"The most dangerous phrase in healthcare is 'we've always done it this way.' AI doesn't fix that. It just makes it faster."
Picture this: You're a CIO. Your board just greenlit a $50 million AI initiative. The vendor deck is beautiful. The pilot metrics are stunning. Your CMIO won't stop quoting the ROI projections in board meetings.
Six months later, your clinicians are quietly routing around the system. Not because it crashed. Not because it's inaccurate. Because it works — and when it works, it tells them something they don't want to hear about their own value.
This isn't failure. This is sabotage by success.
And the numbers that capture why? Nobody puts those in their vendor deck.
📊 The Signals That Actually Matter This Week
This newsletter doesn't track deal flow. It tracks what happens after the deal closes.
| AI pilots that reach scaled deployment | ~20% |
| Gen AI pilots delivering measurable financial impact | 5% |
| Health IT projects that fully deliver on stated goals | 30% |
| Time a primary care physician loses daily to AI alerts | 49 min |
| Orgs with formal AI upskilling programs for clinical staff | 14% |
| Clinicians who feel AI serves them — not the org | Not measured anywhere |
That last row is the one that should haunt every CIO. Nobody is measuring whether clinicians feel AI serves them. We measure adoption rates, ROI, and alert volumes. We don't measure the thing that determines whether any of it sticks.
This isn't a technology problem. It's a deployment-without-redesign problem. And it's being replicated at scale, right now, with every new dollar that enters the market.
The Capital Cascade: Funding What We Know Won't Work
Let's not celebrate the investment headlines. Let's interrogate what they're buying.
Billions are flowing into efficiency automation, workflow optimization, clinical decision support. Noble categories. Necessary categories.
Categories that make broken processes faster.
Not better. Faster.
McKinsey estimated up to $1 trillion in annual value could theoretically be unlocked by AI in global healthcare. The word "theoretically" does a lot of work in that sentence. Meanwhile, fewer than 1 in 5 AI use cases successfully move from pilot to scaled deployment.
The hiring signal is even more telling. A Fortune 500 health system is pulling AI engineers from Big Tech — at salary bands that dwarf what their senior clinical staff earns. You know what that says to the people who've been there for 15 years?
"We don't think you can keep up."
That's not a technical strategy. That's a cultural grenade with a slow fuse.
And the partnerships? Absent. Because genuine collaboration requires admitting you can't do it alone. Healthcare systems — built on referral politics and competitive paranoia — would rather buy five point solutions than share one platform.
Three Signals That Tell You Where This Ends
The majority of this week's deals fund "efficiency platforms" promising 20–30% reduction in administrative costs. Some will hit those numbers. For about 18 months. Then the ceiling appears.
Efficiency without role redesign is a paint job on a sinking ship. You look good at the next board meeting. You do not fix the hull.
The real question nobody is asking: Should this process exist at all? Answering that requires cultural courage. Automating around it requires only a contract signature. Guess which one gets funded.
Proof: Menlo Ventures analyzed 52 enterprise healthcare AI deployments in 2025 — 95% of generative AI pilots failed to deliver measurable financial impact. The consistent pattern wasn't technical failure. The workflows were never redesigned to absorb the AI.
When health systems recruit externally for AI talent at scale, they are announcing something: We have given up on our own people.
The outcome is predictable. You get bifurcated teams. Tech outsiders who understand transformers but not triage. Clinical insiders who understand the work but resent the disruption. The AI gets built in a lab. Deployed in a war zone. Adoption flatlines.
The alpha signal: Watch for health systems investing in internal AI literacy programs — reskilling nurses, analysts, and administrators rather than importing engineers. Qualitative research in PMC found a consistent pattern: clinicians believe AI gets "implemented for the organisation, not for them." The systems hiring internally and retraining existing staff are the ones quietly building durable adoption.
Zero major partnerships this week. This has been the pattern for months.
AI thrives on integration — on data flowing across systems, on models that see the full patient rather than one touchpoint. Point solutions, no matter how sophisticated, produce point improvements.
The industry knows this. And keeps buying point solutions anyway. Why? Partnerships require data sharing agreements, legal exposure, and the admission that your EHR is a mess. Buying a clean-looking vendor product lets you announce progress without confessing dysfunction.
The contrarian bet: Whoever builds the integration layer that forces these systems to talk wins the decade. Not the fanciest model. The best plumbing.
The 2027 Reckoning
The consensus view: AI will transform healthcare. The more likely near-term view: a meaningful correction is coming — probably 2027–2028 — that will look like failure but is really just the gap between what was promised and what was built catching up with everyone.
The dominoes:
-
Plateaued ROI becomes impossible to hide. Health systems have been averaging AI investments against pilot-era metrics. When enterprise-scale deployment data hits CFO dashboards, the gap will be stark. Gartner's 2025 Healthcare AI Hype Cycle already places clinical decision support and ambient documentation firmly in the "Trough of Disillusionment." The slide is underway.
-
"Transformation fatigue" will be the clinical burnout story of 2027. Physicians already experiencing record burnout levels (54% in 2025 per AMA data) are now being asked to adopt an average of 3.2 new AI tools per year per institution. Each one comes with training, workflow changes, and the implicit message: adapt or become obsolete. This is not sustainable.
-
Regulatory backlash from overpromising. The FDA cleared 950+ AI-enabled medical devices through 2025. A meaningful number were cleared on the strength of narrow validation studies. When real-world performance diverges — and it will — congressional hearings follow. The sector is one high-profile failure away from a chilling regulatory event.
Five Asymmetric Bets for the Next 12 Months
These aren't consensus views. They're where the market is mispricing risk and opportunity.
Any vendor promising sub-12-month ROI on clinical AI without a dedicated change management component is selling you a number, not a transformation. Their renewal rates will tell the truth. Watch churn data.
The health systems quietly running AI literacy bootcamps for existing staff are building something competitors can't buy: contextual institutional knowledge fused with AI capability. This compounds. Hiring engineers from Meta does not.
The unsexy infrastructure bet: whoever solves interoperability at the AI layer — not just FHIR compliance, but genuine model-to-workflow integration — captures the margin. This is a platform play masquerading as a plumbing problem.
The emerging category almost nobody is funding: AI tools designed explicitly for workforce transformation — helping clinical leaders understand resistance, model change management scenarios, and track adoption psychology. The bottleneck is human. The solution category should be too.
Any vendor projecting meaningful ROI before Month 18 without robust change management embedded in the contract should trigger immediate skepticism. The honest ones will price time into their models. Find them.
The Mirror Problem
Healthcare AI isn't stalling because of model quality. The models are extraordinary.
It's stalling because success makes the human problem visible.
When an AI coding tool starts closing more prior authorization appeals than the team that's been doing it for a decade, the question it raises isn't operational. It's personal. What am I for? That's not a technology question. It's an identity question. And no implementation plan addresses it.
Organizations that treat AI as a mirror — using it to surface dysfunction and then doing the hard cultural work to redesign — will compound. Organizations that treat AI as a magic wand will plateau and blame the vendor.
The alpha is in who's willing to look in the mirror. That's the edge. Watch for it.
A few of you asked for something more hands-on — less analysis, more doing. So we're trying something. 13 weeks, one working app, no code required. A pharmacy price tracker that you actually build and use.
This is Week 1. Here's exactly what we're building, where we're going, and what "done" looks like by the time you put your phone down tonight.
Today you're completing Phase 1, Week 1. By the time you're done, you'll have a live Google Sheet tracking real drug prices and your first AI conversation proving this whole thing is possible. That's it. That's the win for this week.
Go to claude.ai and sign up for a free account. This is your AI partner for the entire project — the thing that will do the thinking so you don't have to.
Once you're in, start a new conversation and paste this exact prompt:
Read what it tells you. Screenshot the response. Save it somewhere you'll find it — this is the intellectual blueprint for everything we're building.
Stay in Claude. Send this follow-up prompt:
You'll get something like: price comparison → location suggestions → generic alternatives → insurance integration → automated alerts.
Copy the list into a note (Apple Notes, Google Keep, doesn't matter). This is your 13-week feature roadmap. You just did product strategy with an AI in under 3 minutes.
Open Google Sheets and create a new spreadsheet. Rename it "My Pharmacy Tracker." Set up these exact columns:
| Drug Name | Dosage | Pharmacy | Price | Location | Date Checked | Notes |
|---|---|---|---|---|---|---|
| Metformin | 500mg | CVS | $12 | Chicago, IL | Apr 7 2026 | Generic |
| Lisinopril | 10mg | Walgreens | $8 | Chicago, IL | Apr 7 2026 | Generic |
| Atorvastatin | 20mg | Costco | $6 | Chicago, IL | Apr 7 2026 | Generic |
Enter 3 real drugs relevant to you or your family. Use your actual city. This isn't a demo — this is your real data from day one.
Go to bubble.io and sign up for the free tier. Create a new app called "My Pharmacy Tracker."
Add two elements to your first page using Bubble's drag-and-drop editor:
- A text input box → label it: "Enter Drug Name"
- A button → label it: "Compare Prices"
Don't worry about making it work yet. You're just laying the bones. Take a screenshot. You now have a real app screen. The gap between "I could never build an app" and "I am building an app" just closed.
Back in Claude, send this final prompt for the week:
Copy the response into the Notes column of your sheet. This is the embryonic form of the AI recommendation engine we'll build out over the coming weeks. It's rough. It's manual. And it works right now.
- A Claude account — your AI agent is ready
- A feature roadmap saved in your notes 5 features, ordered by complexity
- A Google Sheet called "My Pharmacy Tracker" 7 columns set up correctly — at least 3 real drug entries with real prices + your city
- A Bubble.io app called "My Pharmacy Tracker" One page · One text input labeled "Enter Drug Name" · One button labeled "Compare Prices"
- A screenshot of each — your Week 1 proof of work
If you have all five of those, you did Week 1. You're not behind. You're not guessing. You have a foundation. If something didn't work — Bubble confused you, Claude gave you a weird answer, the Sheet formatting looks wrong — drop a comment below. I read every one.
Next week: We'll connect real-time pricing from GoodRx and NeedyMeds directly into your Sheet, so your prices update automatically. Week 2 is where the tracker starts to feel like a real tool.