The Gap Between Having Data and Trusting It
On October 7, Teradata published results from a Wakefield Research survey of 1,000 senior technology and data leaders that puts a number on a problem most healthcare CIOs already feel in their gut. Ninety percent of healthcare respondents say twenty percent or less of their data is described well enough for an AI agent to act on it reliably. That is not a storage problem or a compute problem. It is a description problem: healthcare organizations have decades of institutional knowledge sitting in systems where the meaning of a field, the provenance of a value and the rules for using it live in someone's head rather than in machine readable metadata.
Teradata's Chief Operating Officer, Mike Hutchinson, cut straight to the point: "The problem has never been the data." That framing matters because it redirects the conversation away from the usual AI readiness checklist item, more data, more compute, and toward the harder and less glamorous work of describing what you already have. Healthcare is a useful stress test for this argument because the data is simultaneously enormous in volume and extraordinarily sensitive in use, which means governance cannot be an afterthought bolted on once the model works.
Disease Prediction Without the Usual Integration Tax
The first case Teradata highlights involves a European research institution running a compact model, roughly two million parameters, that predicts disease onset up to ten years in advance with seventy percent accuracy at the population level. The headline figure most CTOs will actually care about is the three day deployment window Teradata reports for getting that model running in database, a fraction of the months a typical integration project between a research model and a production data environment usually takes.
That speed comes from avoiding a data movement step entirely. Instead of extracting data to a separate environment, transforming it, and feeding a model somewhere else, the model runs where the governed data already lives. For any CTO who has watched a promising pilot die in the handoff between the data science team and production infrastructure, this is the part of the story worth interrogating with Teradata directly: how much of that three day number depends on data already being well described, versus Teradata's tooling doing the describing for you.
A Five Hundred Percent Jump in Patient Engagement
The second case comes from a large US integrated health system running Python based machine learning models for patient risk stratification, drawing on nearly 150 patient features, directly inside the database. Teradata reports a five hundred twenty percent increase in patient engagement across the health system's digital programs after the move. That is a striking number, and it deserves the same scrutiny any vendor supplied case study deserves: Teradata does not disclose the baseline engagement rate the increase is measured against, nor the time window over which it was observed.
What is verifiable and useful regardless of the exact multiplier is the architectural claim underneath it: running risk models against nearly 150 features without extracting that data out of the governed system removed a step that most health systems treat as unavoidable. If your data science team is routinely pulling patient features into a separate notebook environment to run stratification models, this case is a reasonable prompt to ask your platform vendor why that extraction step still exists.
Compliance Work That Never Leaves the Database
The third engagement is less flashy but arguably more consequential for regulated industries generally. A major US healthcare payer applied large language models, embeddings and in database similarity matching to a high priority regulatory compliance initiative, with Teradata emphasizing that no data moved outside the database boundary throughout the process. For a payer, that boundary directly determines whether an audit passes cleanly or triggers a much longer conversation with regulators about where protected data traveled and who touched it.
This is the pattern across all three cases: the AI capability is almost secondary to the governance posture that makes it deployable at all. Prediction models, risk stratification and LLM based compliance work are all available technology today. What separates the organizations that can actually ship them in healthcare is whether their data platform lets AI operate inside the existing governance perimeter instead of requiring a new, less controlled one.
Metadata Debt Is the Real AI Readiness Gap
Put the survey finding next to the three case studies and a clearer picture emerges than either on its own would give you. The ninety percent of healthcare leaders who say most of their data is insufficiently described are not lacking AI tools, data volume or even executive appetite. They are carrying metadata debt accumulated over years of systems built for human analysts who could ask a colleague what a field meant, rather than for an agent that needs that answer encoded somewhere it can read.
That debt does not get paid down by buying another model or another agent framework. It gets paid down by the unglamorous work of documenting lineage, defining terms and attaching governance rules to data assets, which is exactly the work Teradata, Precisely, Ataccama and several other vendors are all now racing to automate with AI itself. Healthcare CTOs evaluating any AI agent initiative this quarter should budget for that documentation work explicitly, rather than assuming it will happen as a side effect of the AI project.
What to Ask Before Your Next AI Pilot
If your organization is sitting closer to the ninety percent than the ten percent on data description, the honest move is to say so before committing a roadmap quarter to an agentic AI pilot that depends on data your agent cannot actually interpret. Ask your data team what percentage of your production tables have current, accurate business definitions attached, not what percentage have a name in a catalog. Those are very different numbers, and the gap between them is usually where agentic AI projects quietly stall.
Teradata's case studies argue most convincingly for in place governance as a design principle, beyond any pitch for its own platform specifically: keep the data where it is, keep the controls intact, and bring the AI to the data rather than the other way around. Whichever vendor you build this on, that principle, not the choice of model or framework, is what will determine whether your next healthcare AI pilot becomes a production system or a slide in next year's postmortem.
