This piece was published in the August 2026 Vega Health newsletter, The Signal. Sign up here.
Ask most people in health tech what a university tech transfer office does, and you'll get a shrug. These offices sit inside nearly every research university, and they are the reason a discovery made in a lab or a clinic ever becomes something a patient benefits from. They're also, right now, in the middle of a real transformation driven by the flood of AI models and solutions being built by individuals in academic medicine.
What a Tech Transfer Office Actually Does
Tech transfer offices trace back to the 1980 Bayh-Dole Act, which gave universities the right to own and license inventions developed with federal research funding. "It really didn't pick up until the 90s," says Robin Rasor, associate vice president for Duke's Office for Translation and Commercialization, "and now it's pretty much any research university has a tech transfer office."
The job started narrow: "In the old days we were the patent office," Rasor says. But the job has evolved. Offices now handle software and data alongside patents, and many, Duke included, have grown into full entrepreneurial engines, offering startup support, translational funding, and even venture capital. Duke Capital Partners, the university's angel arm, has invested more than $20 million per year in recent years across deals with a Duke connection.
The scale involved is surprisingly large. Drew Bennett, who has led the University of Michigan's medical school tech transfer practice for close to a decade, points out that Michigan fields between 600 and 700 invention disclosures a year, a volume matched by only two other schools in the country, Stanford and MIT. Before moving into tech transfer, Bennett spent six years as a respiratory therapist in trauma and neurological intensive care, then fifteen years building software startups. Drew is a reminder that the people running these offices often come to the work from inside medicine and inside the technology industry, not from a purely administrative track.
How AI Is Changing the Deal
The intellectual property moving through these offices today looks nothing like the patents the system was built around. Tech transfer offices have adapted from working largely with physical sciences to early software and now AI. "Most of them probably aren't going to get covered by patents," Rasor says of AI tools. Software is more often protected through copyright, Rasor says, and AI may not meet the criteria for patents or copyright. This changes licensing deals significantly.
Many of Duke's AI inventors are clinicians and researchers first: deeply expert in the clinical problem they're solving, and resourceful enough to have already started building a solution themselves, even without a background in shipping commercial software.
Rasor says a platform partner like Vega Health can turn that early work into something usable: "if you can have somebody who's putting them together in a platform... it's kind of helpful." What arrives at her office, she says, is often less a finished product than a strong idea partly built.
"They come to us with either an idea or a half-baked AI agent," she says, "our clinicians are testing it, maybe they're helping with the programming, maybe they're giving ideas how to make it better. It turns into more of a collaboration than a typical licensing agreement.”
Most clinical AI needs exactly this kind of sustained back-and-forth with the clinicians who understand the underlying problem before it's ready to run reliably. A half-baked starting point isn't a mark against the inventor; it's simply where useful clinical AI tends to begin. Getting it the rest of the way takes a partner willing to put in that iteration.
The Case for Academic AI
If there's a case for why this work is worth all the trouble, it's that academic medicine often builds AI better than the tech-first vendors selling into health systems today. Michigan has run its own deterioration-risk tool, called Picture, head-to-head against equivalent products from major electronic health record vendors.
“We outperformed, outperformed substantially," Bennett says. He doesn't think that's a coincidence. An analytic built by "somebody who lives, breathes, and spends all their time in a health setting" will generally beat one built by a vendor managing dozens of other priorities at once, he says: "It's certainly a practitioner-based model. And the reality is they do tend to outperform, for the most part."
Michigan has gone as far as building two institutes, the Weil Institute and its AI and Digital Health Innovation Initiative, staffed by people who've spent their careers on exactly this kind of clinical problem-solving. Duke runs something similar. It's a real, sustained institutional bet that the best health AI comes from people embedded in the clinical work itself, not from generalist software teams working at a distance from it.
The Last Mile Problem
But even if a tool built inside a health system outperforms its competition, it’s not a guarantee the resulting tool is ready to run anywhere else. Rasor points to Duke's own AI innovation group, known as DIHI, which builds tools to solve specific problems inside Duke's health system and does it well. But a model built and proven at Duke doesn't automatically work somewhere else. "The assumption has been: if we fix it at Duke and it works at Duke, it must have value at other health systems," she says. "Well, yes and no. Other health systems, everybody's a little different." Most run Epic, she notes, but how they actually use it, and what their data and workflows look like, varies enough that a Duke-tuned tool isn't automatically deployable at, say, a community hospital in Iowa.
Bennett sees the same problem from the buyer's side, multiplied. Health systems evaluating several AI tools at once, not just one, quickly run into a different kind of ceiling. "If you had to do that with fifteen different vendors for fifteen different analytics, it gets to be really difficult," he says, since each new vendor relationship brings its own integration work and, in his words, "changes your entire security surface that you have to monitor and manage."
And tech transfer offices, for all their scale, were never built to solve either problem. They were built to protect and license IP, not to run sales operations or handle integration work for hundreds of hospitals. Rasor is candid about the mismatch: "The biggest problem tech transfer offices have, especially universities as big as Duke, is we don't have the staff to go out and market these individual things."
Duke has now licensed more than a dozen AI models through Vega Health rather than trying to sell each one on its own. "Having us try and license those individually is not so easy," she says. That partner, in effect, "becomes the sales agent" for the university.
Both Rasor and Bennett describe this shift with genuine optimism, not as a stopgap but as a better-fitting model for a category of technology their offices were never designed to distribute on their own. "It's not like the universities have a lot of choices for these things... let's go, see how it works, " she says of Vega Health.
Bennett, comparing Vega to how mobile app stores solved a similar aggregation problem, likes that a shared marketplace also creates healthy competition between models rather than locking a hospital into just one: "Let the best analytic win," he says, "and make the platform work really hard to maintain trust and engagement with people."
The two conversations point to the same conclusion from different sides of the country. Academic medicine is, by design, an unusually good place to build health AI: specialized for the practice of medicine and considerate of the strains on healthcare organizations. What tech transfer offices aren't built to do is get that work into every hospital that could use it. As Bennett puts it, "You can build the greatest thing on earth, but if you cannot get it to the marketplace, then it's something with tremendous latent potential, but not real impact." Closing that gap is a different job entirely, and it's the one companies like Vega Health now exist to do.
