In THIS ISSUE
Signal of the Week, ambient AI moves from pilot to infrastructure in the exam room.
Framework of the Week, The Exam Room Test: four checks when AI sits beside the prescriber.
5 Things Worth Reading, scribes, regulators, and the evidence front door.
Tool of the Week, Undermind.
Reading time: ~8 min read
🎯 Signal of the Week
Ambient AI crossed from pilot to infrastructure this week, and it is quietly becoming the channel through which evidence reaches the prescriber.
Welcome to Issue 09,
Seventy-five medical centers, nine million veterans, and a contract ceiling of $775.72 million. On Tuesday, Abridge announced it had been selected, through a distribution partner, to provide ambient clinical AI under a new enterprise contract at the US Department of Veterans Affairs, the largest integrated health system in the country. The multi-award, five-year IDIQ lets individual VA medical centers and regional networks buy task orders, and Knowtex won a seat too. The same day, Heidi closed $340 million ($100 million in Series C equity plus $240 million in growth financing from General Catalyst's Customer Value Fund) at a $900 million valuation, shortly after Stockholm's Tandem Health raised a $100 million Series B. Ambient AI, the software that listens to a consultation and writes the note, is no longer a pilot. It is procurement.
The details show where this is heading. Abridge runs on both VA record systems, legacy VistA/CPRS and the new Federal EHR, across primary care, twelve specialties, and the virtual Clinical Resource Hubs, in more than 28 languages. Conversations are captured only with the veteran's consent and turned into SOAP notes the clinician must approve before anything is committed. But the note is only the entry ticket. Abridge now pitches linked-evidence support that synthesizes the live encounter, the chart, and peer-reviewed literature inside the workflow, and expects to handle more than 100 million patient-clinician conversations this year. Heidi's Evidence engine has already answered more than 10 million point-of-care queries. The scribe is becoming a clinic operating system, and that operating system is starting to decide which evidence appears while the patient is still in the room.
Figure 1. Ambient AI from pilot to infrastructure: twelve months of procurement, regulation, and capital. Sources: Abridge VA enterprise contract, Nextgov/FCW, 22 Sep 2026; deployment details, HIT Consultant; MHRA, HSSIB, Heidi and Tandem, Nelson Advisors; FDA GenAI discussion paper.
Here is why this belongs on a Medical Affairs agenda and not only on a health-system CIO's. For a decade, the moment of scientific influence was a conversation you could see: the MSL meeting, the congress symposium, the medical information call. That moment is moving upstream, into an AI layer that listens to the consultation, reads the chart, and surfaces a cited answer before anyone from your company is in the picture. IQVIA's survey work puts 54% of HCPs already using generative AI to access scientific information, and 94% of those users say it makes information easier to find. Ambient tools carry the same logic into the exam room itself. Whether your latest data, your label update, or your safety communication is in the corpus those tools read, and how faithfully they summarize it, is now a medical question with patient consequences.
Now the opinionated part. Most Medical Affairs teams will read this as health-tech news and file it. That is a mistake, for two reasons. First, the evidence layer inside ambient platforms is being built right now, with vendors choosing sources, guidelines, and ranking logic, and nobody from your function is in the room. Share of voice is giving way to what IQVIA calls share of algorithmic trust, and you earn that with strong publications, guideline inclusion, and clean, current evidence, not with more detailing. Second, the same technology is coming for your own field teams. Ambient capture of MSL conversations and automatic insight structuring are already on vendor roadmaps, and the questions the VA answered for veterans (explicit consent, clinician review before anything is committed, formal security authorization) will be asked of you. The teams that treat the exam room AI as a stakeholder, and hold their own recordings to the same bar, will set the terms. Everyone else will discover that their evidence was summarized by someone else.
🧠 Framework of the Week
The Exam Room Test: four checks for Medical Affairs once AI sits between the evidence and the prescriber.
Who writes the first draft of a clinical decision? Increasingly, it is software that was listening. The ambient platforms scaling this week start with the note, but their roadmaps run straight through pre-visit briefings, point-of-care evidence retrieval, coding, and order preparation. Each step puts an AI between the published evidence and the prescriber, and each one is a place where your data can be surfaced, compressed, delayed, or distorted.
This picks up a thread from recent issues. We have followed the agent from a bolt-on tool to a native feature, a shared rental, a model maker with its own agenda, and last week a governed platform asking to run all of Medical Affairs. This week the AI shows up on the other side of the table, in the HCP's workflow, where you do not buy it, configure it, or control it. That changes the job: you cannot govern the tool, so you have to govern your evidence and your own behavior around it.
Read the framework as a staircase. The first two steps ask whether your evidence gets into the point-of-care layer at all and stays current there. The last two ask whether it arrives intact, and whether you apply the same standard to the ambient tools your own teams use. Skip a step and the ones above it wobble.
Figure 2. The Exam Room Test: four steps for Medical Affairs when AI sits beside the prescriber. Source: iCHealth Pathway framework (original).
Source. Which corpus does the point-of-care AI actually read: guidelines, PubMed, labels, its own licensed content? Map the ambient and evidence tools your HCPs use in each market, and find out whether your pivotal publications, label, and guideline positions are in scope. If your evidence is not in the corpus, it does not exist at the moment of decision, however strong it is.
Freshness. How fast does a new label, a safety communication, or a data readout reach that layer? A tool that surfaces last year's guideline or a superseded dosing statement is a patient-safety problem, not a marketing one. Treat AI indexing lag as a real delay in your publication and disclosure plan, and check it after every major update.
Fidelity. When the AI summarizes your data, who checks that it is right? Run structured audits of how leading tools answer the questions HCPs actually ask in your therapy area, log misstatements, and route them through medical information and pharmacovigilance where appropriate. The I3LUNG study is the warning: clinicians sometimes accept incorrect AI suggestions, so an error in the summary can become an error in the prescription.
Mirror. Do your own ambient tools meet the bar you expect of theirs? If MSL conversations or advisory boards are recorded and structured by AI, you need explicit consent, human review before anything enters a system of record, clear retention and deletion rules, and adverse-event detection. The VA's approach for veterans is a sensible floor: consent first, clinician sign-off before commit, formal security authorization.
The first two steps decide whether your evidence gets into the room at all. The last two decide whether it arrives as you published it, and whether you have earned the right to ask others for accuracy. Most teams will jump straight to fidelity audits, because errors make headlines. Start with source instead: you cannot correct an answer drawn from a corpus you are not in.
Figure 3. The numbers behind the new front door. Sources: HCP generative AI use, IQVIA (EPG survey); NHS ambient AI evaluation, Nelson Advisors.
The exam room AI is now a stakeholder: earn its trust with evidence, audit what it says, and hold your own recordings to the same bar.
📚 5 things worth to read
Five paywall-free reads, each summarized so you get the value without the click
01
Abridge wins seat on $775.7M VA enterprise contract to power ambient clinical AIThe most detailed account of this week's signal, and worth reading for the fine print rather than the headline number. The $775.72 million is a five-year ceiling shared across all eligible vendors, and individual VA medical centers buy task orders, so this is a licence to compete, not a guaranteed rollout. What makes it matter is the footprint: Abridge already runs in more than 75 VA medical centers on both the legacy VistA/CPRS system and the new Federal EHR, in over 28 languages, with capture only on explicit veteran consent and every note approved by the clinician before it is committed. The piece also spells out the next step, evidence support that combines the encounter, the chart, and peer-reviewed literature at the point of care. The line to keep: the scribe is the door, and the evidence layer behind it is the room Medical Affairs needs to be in.
02
Abridge wins seat on $775.7M VA enterprise contract to power ambient clinical AI
The most detailed account of this week's signal, and worth reading for the fine print rather than the headline number. The $775.72 million is a five-year ceiling shared across all eligible vendors, and individual VA medical centers buy task orders, so this is a licence to compete, not a guaranteed rollout. What makes it matter is the footprint: Abridge already runs in more than 75 VA medical centers on both the legacy VistA/CPRS system and the new Federal EHR, in over 28 languages, with capture only on explicit veteran consent and every note approved by the clinician before it is committed. The piece also spells out the next step, evidence support that combines the encounter, the chart, and peer-reviewed literature at the point of care. The line to keep: the scribe is the door, and the evidence layer behind it is the room Medical Affairs needs to be in.
03
Considerations for the Regulation of Generative AI-Enabled Medical DevicesThe primary source every regulatory and medical governance lead should skim before 19 October. The FDA's Digital Health Center of Excellence released this discussion paper in August and is collecting feedback under docket FDA-2026-N-7874. It covers four areas: how to assess risk for GenAI-enabled devices, what premarket evaluation should look like, how to monitor these products once they are in use, and other open questions. It is explicit that it is neither draft nor final guidance and proposes no policy, but its questions show where CDRH is heading as ambient tools drift from documentation toward decision support. For Medical Affairs, the postmarket monitoring section is the one to read, because that is where hallucination, omission, and drift will be defined. The line to keep: the moment a scribe starts suggesting, it may become a device, and the evidence it cites becomes part of the risk story.
04
Healthcare AI News and Regulation: September 2026 Evidence Briefing
A carefully dated, skeptical roundup that models how to read AI news, which is itself the lesson. It labels each item by evidence type: an FDA final order confirming that specified radiology AI still needs 510(k) clearance, OpenAI's two new routes into healthcare (Epic patient context for organizational deployments, and a public data plugin covering PubMed, DailyMed, and CMS coverage for eligible US clinicians), ARPA-H's $62.7 million ADVOCATE program for agentic cardiovascular AI, and two studies. The I3LUNG study is the one to remember: explainable model support lifted physicians' disease-control prediction accuracy from 57% to 65%, yet clinicians also accepted incorrect AI suggestions. Note that the publisher sells an AI scribe, and says so. The line to keep: a product launch does not establish patient benefit, and a research result does not authorize a deployment.
05
Heidi Health's $340M and Tandem Health's $100M mark a structural inflection point
The European lens on ambient AI, and the best single map of its regulatory split. The analysis tracks Tandem across 14 European markets and more than 130 record-system integrations with three EU MDR Class IIa certifications, while Heidi holds the NHS England Midlands framework covering up to 70,000 clinicians. Two dates matter: on 29 July the UK MHRA clarified that transcribing or summarizing alone does not make a tool a medical device, and on 6 August the Health Services Safety Investigations Body opened a national inquiry into hallucination, omission, and automation bias. An NHS evaluation across more than 17,000 encounters found 23.5% more direct patient interaction time. The author is an investment bank, so read the market claims accordingly. The line to keep: scribes are becoming clinic operating systems, and whoever certifies the decision-support layer shapes how evidence enters European care.
🔧 TOOL OF THE WEEK Undermind
Undermind
What it is: an AI search agent for the scientific literature, built by two MIT-trained physicists and backed by Y Combinator. Instead of matching keywords, you describe your question in a full paragraph and its Deep Search agent runs several rounds of searching, reading candidate papers and refining as it goes, across a Semantic Scholar index of roughly 200 million papers. A typical search takes a few minutes and returns a ranked set of papers, each with a match score and a stated reason for inclusion, so you can see why a paper surfaced. GSK reports more than 1,000 of its scientists using it inside its internal research stack (a vendor-published case study), and a librarian's review in the Journal of the Canadian Health Libraries Association judged it a useful way to start a biomedical search.
How to use it for Medical Affairs and MSLs: when a field insight or an unsolicited request raises a question that keyword search handles badly (a mechanism across indications, an emerging safety signal, a subgroup that nobody names consistently), write it out as a paragraph and let Deep Search find the seed papers; before refreshing a publication plan, use it to see which of your own studies surface for the questions HCPs actually ask, a rough proxy for this week's Source check; and before trusting it in your therapy area, describe a published systematic review you know well and count how many of its included studies come back.
Where to access it: undermind.ai on the web. A Free tier includes Deep Search with standard limits; Pro is about $16 per month billed annually, Team about $15 per person per month, and Enterprise is custom-quoted. Confirm current limits on the site before you buy.
COMPLIANCE NOTE
Undermind is a discovery aid, not a validated systematic review method or a source of MLR-approved copy. Its relevance screening runs mainly on titles and abstracts, so read every paper before you cite it, and never treat a Deep Search as the documented, reproducible search strategy that a regulatory submission or PRISMA-style review requires. Do not paste unpublished data, patient-identifiable information, or pipeline detail into the query box, and check your company's data-handling rules before connecting it to other AI agents. Match scores and explanations make results easier to audit, not correct by default, so every output stays a draft until a human has verified it at the source.
That is Issue 09. 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 09, Monday, September 27th, 2026.
Parts of this issue were produced with AI assistance and checked before publication.