On Centene’s July 28 earnings call, Sarah James of Cantor Fitzgerald asked Sarah London, Centene’s CEO, about the practicalities of Centene’s AI strategy. In response, London said this:
“Thanks, Sarah. My comments were to go a little bit deeper on strategy, less so a shift, but more to point out that we've made good progress in terms of high ROI use cases and deployment of AI. It is still early. It's a little bit more of a philosophy, which is, I think it's easy to get on an earnings call and say, ‘We're doing AI everywhere, and we've got all these great partnerships.’ Our view is that there is also a risk of spending a lot of money on AI and getting no return for it. That's not something that we can afford to do as a Medicaid first company, in a margin compressed environment.
We are really strategically thinking about this around first, the greatest value that will come from all of this and sort of the ability to maximize AI is really predicated on the command you have of your data, the ability to build differentiated data products and to maintain and own the context layer. If you think about being the largest Medicaid managed care, the largest government sponsored programs company in the market with 25 years of history and context, that's really where we are focusing in terms of the foundational work. While we'll continue to highlight places where we're then leveraging that to deploy use cases, we are also holding ourselves to an extraordinarily high bar in terms of the ROI.
That we aren't just going to deploy AI to talk about AI, we're going to deploy it where there is very clear, tangible return on that investment. We think the way to do that is by this foundational focus on data and context, in the short term.”1
(Emphasis added)
Bingo.
If you understand that – and why – Sarah London is correct, feel free to skip to the “Predictable and Predicted” section. If you don’t, then I encourage you to keep reading.
Chuck D famously reminded us not to believe the hype, and perhaps nothing in global history has been more hyped than “AI” has been in the past 12 months. Regrettably, some significant portion of Americans – maybe a majority – believe that “AI” means ChatGPT and Claude. All those people are incorrect.
When the von Trapp children do not know how to sing in The Sound of Music, Fraulein Maria teaches them to “start at the very beginning, a very good place to start.” Elon Musk calls that first principles, which is a good place to start with respect to “AI.”
IBM provides a useful definition and graphic.
“Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.
Applications and devices equipped with AI can see and identify objects. They can understand and respond to human language. They can learn from new information and experience. They can make detailed recommendations to users and experts. They can act independently, replacing the need for human intelligence or intervention (a classic example being a self-driving car).
But [today], most AI researchers, practitioners and most AI-related headlines are focused on breakthroughs in generative AI (gen AI), a technology that can create original text, images, video and other content. To fully understand generative AI, it’s important to first understand the technologies on which generative AI tools are built: machine learning (ML) and deep learning.
A simple way to think about AI is as a series of nested or derivative concepts that have emerged over more than 70 years...”2
It should be self-evident which type of AI is the most applicable to healthcare, probably the industry with more rules than any other. It should also be self-evident which type of AI is least suited to clinical use cases, where variation is undeniably unacceptable and undesirable. Most importantly, it should be self-evident that Centene is not the only health economy stakeholder operating in a margin compressed environment.
Large language models, i.e., LLMs, are probabilistic, the result of math, not magic. Even Gretl von Trapp knew that when you read you begin with ABC, and first principles thinking compels understanding the meaning of each word in the phrase “large language models are probabilistic.” The online version of Merriam-Webster provides these definitions:
“Large” - exceeding most other things of like kind especially in quantity or size : big
“Language” - an organically developed system of communication used by groups of humans: such as the words, their pronunciation, their written representation, and the methods of combining them as used and understood by a community
More simply for a redneck like me, “model” is a fancy word for “made up,” and “probability” is a fancy word for “could be.” In summary, LLMs represent a whole lot of made-up answers that may or may not be true. Or, as George Box said more eloquently:
“All models are wrong, but some are useful.”
More practically, LLMs are the aggregation of information that OpenAI and Anthropic hoovered up from the internet, i.e., other people’s intellectual property. You may know that the internet is full of stupid, incorrect and grossly offensive things, and ChatGPT and Claude have read all of it. If you believe everything you read on the internet, then, in the words of Journey, “don’t stop believing” in whatever ChatGPT and Claude tell you.
If, on the other hand, you do not trust everything on the internet, then you shouldn’t trust ChatGPT and Claude for something as complex as healthcare, especially clinical use cases. The “probability” they express is, by definition, an approximation of the “average” answer on the internet, and that answer is, according to Anthropic, surprisingly easy to manipulate. If you want “an” answer, type away. If you want “the” answer, then you must understand Sarah London’s “foundational focus on data and context.”
Some of you might ask, what about Codex and Claude Code and the claimed gain in efficiencies for your engineering teams? The adage of “trust but verify” still holds true. The bottleneck in software development has simply shifted; code generation used to be the slow and expensive part, which meant plenty of time for planning and verification. Now, code is generated at the rate of millions of tokens per second, but the question of what your system should do, and whether it actually does that thing is still largely left up for humans to decide and verify. More than ever, “garbage in, garbage out” rules the day.
When you learn to read, you begin with ABC. When learning to harness the power of AI, health economy stakeholders should begin with NLP and ML, not LLM. The true opportunity for AI in healthcare is using NLP and ML to automate the numerous mind-numbing, low-value, repetitive, rules-driven tasks that represent a significant portion of the oft-cited, never-proven assertion that “30% of healthcare is waste.”
It is telling that the most talked about AI application in the health economy is “ambient AI,” a “magic trick” to relieve physicians of a bit of administrative and technological fatigue. Penn and Teller’s revelation of the supposed magic would reveal that “ambient” means automated transcription through speech recognition, a technology first deployed in 1987 using hidden Markov models, aka Dragon Dictate, and “AI” means processing the transcribed conversation through an LLM to predict the most likely clinical code for submitting a claim.
The fact that providers and payers are at war over “AI-driven coding” simply reveals that the industry lacks a shared definition of clinical coding. Similarly, the fact that health economy stakeholders cannot or will not automate prior authorization reveals a lack of consensus about the relevant rules for prior authorization, which is a political problem, not a technical one.
Beneath the surface lurks the reality that if, in fact, the U.S. economy is creating new jobs, healthcare is the largest contributor, and every member of Congress represents at least one hospital, and every Senator represents a Blue Cross Blue Shield licensee. The fear that AI will eliminate jobs in the health economy is the real obstacle to deploying NLP and ML at scale.
Curiously, health economy stakeholders are seemingly unaware of the Silicon Valley axiom that if you aren’t paying for the product, you are the product. The same concept applies when a technology company subsidizes the product by offering it below the production cost.
In contrast, Sarah London understands that what Centene knows about the Medicaid market is valuable, which presumably means that Centene is contemplating how to protect that intellectual property – and PHI – from the companies offering “frontier” LLMs. Are you certain that your LLM vendor isn’t using the data that you have exposed to them to train those frontier models? And are you certain that your LLM vendor’s Business Associate Agreement fully complies with HIPAA?
If you are evaluating the deployment of AI in your organization and don’t know what Sarah London knows, you should heed the words of the late Stevie Ray Vaughan: Instead of being “blinded by the neon light,” you should “trust nobody, don’t be no fool.”