What I’m thinking about: Healthcare practices are buying AI one tool at a time, and the gaps between those tools are where the money quietly leaks out.
None of the 14 conversations I planned for San Antonio happened, and that was my call.
Quick recap for anyone who missed last week’s issue. With healthcare the lead vertical for Serious AI Partners, I bought a ticket to MGMA’s annual conference (the Medical Group Management Association, the flagship event for medical practice leaders) just a few days before it began. There’s no full attendee list, so Claude researched 100+ speakers and their organizations and tiered them down to 14 targets.
Each of the 14 got a printed, 4-page outside read (how I’d use AI on what was surfacing in their reviews, complaints and website gaps), a handwritten note on top, and a time-gated plan for when I was most likely to cross paths with them.
Then, minutes before I planned to send those reads…MGMA emailed attendees a pointed reminder on its strict no-solicitation rules, prohibiting “suitcasing” (working the aisles to land clients without paying to exhibit), noting it “reserves the right to revoke conference credentials and ask the individual to leave the conference without refund.”
(In this case, my procrastination saved me! I had forgotten how much medical providers tend to be pitched by pharma, med devices, and vendors of all kinds.)
My wife and I drily joked that getting embarrassingly walked out of a room full of potential allies or clients would have torched my credibility before the healthcare vertical even started, so I went purely as a participant (how we reworked the conference plan with Claude Code at the last-minute, both before and during, is next week’s story).
That still left the floor under the bet I described last week, consisting of healthcare operators describing their AI problems in their own words…and I attended 18 sessions over 2.5 days, had several illuminating conversations, and joined ~1,700 attendees at MGMA’s 100th annual conference.
If you don’t run a medical practice, stay with me…every problem below has a version in your business.
AI was in EVERY room
AI was the topic du jour, the AI-specific sessions ran standing room only, and some of the busiest operators in the country were left wondering whether to do more with AI (or at least feeling FOMO), while doubting they even fully use what they already bought.
The numbers backed that up. In one session, only 6% of ~100 practice leaders said they run a mature AI program (the largest group, 41%, said they were still in early pilots). MGMA’s national poll in August found AI in patient visits at 83% of medical groups, yet fewer than 3 in 10 users said their workload actually went down.
Part of the reason is how it’s being bought…one tool at a time, a phone vendor here, an ambient scribe there (software that listens to the patient visit and drafts the note, like the AI notetaker on your Zoom calls), with no plan tying them together.
Part of it is timing, too. AI agents only got good enough for ‘useful’ work in 2026 (after roughly the first quarter IMO), so adoption is still running well ahead of results in many cases, even with a TON of low-hanging fruit in this industry.
Meanwhile, the other side of the table is moving fast. Insurers are bulk-automating denials (when an insurance company refuses to pay a claim), startups like Assort Health (which was present at the conference) raised $120M in June, and Oracle is folding AI into its electronic health record (EHR).
These buying decisions are getting made this quarter, whether practices are ready or not.
The blind-spot triangle
Across those 18 sessions, the biggest gaps sorted into three sides…whether you’re paid what you’re owed, whether your AI’s work is right, and whether anyone owns the whole picture. Each one comes down to a single number most practices can’t name yet.
1. Getting paid what you’re owed
Insurers (payers, in healthcare speak) can underpay without sending a denial. A practice bills a visit as a 99214 (a CPT code, the five-digit codes that define each service and what it pays), and the payer pays it as a lower-paying 99213. No denial, no red flag…it’s called silent downcoding, and Maryland fined Cigna $80,000 (that seems…low) in March for doing it automatically.
Contracts hide it too. Many commercial insurance contracts price each service as a percentage of Medicare’s rates, since Medicare publishes a fee schedule (what it pays for every billing code) that the whole industry uses as its yardstick. So “150% of Medicare” should mean getting paid one and a half times what Medicare pays for the same visit.
One contracting session walked through a doctor whose contract read exactly that, but an addendum tied those rates to the payer’s own fee schedule (a price list the payer sets itself), so he was collecting 89% of Medicare on his top 25 codes…less than Medicare itself pays.
When the presenters asked who checks what they’re paid against their contract, fewer than 10% of hands went up. That’s one room (MGMA’s 2025 poll had 48% auditing monthly or quarterly), but they also said no practice they’d walked into as consultants could produce all of its contracts…and you can’t check a rate you can’t find.
To be fair to the payers (typically viewed as punching bags), Scott Galloway’s piece on Friday cites new JAMA research showing rising spending on care accounted for 91% of private premium growth from 2011 to 2024, while insurers’ cut shrank as a share (though it still grew ~48% in dollars per enrollee). Payers now blame providers’ AI too, with Blue Cross Blue Shield claiming hospital AI coding added $942M in costs (hospitals dispute it).
Both sides are automating, and each points to the other’s AI as the reason. The side holding a documented record (or using overwhelming processing speed, combined with volume) tends to win the dispute.
The number here is paid versus contracted, code by code. Outside healthcare, swap “payer” for any customer or platform paying you on terms you don’t audit.
2. Whether the AI’s work is right
A specialty practice said its AI scribe ignored their documentation rules 20 to 25% of the time, so fixing notes took about as long as writing them.
Headlines need checking too. One session’s slide claimed ambient AI cuts documentation burden 40%, with no source. To the presenters’ credit, a later slide showed one health system’s actual rollout across 1,900+ clinicians, where time in notes fell 15% (obviously still directionally positive, but less extreme than the initial claim).
The bigger win was one no headline mentioned…charts closed in 0.9 days instead of 15, so claims go out (and cash comes in) far sooner. That’s the number I’d focus on.
AI can produce the note…but it can’t be wrong. As in, AI can’t sign the chart, hold a license or carry the insurance. Per one healthcare privacy attorney, AI vendors disclaim responsibility and malpractice carriers are starting to carve scribe errors out of coverage (her slide read “Do not rely on your malpractice policy”), so the clinician holds all of it, and verification belongs right at that signature.
Coding is where that verification gap gets expensive. Coders are the staff who translate each visit into billing codes, and one group with 250+ providers runs ~50 of them, held to 95% accuracy, with much of the checking still done by hand in spreadsheets. That just doesn’t make sense to me…coding is a verification task, perfectly positioned for AI to take the first pass and people to check a sample.
The upside of checking hides in plain sight, too. A senior operations analyst at a 200-provider group found ~$753K a year in missed revenue (her numbers, my math) with, in her words, a “Simple report, a pivot table, and CURIOSITY.” Now imagine AI running that same scan every month, across every payer and every code, with her reviewing what it flags.
Plenty of people in those rooms are worried about what that means for their team’s jobs (…and silently, probably their own). The smarter practices I heard from are handling it through attrition, so when someone leaves, the role simply isn’t refilled, which I think is smart, good for morale and a more humane approach. Cash-flow pressure may still force tougher calls, though, especially for practices competing against groups willing to cut staff fast.
3. Who owns the whole picture
Credit where it’s due, many vendors I met or listened to are sharp and well-funded, and several measured ROI with the practice’s own numbers (most practices would rather grow throughput, meaning more patients booked, than cut costs). Most of them work the front door of a practice…phones, scheduling, referrals, faxes. Only a handful were tackling coding, and plenty are selling virtually the same thing as the next booth.
At the moment, there’s still plenty of room for them. Healthcare is a huge industry that runs on relationships, so a vendor can build a large business serving its own selection of operators, and whether this market ends up winner-takes-all or winner-takes-most remains to be seen.
Somebody still has to own the whole picture. My read on where this is heading is that as the cost of high-powered AI output keeps falling, it makes more and more sense to shape the underlying models (e.g., Claude Code, OpenAI’s Codex) around how a specific practice actually runs, layering in vendors where they’re clearly best (AI voice agents are already very well done), versus stitching together a purely vendor-based stack.
That takes someone owning AI implementation in-house, and if nobody on the team has the capacity, a fractional Head of AI. I think pretty much every mid-market or larger business is going to need that seat going forward.
Your single number
Pick one AI tool you pay for (or are about to) and write down the single number that would prove it’s working. If you can’t name it, that’s your finding, and it’s the constraint to go after first.
Healthcare is where I’m starting, but the triangle travels…any business that buys AI, signs contracts or trusts a vendor’s dashboard has its own version, and the operators who can prove their numbers are the ones who are best positioned to win in this rapidly changing business environment.
=====
Whatever your industry, the work starts the same way…name the number, find the constraint actually gating it, then build with verification right where a human still signs. It’s how you become the operator whose AI investments come with receipts.
Healthcare is where we’re going deepest first (see how), and everything else lives at seriousai.partners.