In THIS ISSUE
Signal of the Week, OpenAI buys the workflow layer of drug development.
Framework of the Week, The Landlord Test: four checks before a model-maker owns your workflow.
5 Things Worth Reading, deals, regulators, and Medical Affairs strategy.
Tool of the Week, Google Scholar Labs.
Reading time: ~8 min read
🎯 Signal of the Week
OpenAI's first pharma acquisition buys the operating layer, not another model.
Welcome to Issue 12,
Trial scheduling is not a keynote topic, and that is exactly what OpenAI just bought. On October 20, OpenAI announced it has acquired Orchestra Bio, a roughly four-year-old platform that coordinates drug development across discovery, preclinical, manufacturing, and clinical stages. Third-party reporting says its agents help with study monitoring, financial tracking, and trial scheduling. Orchestra's team and technology will be combined with Rosalind, OpenAI's life-sciences model and tools. Financial terms were not disclosed. TechCrunch reported that Orchestra had raised about $20 million from backers including Andreessen Horowitz, Atria, and 8VC, a figure OpenAI did not confirm.
It is OpenAI's first known acquisition in pharma and life sciences, and it follows Anthropic, which bought Coefficient Bio in April and launched its Claude Science platform in June. Three days after the OpenAI announcement, STAT reported that Bristol Myers Squibb is leading a $237 million Series E in Pathos, an AI drug developer valued at about $3 billion, whose Foundry model runs thousands of agents that estimate the probability of success for clinical assets. Read the caveats too: OpenAI's announcement names no customers, gives no integration timeline, and does not say whether Orchestra stays a separate product. Data handling, agent reliability, and customer retention are all open questions.
Figure 1. The model-makers moved into operations: four moves between April and October 2026. Sources: eMarketer; Tech Insider; STAT.
Here is why this belongs on a Medical Affairs agenda. Scheduling, monitoring, and tracking sound like clinical operations, but they are where evidence gets created and recorded. Whoever runs the agents that watch a study also holds the trail of what was seen, when, and by whom. Your function lives on that trail: it is what you will be asked to explain to HCPs, payers, and regulators later. When the model vendor moves from supplying answers to running the workflow, the question changes from whether the model is good to who holds the record.
Now the opinionated part. A model you rent is easy to swap. A workflow you rent is not, because your history, your habits, and your audit trail live inside it. That is not automatically bad, and several of the best tools will come from exactly these companies. But treat every operating-layer deal as a landlord relationship and read the lease. Last week's issue asked what a headline deal is actually worth. This week's asks what happens when the vendor owns the building.
🧠 Framework of the Week
The Landlord Test: four checks before a model-maker owns your workflow.
Renting a brain is easy to reason about. You pick a model, you test it, you swap it when a better one arrives. Renting the building is different, because the building keeps your records.
That is the shift behind this week's signal. In recent issues we looked at the agent in the workflow, the badge office, and the money behind deal headlines. The Landlord Test applies the same discipline to ownership: when a model-maker moves from supplying a model to operating a layer of your process, which parts of your function now sit inside its walls?
Read the figure from the bottom up: start with the lease, then the tools inside, then the agents running the work, then the evidence record on top. Checks 1 and 2 are for procurement and IT. Checks 3 and 4 are where Medical Affairs has to be in the room, because you know which records the field and the regulators will ask for. The same test works on any vendor, not only the big model-makers.
Figure 2. The Landlord Test: four layers a model-maker can own, from the lease to the evidence record. Source: iCHealth Pathway framework (original).
Contract and exit. Can you leave with your data, prompts, and history intact, in a usable format? Ask for export rights, a deletion clause, and a notice period before signing, not at renewal. If the exit costs more than the subscription, you are not a customer, you are a tenant.
Science tools. Which models and tools sit inside the process, and can you swap one without rebuilding the workflow? Orchestra will be combined with Rosalind, so ask whether a team could still run another model inside the same agents. Prefer vendors that let you bring your own.
Workflow agents. Who runs scheduling, monitoring, and tracking, and can a human see what each agent did and why? Request activity logs you can read without the vendor's help. An agent you cannot audit is a risk in any GxP-adjacent process.
Evidence record. Who holds the audit trail and source data after the deal, and after the vendor's next acquisition? Write down which records must stay under your control, who can access them, and what happens if the vendor is bought. This is the check only Medical Affairs will think to ask.
The first two checks tell you whether you can leave. The last two tell you whether you can explain what happened while you were there. Most teams negotiate the first and forget the fourth. Start from the record, because that is what you will be asked to produce.
Figure 3. Why the operating layer is worth buying: the trial failure rate and timeline OpenAI cites, set beside the money moving in this week's deals. Sources: eMarketer; Tech Insider; STAT.
A model is a supplier. A workflow is a landlord. Read the lease before you move in.
📚 5 things worth to read
Five paywall-free reads, each summarized so you get the value without the click
01
OpenAI Expands Its Pharma Ambitions With Drug Development AcquisitionThe most compact account of the signal. It frames the Orchestra deal as OpenAI's first known acquisition in pharma, set against earlier partnerships with Novo Nordisk, Sanofi, and Moderna, and notes that Rosalind will work alongside Orchestra's agents on tasks such as trial scheduling and monitoring complex lab studies. It repeats OpenAI's numbers: 10 to 15 years from target discovery to US approval, and a trial failure rate of about 90%. The useful caution comes at the end: AI's impact on clinical outcomes and development costs remains limited so far, and the pressure to show returns will grow. Terms were not disclosed. The line to keep: predictions that AI will cut discovery from years to months are ambitious, so ask for the returns before the roadmap.
02
Microsoft's Mustafa Suleyman Says Its AI Outperforms Doctors on Hard Cases
Microsoft's MAI-DxO coordinates several AI agents that question a patient, order tests, and debate a diagnosis, with results published in Science. On complex New England Journal of Medicine cases it was correct 83% of the time, against 36% for primary care doctors and general internists working alone with reference materials. Read the fine print the article includes: the cases were simulated, specialists and physician teams were not the comparison, and most of the gain came from the underlying model, with the orchestration adding on top. It showed little or no improvement on emergency cases, and a test with consenting patients at Beth Israel Deaconess is planned. The line to keep: the wrapper helps, but the model sets the ceiling.
03
From Vision to Velocity: Architecting a Functional AI Strategy for Medical AffairsA short roundtable summary by Laura Orengo of Digital Science, published on October 23, with no statistics. Its argument is that Medical Affairs teams move faster through deliberate, incremental progress than through one large leap: define clear use cases and align stakeholders before scaling, keep humans in the loop and insist on provenance for AI outputs, and invest in change management so adoption becomes routine. It lists six discussion topics, among them trust and adoption, ROI, governance, and systematic literature review. It is vendor-published, with no methodology or participant count, so treat it as a list of questions, not evidence. The line to keep: traceable sources are the price of admission, which is the fourth check of this week's framework in one phrase.
04
FDA AI Credibility Guidance Still Draft: What to Do Now
A consultant's status check, dated September 5, on the FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs and biologics, which still reads as a draft not for implementation. It walks through the seven-step credibility framework, from defining the question of interest and the context of use to rating model influence and decision consequence and judging whether credibility is adequate. It expects the core structure to survive into the final version and advises building the records now, without tying SOPs to draft line numbers. The author's firm sells AI governance services, and the predictions are opinions, not FDA statements. The line to keep: your context of use is a document you can write today, whatever the final text says.
05
Use of Artificial Intelligence (AI) in the Medicinal Product Lifecycle
The EMA's page for its reflection paper on AI across the medicines lifecycle, from drug discovery through post-authorisation, covering human and veterinary medicines. It records the path: a draft and public consultation in 2023 (July 19 to December 31), a workshop in November 2023, and the final paper published on September 30, 2024, together with a document showing how comments were addressed. It is a reflection paper, not a binding guideline, and the page is only a summary, so the detail sits in the PDF. It is older than the rest of this issue, but it remains the reference point for how European regulators think about AI in your products. The line to keep: the paper is principle-based, so turning principles into records is your job.
🔧 TOOL OF THE WEEK
Google Scholar Labs
What it is: an experimental, AI-powered mode of Google Scholar, announced on November 18, 2025, for research questions that need a topic examined from several angles. It breaks your question into key topics and the relationships between them, searches Google Scholar for each, reviews the results, and explains how each paper addresses your question. An early independent review calls it a deep search tool, not deep research: it screens papers one by one and does not synthesise across them, with a limit of about 50 relevant papers per search. Its strengths are a very large index and few false positives. It can be strict about query types and is sensitive to exact wording.
How to use it for Medical Affairs and MSLs: when a deal or guidance lands, such as this week's Orchestra news, ask a full-sentence question like what evidence exists on AI agents for clinical trial monitoring and data quality, then read the per-paper explanations to see which papers actually answer it; use it to find the paper behind a concept you half remember, which the early review names as one of its best uses; and treat it as a first pass before your regular multi-database search, because the same review warns it is not enough alone for systematic evidence synthesis and that missed papers are harder to spot when precision is high.
Where to access it: scholar.google.com, signed in with a Google account. Google said at launch that it was available to a limited number of logged-in users, and early testers reported a waitlist, so check whether it is enabled for your account. Features may have changed since the launch coverage, so confirm what is available before you build a process around it.
COMPLIANCE NOTE
Scholar Labs summaries are generated explanations, so verify every claim against the paper itself before quoting it, and never present a generated explanation as MLR-approved content or as a systematic review. Because the tool does not show how many results it evaluated, document your search strategy and query wording if the output informs a deliverable. Use a Google account only if your company allows it, and do not type confidential pipeline plans, unpublished data, or patient-identifiable information into the search box. Check your company's policy on third-party AI tools before using it on a work device.
That is Issue 13. 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]
Issue 13, Monday, October 16th, 2026.
Parts of this issue were produced with AI assistance and checked before publication.