Inside Versiti's Data Strategy: Building AI-Ready Research Systems

By John Oncea, Chief Editor, Clinical Tech Leader

Versiti’s Banu Santebennur explains how clean data, semantic layers, and cautious tech adoption prepare clinical research for AI.

Ask Banu Santebennur what technology problem she thinks the clinical research industry is most underestimating, and she doesn’t hesitate: “It’s going to be all data. Data cleanliness, data usability is going to be key.”

As director of clinical and research systems and strategy at Versiti, the Milwaukee-based blood health nonprofit whose work spans donor collection, diagnostic testing, clinical trials, medical device testing, and rare-disease research, Santebennur has spent the past several years building the data infrastructure that will determine whether Versiti’s investments in AI actually pay off. Her team’s approach and the mission it supports are profiled in a companion piece, How Versiti Turns Data Into A Lifeline For Rare Blood Patients.

Why “Clean Data” Is Harder Than It Sounds

Versiti’s research portfolio is unusually deep for an organization its size. Its Blood Research Institute studies sickle cell disease and thrombocytopenia, among other conditions, and maintains one of the field’s leading centers for glycomics, the study of the sugar structures on red blood cells that define blood group variation beyond the familiar ABO and Rh systems. Versiti also curates the platelet variant registry, a classification system for genetic differences in platelets built under a researcher Santebennur describes as internationally recognized in the field.

That research generates enormous volumes of structured and genomic data, and Santebennur’s team has learned the hard way that data quality problems rarely start as technology problems. “We found out that you think people are saying the same thing using the same word, but in their minds it’s completely different,” she said. Her go-to example: the term “donor.” Is someone a donor the moment they register online, when staff call them, or once they’ve completed a donation? Every answer changes what a metric means, and, by extension, what any AI system trained on that metric will tell you.

Versiti’s response has been to build data change advisory boards: standing groups of subject matter experts, organized by service line, with explicit authority to settle definitional questions like these. It’s slow, deliberately so. “We’re not doing it all at once,” Santebennur said, but the organization has set a target of having at least one service line’s data clean enough for AI-driven self-service querying by the end of 2026.

The Semantic Layer As Infrastructure

The technical piece of that work is what Santebennur calls a semantic layer, a translation layer sitting above raw data tables that expresses what each field actually means, rather than leaving that meaning implicit in a column header. “We have to make sure that our data is clear enough for a dumb AI to understand it,” she said. “It has to be broken down to that extent.” Rather than a field simply labeled “donor,” the goal is a fully explicit definition attached to every measure, dimension, and source Versiti’s teams rely on.

That investment reflects a specific, well-founded skepticism about how large language models handle the kind of data clinical research organizations actually have. “Most LLMs don’t do well with structured data,” Santebennur said. “It doesn’t have context when you have structured tables the way it does if you have a narrative.” Given that nearly all of Versiti’s clinical, laboratory, and donor data lives in structured tables, that’s not a minor caveat; it’s the central design constraint her team is building around.

Interoperability, Beyond The Acronyms

Versiti’s structure, six service lines spanning collection, diagnostics, trials, device testing, and research, means data has to move cleanly across boundaries that were never built to align. Santebennur points to standards like FHIR and HL7 as necessary but insufficient on their own, because implementations vary enough that “flavors” of the same standard can still break interoperability in practice. Her team’s fix is granular: standardizing something as specific as the label printed on a test tube so that any Versiti employee handling it, regardless of which service line they work in, can trust what it says without having to ask.

That discipline connects directly to two of Versiti’s larger research infrastructure commitments: the NIH’s All of Us Genomic Project and the REDS Program, a federally funded initiative that links donor data to transfused-patient outcomes across research hubs in the U.S. and Brazil. Versiti has participated in four phases of REDS, acting as what Santebennur calls an “honest broker”, aggregating donor and patient medical history, de-identifying it, and making it usable for researchers studying blood and disease at a population level. None of that works without the underlying interoperability and data-definition discipline her team has been building service line by service line.

Adopting Technology On Versiti’s Own Timeline

Versiti’s approach to research technology adoption is deliberately unhurried, a function of its nonprofit funding model as much as its risk tolerance. “One of the decisions our CIOs made is that we are not going to be the early adopters of new technology,” Santebennur said. “We look for that sweet spot when we know that a technology is getting established enough for it to last longer.” Next-generation genetic sequencing is the clearest example: a technology Versiti brought in once it matured enough to analyze the unusually complex genetic structures behind some of the rare diseases it studies, at a cost that had fallen from prohibitive to workable.

AI tools get the same scrutiny, filtered through a practical cost question. Santebennur described weighing a specialized AI platform that might cost roughly $100,000 a year against a straightforward SQL server running analyst-built queries for a fraction of the cost. “It’s not doing anything great enough for me to move things over,” she said of premature AI investment. “We’ll still have analysts run reports. It’s okay. We’ll do that a little longer until all of this settles.”

That caution isn’t resistance to AI; it’s sequencing. Santebennur’s clearest articulation of the strategy: get the data layer right first, because a large language model trained on bad or inconsistent data will still return a confident-sounding, plausible, and wrong answer. “The more bad data you have, that affects the likelihood of it telling you the right answer,” she said.

The Payoff, When It Comes

The eventual goal, per Santebennur, is a research environment where curated, trustworthy data carries its own “stamp” of integrity, usable with confidence by researchers and AI systems alike, while anything outside that curated environment carries none. It’s an ambitious standard, built one defined workflow at a time, and it’s explicitly tied to patient outcomes: faster test turnaround, more precise blood matching for patients with rare antibody profiles, and research infrastructure sturdy enough to support the next gene therapy or the next diagnostic breakthrough.

For technology leaders elsewhere in clinical research, Santebennur’s data-first sequencing offers a useful corrective to AI-adoption pressure: the infrastructure work that looks the least exciting – definitions, semantic layers, governance boards – is also the work that determines whether every downstream AI investment succeeds or quietly fails. As she put it, describing the industry’s broader underestimation of the problem: “I think that’s going to start surfacing up after people invest a lot of money into these AI platforms. They've got to realize, oh my gosh, there’s a whole other thing we need to do.”