|
Who Owns the Model Trained on Your Data? Recursion just cut its patient-data bill in half, paid part in stock, and licensed its own AI model to the same partner. Merck, UCB and Roche signed AI partners this week too. The terms are worth a closer look. Life Sciences CIO Weekly • Coverage: November 2 to November 8, 2026 For years the question in an AI partnership was what the model could do. This week the public terms shifted to what each side gets to keep.
The data and model terms in these deals are often settled by business development and legal. Below are the details, and where IT may want a seat at that table. 🔍 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 LeadRecursion halves its Tempus data bill and licenses Tempus its own AI model ↗The price of patient data for AI training is being renegotiated, and models are now part of the trade. Tempus AI reported third-quarter revenue of $411.5 million on November 2, up 23%. Its data and applications business grew 33% to $108.1 million, and 2026 data-licensing bookings now top $550 million (Tempus). Tempus also sells “AI modeling,” where customers build models on its data using dedicated GPU capacity. Recursion’s quarterly filing shows the buyer’s side (10-Q). In September it amended its 2023 Tempus data contract. Two remaining payments of $42 million became three payments of $14 million, partly payable in Recursion stock. Recursion added a year to the term and gave up its right to exit for convenience. It also granted Tempus a two-year, non-exclusive license to its TxFM RNA foundation model, for which Tempus will pay $12 million. What to watch: Check whether your real-world data contracts allow an exit for convenience, and who owns any model trained on licensed data when the contract ends. Merck, UCB and Roche each add an AI design partner on different terms ↗Big pharma now runs several outside design models at once, and some of them learn from the sponsor’s own data. On November 2, Variational AI said it will design small molecules for Merck against two undisclosed targets. It will use a version of its Enki platform fine-tuned on Merck’s proprietary data, in a deal worth up to $349 million with milestones (Variational AI). On November 4, UCB said it will use Cradle’s protein design platform in selected antibody projects. UCB’s lab results feed back into program-specific models in a “lab-in-the-loop” cycle, and UCB keeps ownership of the proteins it designs (UCB). On November 5, Chai Discovery said Roche will use its models in “a number of drug discovery programs” (Chai Discovery). UCB Chief Scientific Officer Alistair Henry called AI “a scientific force multiplier.” For IT, lab-in-the-loop models only improve if assay results are captured, structured and returned fast. That puts ELN, LIMS and data pipelines on the critical path (BioPharm International). What to watch: Where a fine-tuned model runs, who owns its weights, and whether what it learns can reach the vendor’s other clients are becoming standard contract questions. 🤖 AI & DataVeeva will charge separately for AI: 10% on top of Vault, and agents licensed on their own ↗AI from your core regulated-content vendor is about to show up as a new line on the renewal. At its November 5 investor day, Veeva said Vault AI, an assistant inside Vault applications, will cost 10% of the Vault subscription. Use is unlimited. Falcon agents, which do work for customers, carry a separate annual license per agent. CEO Peter Gassner gave a top-10 biopharma clinical intake agent as a seven-figure example, versus about $100,000 for a small biotech. He said a few customers are live and “we have to make these things work first” (transcript). Joanne Byrne, Roche’s head of global regulatory operations, said Roche joined early partly to test whether Veeva could bring AI into the Vault systems that hold much of its data. Veeva raised its market estimate to $30 billion and kept its $6 billion revenue target for 2030. What to watch: Ask for usage and productivity measures before accepting a 10% uplift, and plan agent licenses into 2027 budgets. Industry, academics and Anthropic split on how FDA should regulate generative AI devices ↗How FDA handles model updates will decide how much validation work every new LLM version triggers. FDA received more than 500 comments on its August discussion paper on generative AI medical devices. RAPS and MedTech Dive analyzed them on November 3 (RAPS, MedTech Dive). AdvaMed said existing frameworks “can be adapted.” It called predetermined change control plans “too restrictive.” Anthropic asked FDA to keep benchmarks for an authorization, or for a move to a new model version, independent of both the device maker and the model developer. What to watch: If FDA adopts independent re-benchmarking for model-version changes, teams embedding third-party LLMs in regulated software will need version pinning and change control for every update. ⚖ Regulatory & PolicyFDA’s surprise inspections in India and China jump, and inspection records go public ↗Overseas suppliers now need to be inspection-ready on any given day, and their findings will be easy to look up. Unannounced visits made up 87% of FDA drug inspections in India in fiscal 2026, up from 47% a year earlier. In China the share rose from 19% to 61%, said FDA’s Elizabeth Miller (RAPS). Acting chief of staff Lowell Zeta said FDA now posts redacted Form 483s and inspection reports online (RAPS). This follows last week’s formal shift toward remote records requests. What to watch: An unannounced visit leaves no time to tidy audit trails or access reviews. Quality agreements with CDMOs and API makers should require prompt sharing of any 483. FDA asks how to judge autonomous and remote-controlled robotic devices, cyber resilience included ↗For connected robots, network reliability and cyber resilience are now part of the benefit-risk question. On November 6, FDA’s device center published a discussion paper on robotic devices that act with autonomy or are operated remotely. It treats them as “cyber devices” and names network latency, loss of connectivity and AI behavior outside training data as core risks. One of its 15 questions asks what evidence shows resilience to connectivity loss or cyber breaches (discussion paper). Comments are due January 5, with a public workshop December 2 to 3. What to watch: Medtech product security and IT leads can shape these expectations now, before they reach guidance and premarket review. EMA updates its notice on validating computerised systems in clinical trials ↗EU expectations for eClinical systems now point formally to the risk-based thinking in ICH E6(R3). EMA revised its notice to sponsors on validating and qualifying computerised systems used in clinical trials. The revision is dated October 29 and was reported November 3 (LexisNexis). It aligns the text with ICH E6(R3): systems should be fit for purpose, with validation proportionate to risk. EMA calls the changes administrative. Sponsors stay responsible for validation even when they rely on vendor documents, and inspectors must be able to see validation records (EMA). What to watch: Confirm vendor qualification files for EDC, eCOA, IRT and eTMF are ready for an inspector to see. MHRA opens a careful route for trials built on non-clinical data from outside the OECD system ↗UK trial applications may now rest on toxicology from non-OECD labs, but the sponsor carries the data-integrity burden. On November 4, the MHRA said it may accept non-clinical safety data generated outside the OECD Mutual Acceptance of Data system, case by case. Examples include first-in-human trials in advanced cancers with high unmet need (MHRA). Standards do not change, and the agency will inspect test facilities that appear often in applications (Clinical Trials Arena). What to watch: Contracts with toxicology CROs outside the OECD system should guarantee raw-data access and audit trails. 🔒 Cybersecurity & RiskFive US agencies warn of attacks on cold-chain controllers after French plasma storage is hit ↗Attackers are changing temperature settings on refrigeration systems, and medicine storage is already in the blast radius. On November 5, CISA, the FBI, HHS, USDA and DOT warned that attackers are targeting internet-facing refrigeration and temperature controllers. They can force heating or defrost cycles and hide real temperatures from operators (joint advisory). France’s blood service, EFS, said its refrigeration vendor was compromised and settings were changed at three sites storing blood products, including plasma for medicine manufacturing. Monitoring caught it in time (SecurityWeek). What to watch: Biorepositories, sample freezers, cell and gene therapy storage and 3PL warehouses use the same kinds of controllers. Ask facilities and logistics partners which ones are reachable from the internet. Myriad Genetics shuts down testing and order systems to contain an intrusion ↗Containing an attack can mean stopping the lab, so the decision to pull the plug needs an owner before it happens. Myriad detected unauthorized access on November 2 and temporarily suspended its testing operations and order management systems while it reset credentials. It is still assessing whether patient records were accessed, and most systems are back (8-K filing). Myriad filed under Item 8.01 rather than the SEC’s material-incident item, saying it does not expect a material impact. Its tests also support companion diagnostic work with pharma partners. What to watch: Settle in advance who can take systems offline, how specimens and orders are handled while they are down, and how legal decides the disclosure item. Attackers exploit a critical Node-RED flaw in Siemens industrial gateways ↗A tool often used to wire up lab and plant data is being hijacked with a single web request. VulnCheck reported in-the-wild attacks on November 3 against CVE-2026-58115, a maximum-severity missing-authentication flaw in Node-RED on Siemens SIMATIC IOT2050 devices (SecurityWeek). Siemens issued a fix in August (Siemens advisory). Attackers are also scanning for any Node-RED instance left without a password. What to watch: Node-RED often runs in lab automation and bioreactor data pilots outside IT change control. Find every instance and lock it down. Three charged in a $129 million scheme that used AI to impersonate CEOs and lawyers ↗Deepfaked executives on a call are now a documented route to a fraudulent wire. Federal prosecutors in New York charged three men on November 5 with stealing about $129 million from 2024 to 2026 (The Record). It started with an email from the “CEO” about a confidential deal. Calls followed, using AI to mimic the executive’s voice, and then fake outside lawyers read from scripts (SecurityWeek). No life sciences victims were named. What to watch: Licensing and milestone payments suit the “confidential deal” story. Require a callback on a known number before any wire an executive or outside counsel requests. 🎯 Leadership & Operating ModelPfizer opens a new $2.5 billion savings phase and names AI in manufacturing, R&D and drug safety ↗A top-five pharma has written AI into a multiyear cost commitment that peers’ boards will notice. On November 3, Pfizer said it will deliver nearly all of its $7.2 billion savings program by the end of 2026. A next phase targets $2.5 billion in net savings from 2027 through 2029 (prepared remarks). CEO Albert Bourla said AI is improving manufacturing efficiency and speeding target validation. He described GENESIS, a new AI pharmacovigilance platform used in more than 120 countries. It “takes on the administrative burden so our safety scientists can spend their time where it matters most” (transcript). What to watch: Expect questions about which savings come from technology, and about the validation and human review behind AI in safety work. Charles River ties $300 million in savings to digitized labs and launches a new sponsor data channel ↗How preclinical CRO data reaches sponsors is changing. Charles River said on November 4 that its “Create the Future” program should save over $300 million over four years (transcript). It partnered with Autolomous to launch Apollo Orchestrate, which digitizes the biologics lab workflow from sample submission to results delivery (release). Asked about AI demand, CEO Birgit Girshick said it makes little sense to “split out AI versus non-AI.” What to watch: Check how Orchestrate fits your LIMS and data pipelines before it becomes the default way CRO results arrive. Zimmer Biomet opens an India capability center and aims AI at shared services ↗A medtech company has put a global capability center and AI process work into one cost story. On November 4, Zimmer Biomet’s interim CFO Paul Stellato said the company “recently opened up a global capability center in India.” He added that “technology will be a big part of it.” CEO Ivan Tornos said it is working with outside AI partners to rework shared services, demand planning and customer service (transcript). What to watch: Expect capability center and AI process programs to be scoped together at other device makers. Bayer credits its operating model and AI with getting a drug to patients a year early ↗Bayer tied AI and a leaner way of working to a specific pipeline gain, though without detail. Speaking on November 3 about asundexian, approved in China, CEO Bill Anderson said Bayer’s Wuppertal scientists “embraced into our operating model and AI, which allowed them to substantially compress the timeline, getting this medicine to patients a year early” (Bayer). Bayer raised its pharma margin outlook to 25% to 27% (Q3 release). It has not said which steps AI shortened. What to watch: Whether Bayer explains what was compressed, and puts a number on AI productivity in its 2027 guidance. On Q3 calls, Amgen talks AI, Teva says “not AI,” and Novo stays quiet ↗AI ambition is uneven across the sector, and so is the mandate it gives IT. Amgen CEO Robert Bradway said Amgen is “leveraging technology, data-driven insights and artificial intelligence” across discovery, development, manufacturing and delivery (transcript). Teva’s R&D head Eric Hughes described his approach as “shoe leather, not AI” (transcript). Novo Nordisk’s call did not mention AI (Q3 report). What to watch: Where the CEO ties AI to margin and speed, IT gets both a mandate and a target. Where leadership credits execution, AI spending will be judged project by project. 💬 The Bottom Line — Joe’s TakeSeveral stories this week point to the same need: people close to the business who truly understand how AI works. When a vendor fine-tunes its model on Merck’s chemistry, or Recursion hands Tempus a license to its own model, a company can give away more than data. It can give away the know-how behind how it trains its own models, along with trade secrets and security exposure. The company as a whole needs to understand what is really on the table, including whether what’s being promised matches what the technology underneath actually does. The CIO doesn’t need to be the deepest AI expert in the company. The CIO does need to know who is, and bring those people into the room before the terms are set. The good news is that the door is often already open. Many of the CIOs I talk with are being asked to guide their company’s AI efforts, and some are seeing their role expanded to include AI. That invitation is real, but it is usually loosely defined. That’s the opportunity. The hard part is making the time, because the day-to-day pulls hard. So here’s my advice: make sure you have strong leaders underneath you who can run the day-to-day, so you’re free to be that voice in the room. Then earn that seat by being the one who asks the questions no one else thought to ask. — Joe Ready to move beyond the digest? The LS CIO Community is where these conversations continue. Subscriber sources (often licensed inside life sciences companies): Bloomberg, UCB taps Cradle (Nov 3); Citeline Medtech Insight, FDA gen-AI comments (Nov 3); STAT Pharmalot, FDA inspections in India and China (Nov 4). 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 |
