IBM Watson might not deliver on its lofty goal to revolutionize medicine with artificial intelligence, as a recent investigation suggests.

But that doesn’t mean new technology will be incapable of improving health care.

IBM put a heavy dose of investment and hype into the medical potential of its machine learning system Watson. It was supposed to help doctors make decisions on how to identify and treat diseases, most specifically cancer.

STAT News reported this week that Watson has failed to gain much adoption in hospitals and has proven ineffective at recognizing various cancers.

Experts who have spent years researching the space told CNBC that IBM blundered by prioritizing headlines over peer-reviewed evidence. Still, there are plenty of reasons to be optimistic about the potential for machine learning to help detect patterns of disease.

“What bothers me is that exposing hype makes people wonder whether any of this is real and whether the whole thing is marketing,” said Robert Wachter, a practicing physician and chief of the Division of Hospital Medicine at University of California, San Francisco Medical Center. To that question, “the answer is a strong no,” he said.

In the past five years, doctors across the country have moved from paper-based systems to electronic ones, with more than 90 percent of hospitals having made the shift. All this new digital health data is still messy and challenging to access, but it’s more available than ever.

The opportunities to derive recommendations from these data sets are clear, Wachter said. UCSF is partnering with Alphabet‘s machine learning group Google Brain to make better use of this data, once it’s stripped of personal information.

Wachter is confident that these tools can support specialties like radiology, where clinicians are saddled with a huge amount of data and too little time.

Failure in health care is commonplace for technology companies. Google and Microsoft are among businesses that have tried to build tools for the industry but have fallen short.

Wachter said that too often technology companies ignore real-world circumstances, whether it’s the patient population or the physician’s workflow. Throwing sexy tech at the problem won’t prove useful in the day-to-day operations of a hospital, and it certainly won’t replace doctors, as some have claimed.

Technologists also need to understand the needs of the intended buyer, said Leonard D’Avolio, an assistant professor at Harvard Medical School and the founder of a health-technology start-up called Cyft.

If a company is selling a computer program to a community health clinic serving low-income patients, for instance, the system needs to understand if a patient can afford a treatment option or if a surgeon is available to perform a procedure. Otherwise, it won’t offer useful or actionable suggestions.

One area where Wachter sees potential is in the ability of analytics to help doctors understand a patient’s clinical trajectory.

Suppose a patient is suspected of having pneumonia. After three days on antibiotics, the patient hasn’t really improved. Could a system that’s ingested millions of previous cases help a doctor figure out whether the trajectory is normal or unusual? And whether a different medicine is needed?

D’Avolio said his process at Cyft is to sit down with doctors, nurses and other practitioners to identify their highest priorities, the data at their disposal and their willingness to use new tools.

“Thirty years of research suggests AI does work,” he said. “But what doesn’t work is jamming a new tool down throats of health care and not taking time to learn their workflow.”