23andMe’s Data Readiness: A Billion-Dollar Imperative for AI Scale

The Unmoved Data Readiness Baseline

Back in 2022, Innovaccer published research that put a number on what we all knew: chronic data readiness issues were kneecapping digital transformation in healthcare Innovaccer research on chronic data readiness. That report became the baseline for how bad things were. Fast forward to 2026, and there’s no big new study saying we’ve fixed everything. Instead, new Innovaccer reports just keep proving the original point. A recent one on AI for payers found 86% of them aren’t ready to use AI at scale, and the top barrier is still interoperability. This tells me the fundamental problems from 2022 haven’t gone anywhere. The original research pointed to the usual suspects: fragmented data, no standardization, poor interoperability, and just plain bad data quality. An AI model needs clean, consistent, and complete data to function, so feeding it the average health system’s data is like building a skyscraper on sand. No matter how sophisticated the model, failure is baked in from the start. That 2022 baseline is still so relevant because these aren’t problems you fix with a software patch. They’re embedded in the very structure of how hospitals operate. So, for a health system leader, the question isn’t whether AI can work in a lab. The question is, can your specific data environment reliably feed and sustain it? The fact that the baseline hasn’t budged tells me the answer for most is still no, which makes deploying AI risky at best and dangerous at worst.

Julia Adler-Milstein and Data Governance as Infrastructure

To get why this static baseline is such a big deal, you have to look at the work of informatics researcher Julia Adler-Milstein. Her name popping up in 23andMe’s own materials around their data breach tells you everything. Her research frames data governance not as administrative busywork but as the load-bearing infrastructure it has to be, especially when you’re dealing with irreversible genetic data. Adler-Milstein’s work consistently shows that good data governance is the architectural blueprint for how data is collected, stored, secured, and used, it isn’t just a set of policies. When a company like 23andMe mentions her work while talking about its own breach, it’s an admission that governance is directly tied to security, data integrity, and patient trust. For anyone trying to deploy AI, this changes everything. The goal shifts from just having “data” to having “governed data.” An AI fed poorly governed data can cause a cascade of failures, from biased results because the data wasn’t representative, to privacy violations because access controls were a mess, to bad clinical decisions because the source data was wrong. Thinking about it this way reframes governance from paperwork into the actual foundation that determines if your AI tool is safe and effective. Julia Adler-Milstein’s research on health data governance So, when you put the Innovaccer baseline next to Adler-Milstein’s insights, you’re left with a very stark question: Can the data layer actually support what you’re trying to build on top of it? This is pure deployment risk assessment. The risk is deploying AI on data foundations that were never ready, and that risk gets a lot bigger when governance is an afterthought instead of the first thing you build.

What Still Stands in the Way

With the 2022 baseline holding firm and the idea of governance-as-infrastructure in mind, it’s clear what’s still stopping health systems from being ready for AI.

• Fragmented Data Silos: After decades of trying, a patient’s data is still scattered. EHRs, lab systems, imaging archives, and billing databases are often totally isolated. This makes it impossible to build the complete, longitudinal patient record an AI needs to make accurate predictions. Operating with an incomplete picture, an AI will always produce incomplete or just plain wrong outputs.
• Lack of Data Standardization and Interoperability: Even if you can get to the data, it’s often a mess of conflicting terms, codes, and formats. One system’s code for a diagnosis won’t match another’s, making it impossible for an algorithm to reliably pull everything together. Despite initiatives like FHIR, data exchange is still a complex, often manual process that can’t support the real-time data flow that advanced AI requires.
• Data Quality and Integrity Issues: Data readiness is about quality, not just quantity. Our datasets are riddled with missing values, typos, old info, and duplicates. These flaws inject bias directly into AI models, which is how you get AI health misinformation news and bad recommendations. For instance, an AI chatbot giving drug interaction advice from an incomplete medication list is a patient safety nightmare waiting to happen. NIST guidance on AI trustworthiness and data quality
• Insufficient Data Governance Frameworks: Most places have policies, but the actual implementation of those policies as working infrastructure is what lags. This means having clear data owners, strict access controls that go beyond basic HIPAA, and a way to trace data lineage. Without these structural pieces, the entire AI data lifecycle, from collection and annotation to model training, is precarious and open to breaches or non-compliant use.
• Skills Gap in Data Science and Engineering: There’s also a people problem. Most health systems don’t have enough in-house data science, ML engineering, and clinical informatics talent. This forces them to trust vendors completely, without the internal expertise to ask hard questions about the data requirements, the model’s performance, or its hidden biases. These are the deep, structural problems that explain why the 2022 baseline hasn’t moved. Fixing them takes more than buying new software. It requires sustained investment and a total shift in how we think about data.
  

  What an Updated Baseline Would Need to Show

  Any health system leader listening to an AI vendor’s pitch needs to know what a good data readiness baseline would actually look like. A vendor’s claim that their AI is “clinically validated” is meaningless if your data environment can’t support it. Validation has to apply to the real-world conditions of your system. A new, meaningful readiness metric would go beyond just having data and would have to include:

• Quantifiable Interoperability: Hard numbers showing the automated flow of data between systems, with real semantic interoperability. You’d be able to measure the reduction in hours spent on manual data entry just to feed the AI.
• Data Quality Scores: Objective, audited scores for data completeness, accuracy, and timeliness in the specific clinical areas you’re targeting with AI. This isn’t a one-time check. It’s continuous monitoring.
• Governance Maturity Index: An assessment that shows your governance framework is actually being enforced, clear data ownership, access logs, and audit trails for every piece of data an AI touches. This is about proof, not policies.
• AI-Specific Data Curation Capacity: You need to show you have teams that can actually do the work of labeling, de-identifying, and checking data for bias before it ever gets to a machine learning model. It’s a proactive capability.
• Scalability of Data Infrastructure: Proof that your data lake or warehouse can handle the demands of multiple AI applications at once without slowing to a crawl or compromising data integrity. Without these kinds of measurable signs of progress, the promise of AI is still stuck in that 2022 reality where the foundational data problems make every deployment a gamble. Julia Adler-Milstein’s work is a constant reminder that treating data governance as anything less than core infrastructure isn’t just a policy mistake, it’s an architectural flaw that leads directly to AI failures. So, the question for data and analytics leaders hasn’t changed: is your data layer ready for the AI you’re putting on it, or are you building on a foundation you know was never prepared?