced a collaboration with AstraZeneca. The Mass General Brigham spinout will test AI that reads whole-slide images from drug safety studies (Business Wire). Today pathologists review these slides by hand, and results vary between reviewers (Drug Discovery World). What to watch: Whether slide scanning, image storage and GLP archives are ready to feed these models. ⚖ Regulatory & PolicyFDA makes remote assessments a formal tool for evaluating drug sites ↗What a site can produce electronically, and how fast, now counts toward its evaluation. On October 29, FDA posted an updated concept of operations for drug facility evaluation between CDER and its Office of Inspections and Investigations (FDA). Remote regulatory assessments and inspection reports from trusted foreign regulators are now formal evaluation tools (PharmTech). The same week, FDA posted warning letters to two overseas firms based only on records they submitted, with no site visit (FDA). What to watch: Whether your sites and CDMOs can produce controlled, audit-ready record extracts on short notice. FDA drafts guidance that would make eSTAR mandatory for device IDEs ↗Structured electronic submission is moving to the start of the device lifecycle. On October 27, FDA’s device center published draft guidance on an electronic submission template for Investigational Device Exemption applications (FDA). Per RAPS, the final version would require eSTAR after a transition of at least one year, and FDA does not plan to grant waivers. Comments are due December 28. What to watch: Whether regulatory information and publishing systems can produce structured IDE content before the clock starts. Sen. Cassidy floats rolling FDA review of AI software changes ↗A discussion draft could change how AI device makers ship model updates. On October 29, Senate HELP Committee chair Bill Cassidy released a discussion draft of the PATIENT AI Act (Senate HELP). It would clarify when AI tools fall under FDA and allow lighter oversight for lower-risk tools. It would also let FDA “review changes to software on a rolling basis” (RAPS). It is not yet a bill. Comments are due December 1. What to watch: Whether AI device makers weigh in by December 1. FDA would let one license cover many patient-specific gene-editing variants ↗Individualized therapies at scale will need systems that track every variant. On October 27, FDA issued draft guidance on individualized genome-editing products (RAPS). New patient-specific variants that fall inside predefined limits could be added to an approved license with lab data and no new trials. PharmTech notes this shifts much of the evidence work to manufacturing and quality teams. What to watch: How cell and gene therapy sponsors’ quality, LIMS and regulatory systems track each variant. 🔒 Cybersecurity & RiskBoston Scientific and Stryker put a price on their cyberattacks ↗Investors now measure a cyber outage in basis points. On October 28, Boston Scientific said its August cyber incident stopped manufacturing and shipments for eight business days (MPO). It cut third-quarter growth by about 300 basis points, roughly $150 million in sales. The company also cut its full-year outlook for a third time (MedTech Dive). It expects more drag in the fourth quarter. On October 29, Stryker lowered its 2026 outlook. CEO Kevin Lobo said “the cyber incident and SAP implementation have impacted our commercial execution” (transcript). He said a July SAP cutover “really delayed a lot of order entry,” and the company only gained visibility to the quarter around September. What to watch: Board questions about the cost of a one-week order-to-cash outage, and ERP cutover timing in high-risk years. Critical PHP and Node.js flaws are now exploited, and Philips lists affected products ↗Web runtimes buried inside instruments and vendor appliances need checking. CISA added the PHP2Shell flaws to its known-exploited list on October 26 and the Node.js Node2Shell flaw on October 27 (Tenable). Older PHP versions get no fix. Philips listed affected imaging and clinical software, including IntelliSpace Portal and Advanced Visualization Workspace (Philips). PHP and Node.js run inside many lab instrument interfaces, LIMS front ends and validated appliances that IT does not patch directly. What to watch: Written exposure statements from instrument and GxP software vendors, and documented compensating controls where no fix exists. HAPI FHIR flaws could let one tenant read another’s records ↗Multi-tenant trial and real-world data platforms are the most exposed. Four flaws in HAPI FHIR, a widely used open-source server for health data exchange, were disclosed on October 22 and detailed by HIPAA Journal on October 27. The worst lets a user cleared for one tenant read or create records in another. Three of the four had no fix as of October 22 (oss-security). Smile Digital Health says one will be fixed in its November release. No exploitation has been reported. What to watch: Whether your decentralized trial, patient app or real-world data vendors run HAPI FHIR, and how they separate tenants. iRhythm breach reaches 8.2 million people on the federal portal ↗A wearable device maker’s patient data leaked through third-party business apps. Around October 29, the federal breach portal listed 8,179,772 people affected by the June breach at iRhythm, maker of the Zio heart monitor (SecurityWeek). Attackers used social engineering to reach third-party business applications. Device serial numbers and insurance details were among the data taken. What to watch: Which CRM, support and billing apps hold patient or device data, and how help-desk resets are controlled. 🎯 Leadership & Operating ModelBMS writes AI targets into its operating plans ↗AI goals now sit inside the plan the CFO tracks, with ROI required for targeted bets. On its October 29 earnings call, BMS CEO Chris Boerner said the company sets “very specific targets that ultimately can be incorporated into operational plans” (transcript). He restated two: 20% to 30% faster drug testing timelines, and half the time from target selection to lead molecule (Benzinga). Broad AI access continues, while targeted investments need an ROI BMS can articulate. What to watch: Whether BMS reports progress against those targets, and who owns the data that proves them. AstraZeneca expects AI to let it grow without adding as many people ↗A $1.5 billion reshaping program comes with an explicit AI workforce message. On October 30, AstraZeneca launched a three-year program targeting $1.5 billion in annual savings, mostly reinvested in its pipeline (AstraZeneca). On the call, CEO Pascal Soriot said “you grow without adding as many people as you would have if you didn’t have the use of AI” (transcript). Fierce Pharma reports the company promised no net reduction in jobs. What to watch: How IT and shared services are measured when growth must come without matching headcount. Avantor names Mikey Santiago, a four-time CIO, as its CIO ↗A key biopharma lab supplier hires a Danaher-trained technology leader. On October 26, Avantor named Mikey Santiago chief information officer, leading global IT strategy and digital transformation (Avantor). He held senior technology roles at Danaher and its operating companies. What to watch: Changes to Avantor’s ordering and data integration roadmap that feed customers’ procurement and GMP systems. 💬 The Bottom Line — Joe’s TakeHere’s a pattern I’ve seen many times. The CFO and the finance team come up with a savings number or a cost number for technology, often with less input from IT than they could have. That’s understandable. Part of it is the silo between the two functions, and part of it is that finance sees across the whole company in a way IT usually doesn’t. In all fairness, the reverse is just as true. Finance rarely sees how systems actually move orders, shipments and batches. On Stryker’s earnings call this week, the CEO said its SAP cutover “really delayed a lot of order entry,” and that the company didn’t get a clear view of the quarter until September. And let’s not forget, the cost of that gap can show up later. A savings figure IT didn’t help shape can turn into a target IT is expected to hit. A cost-of-outage estimate built without IT’s data may leave resilience underfunded. So here’s my advice. Don’t wait to be asked. If you haven’t already, work with your finance partner to put two numbers in place before anyone needs them: what an hour of downtime costs for each of your core business processes, and what your AI work costs to run and what it saves. Even if you have, it’s worth checking that those numbers cover all your current core processes and still hold up. Then work through the savings claims together and agree on which ones will be used, so finance and IT can both stand behind them. The next number about your systems is going to be written by someone. Let’s make sure you will have helped to write it. Ready to move beyond the digest? The LS CIO Community is where these conversations continue. Subscriber sources (often licensed inside life sciences companies): Endpoints News, Anthropic buys Mithril (Oct 27); Fortune, Waters AI tool (Oct 28); Bloomberg, Boston Scientific (Oct 28); Bloomberg, Stryker (Oct 29); Endpoints News, AstraZeneca reorg (Oct 30). How this is made: each edition is researched two ways in parallel, once with Perplexity and once with Claude, then reconciled into a single verified brief before Joe adds his take. This digest is an interpretive summary of publicly available information and does not constitute legal, regulatory, cybersecurity, or investment advice. Until next week, Founder, Leadership Inklings |


