Hospitals and health systems are deploying chatbots powered by large language models to search and summarize patient medical records, a technology shift driven by how bloated those records have become. A large language model is software trained on vast amounts of text that reads and retrieves information the way a research librarian works through a dense archive. A number of health systems are now moving toward broad implementation of these tools, both ones built in-house and ones supplied by outside vendors.
When the record holds the answer
At Stanford, six pathologists had tried and failed to identify a patient's cancer from a recent lymph node biopsy. The team stained the cells 70 times in search of features that might clarify the diagnosis. Nothing worked. A physician called onto the case was testing ChatEHR, one of several LLM-powered tools now in use by health systems to query and summarize patient histories.
The physician asked ChatEHR whether the patient had any prior history of skin lesions. The tool surfaced a record from a different health system: the patient had previously been diagnosed with sarcomatoid squamous cell carcinoma. That history, the treating physician wrote in feedback for the chatbot, completely explained what had been found in the lymph node. "If that doesn't prove the value of ChatEHR, I don't know what does!" the doctor added.
What the tools are actually built for
That kind of diagnostic discovery is the least of what these tools are built to do. Health systems that began experimenting with generative AI chatbots to search and synthesize patient records are now moving toward broad rollout of the same technology. What they want from it, day to day, is simpler: faster access to information that already exists in a file but is buried too deep to find quickly. Modern electronic health records span years and multiple institutions. Relevant clinical history often sits pages deep in records no clinician has time to read in full before making a decision.
Both homegrown and vendor-built chatbots are entering service. ChatEHR is one named product. The common constraint driving adoption is record bloat: the answer is in the system, but getting to it takes too long. The Stanford biopsy case made that constraint visible. The patient's prior diagnosis was already on record somewhere. It just took a query to find it.