In THIS ISSUE
Signal of the Week, OpenEvidence takes $250M and plans its own cancer drugs.
Framework of the Week, The Referee Test: four checks when the evidence layer has a pipeline.
5 Things Worth Reading, pivots, coding bills, and rules.
Tool of the Week, Litmaps.
Reading time: ~8 min read
🎯 Signal of the Week
The AI that answers your doctors' questions now plans its own cancer drugs: the referee is becoming a player.
Welcome to Issue 10,
What happens when the tool a doctor consults at the bedside also has a drug to sell? We are about to find out. OpenEvidence, the clinical search engine that says roughly 40% of US physicians use it, reportedly raised $250 million at a $15 billion valuation last week, up from $12 billion in January, with Andreessen Horowitz among the backers. Valuation figures vary a little between reports, so treat the number as approximate. The part that matters is not the money. It is what CEO Daniel Nadler said it is for: the company will start developing its own oncology therapies, with a first candidate entering clinical trials before the end of the year and three more to follow, beginning with rare cancers, aimed at, in his words, "the stuff that they are not doing, and maybe can't do." "They" is large pharma.
The details sharpen the picture. Memorial Sloan Kettering is integrating OpenEvidence into its Epic workflows while OpenEvidence folds in MSK's OncoKB precision oncology database, and the company points to data partnerships with MSK and the National Comprehensive Cancer Network as the base for its drug work. Its stated edge is patient finding: a network of physicians who already ask it questions every day. In the same coverage, Anthropic is named as a partner for clinical decision support in nearly 100 low and middle income countries. The announcement itself was muted, a single sentence, while the pivot underneath it is loud. A company that began as a search box for clinicians is now a distribution channel, a data asset, and soon a competitor in the same therapy areas it answers questions about.
Figure 1. Four layers between the evidence and the prescriber, and where the new conflict sits. Sources: Digital Health Wire; Refresh Miami; AI Weekly (Business Insider).
Here is why this belongs on a Medical Affairs agenda. Last week's issue argued that ambient AI is becoming the channel through which evidence reaches the prescriber. This week the channel acquired a motive. For years the answer layer looked like neutral plumbing: you published, it summarized, the physician decided. That model assumed the plumbing had no stake in the outcome. An engine that sees which questions oncologists ask, which trials they cannot find, and which therapies they hesitate over is also sitting on the best map of unmet need in the field, and it has just announced it will act on that map.
Now the opinionated part. I am not predicting that OpenEvidence will tilt its rankings toward its own molecules. I have no evidence of that, and Issue 01 recommended the tool for good reasons that still stand. My point is structural. Journals require conflict of interest statements, trials are registered, and sponsors disclose. The answer layer has no equivalent, and it is now the most-consulted source in the exam room. Medical Affairs should stop treating these platforms as utilities and start treating them as stakeholders with interests, which means asking the questions you would ask any party that sits between your evidence and the prescriber, and writing the answers down.
🧠 Framework of the Week
The Referee Test: four checks when the evidence layer has a pipeline of its own.
A referee does not need to be a saint. A referee needs three things: no stake in the result, visible calls, and a replay you can check. Sport worked that out long ago, and so did publishing, with its conflict statements and trial registries. Medical evidence reaching prescribers through AI answer engines has none of that scaffolding yet, and this week showed why it is overdue: the engine that answers the question may soon own a dog in the fight.
This builds on the thread from recent issues. We followed the agent into the workflow, then into the system of record, then into a shared rental, and last week into the exam room. Each step moved AI closer to the decision. This week asks the follow-up that every step raises: whose interests does the AI represent when it speaks? OpenEvidence is the sharpest example today, but it will not be the last. Any answer layer with pharma partnerships, licensing revenue, or assets of its own will face the same question, and so will the tools your own teams build.
Read the framework as two pairs. Stake and Wall are questions you put to the vendor, in writing, before you rely on or cooperate with the platform. Label and Replay are tests you run yourself, because you should not take disclosure on trust. Start with Stake. You cannot weigh an answer until you know what the answerer wants.
Figure 2. The Referee Test: four checks around every AI answer an HCP receives. Source: iCHealth Pathway framework (original).
Stake. What does the platform own, hold equity in, or earn from in your therapy area? Ask in writing and map it by indication: pipeline assets, licensing deals, sponsored content, and investors with portfolio interests. A conflict you have named is manageable. A conflict you discover after an HCP quotes an answer is a crisis.
Wall. Is there a documented separation between the team that ranks and summarizes evidence and the team that develops or sells products? Ask who can change ranking logic, whether changes are logged, and whether pipeline staff can see the questions clinicians type, which reveal unmet need. A wall nobody can describe is not a wall.
Label. When an answer touches an asset the platform owns or a sponsor has paid to feature, is that marked inside the answer, where the clinician reads it? Test it by asking about indications where the conflict exists. Journals disclose; the exam room should too.
Replay. Can you re-run the same questions next month and compare? Fix a panel of 20 to 30 real questions in your therapy area, log which sources and products appear, and repeat after any vendor announcement. Route misstatements through medical information, and through pharmacovigilance where safety is involved. Drift is only visible if you kept the baseline.
The first two checks decide whether you know what the platform wants. The last two decide whether you can see it in the answers. Most teams will jump to Replay, because testing feels like analysis. Start with Stake instead. And do not assume habit will substitute for discipline: in one survey, 55% of Medical Affairs professionals already use AI daily, yet only structured training moved a team from 29% to 67% daily use and cut KOL research time to 24 to 34 minutes.
Figure 3. Active is not ready: AI use in Medical Affairs before and after structured training. Sources: MSL Mastery, Pellett and Snyder, 10 Aug 2026.
When the answer engine has a pipeline, ask who profits, who can touch the ranking, what is disclosed, and whether you can replay it.
📚 5 things worth to read
Five paywall-free reads, each summarized so you get the value without the click
01
OpenEvidence's Quiet Raise and Loud PivotThe sharpest read on this week's signal, and it is blunt about the contrast in its title. OpenEvidence raised $250 million at a $15 billion valuation, after a $12 billion round in January and July reports of a larger raise at a higher number. The loud part is the plan to develop its own oncology therapies, with a first drug entering trials before year end and three more the following year, starting with rare cancers. It also covers Memorial Sloan Kettering embedding OpenEvidence in Epic alongside OncoKB, and names Wolters Kluwer, Elsevier, Abridge, and Doximity as the competitive pressure. No investor names or usage figures were disclosed. The line to keep: a search engine for evidence that also develops drugs is a referee with a roster.
02
Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care
A reminder that incentives travel with the tool. A Blue Cross Blue Shield Association claims analysis from 2023 to 2025 found that the share of medically complex inpatient cases rose from 37% to 40%, adding $942 million in costs to member plans, $653 million of it from secondary diagnoses lifting cases into higher severity tiers. Hospitals with the highest complexity coding showed similar or lower ICU use and length of stay than peers. A June survey cited in the piece found more than 63% of healthcare organizations use AI in revenue cycle work. The payer association has its own interest, and hospitals call AI a response to denials, so read both sides. The line to keep: whoever operates the AI shapes what the record says.
03
Medical Affairs Is More AI-Active Than AI-ReadyThe most useful practitioner data point this month, and the source of Figure 3. Drawing on 53 or more training sessions with 265 professionals across 10 companies, the authors report that 55% of 336 respondents use AI daily. One team lifted daily use from 29% to 67% in three months, ideation use from 41% to 83%, and feedback incorporation from 32% to 67%, while KOL research time fell from 61 to 137 minutes to 24 to 34 minutes. The key distinction is activity versus readiness: five skills, from prompting to consistent team practice. The data come from the authors' own training programs, so read it as directional. The line to keep: the advantage comes from making good performance repeatable.
04
Regulating AI: Inside the FDA and EMA's road to AI governance
A dated timeline that makes the regulatory picture legible in one sitting. It runs from the FDA's January 2025 draft guidance on AI for regulatory decision-making, through the joint FDA and EMA principles of January 2026, to the EU AI Act's enforcement regime this August and the high-risk deadlines of December 2027 and August 2028. The central insight is that regulators are not treating AI as one product category: sponsors must show credibility for a specific question in a specific context of use. That logic transfers to this week's framework: an AI answer's credibility depends on who built it and why. The line to keep: context of use is the unit of regulation, so it should be the unit of your due diligence.
05
Novo looks to Anthropic's AI models to 'supercharge' drug development
The other side of the same market. On 16 September, Novo Nordisk announced a collaboration with Anthropic, starting with a pilot of Claude Science on R&D workflows where the companies expect the greatest impact, plus frontier models for software development and wider AI adoption. It follows Novo's April partnership with OpenAI, with Bristol Myers Squibb, Eli Lilly, and Roche making similar announcements in what it calls a war of superlatives. Financial terms were not disclosed, and Novo stresses data governance and human oversight. Read it next to the OpenEvidence story: the same model makers are now suppliers to pharma and to the answer layer that pharma depends on. The line to keep: map every supplier relationship in your evidence chain, not only your own.
🔧 TOOL OF THE WEEK Litmaps
Litmaps
What it is: a literature-mapping tool that starts from a few seed papers and draws a visual citation map of the work around them. Its Discover function suggests related papers you have not found, and its Monitor function sends alerts when new papers appear on your topic. The company says more than 350,000 researchers in 150 countries use it. Where keyword searches return a list, Litmaps shows the neighborhood, so you can see which papers anchor a field, which bridge two communities, and which recent work is gaining citations.
How to use it for Medical Affairs and MSLs: seed a map with your pivotal trial and two competitor trials, and look at which papers cluster around each and who is citing whom, a fast way to prepare for a scientific exchange or a data-gap discussion; set a Monitor alert on the seed map for your therapy area, so new papers arrive as a weekly feed for insight meetings; and, as a rough companion to this week's Replay check, seed a map with the papers your label rests on and compare them with the sources an AI answer engine actually cites for the same question.
Where to access it: litmaps.com on the web. The Free plan covers basic search with up to 20 inputs, 100 saved articles, one Litmap, and a monthly alert summary. Pro is $10 per month or $120 per year with unlimited inputs, articles, and maps and configurable alerts. Team pricing is custom, with team-wide collaboration. Confirm current limits on the site before you buy.
COMPLIANCE NOTE
Litmaps is a discovery aid, not a systematic search method and not a source of MLR-approved content. A citation map reflects the coverage of its index and the lag in citations, so very recent papers will look poorly connected and relevant work can be missing. Read every paper in full before you rely on it, and route anything intended for an HCP communication through your normal MLR review. Do not upload unpublished manuscripts, confidential data, or patient-identifiable information, and check your company's policy on third-party tools before using a work account.
That is Issue 10. 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 10, Monday, October 5th, 2026.
Parts of this issue were produced with AI assistance and checked before publication.