- Six nurses told Wired that HCA's AI scheduling tool, Timpani, leaves units short-staffed and ignores their preferences. HCA says nursing leaders make the final decisions.
- The dispute shows that a 'human in the loop' only counts as a control if that person can change the output and the system can be audited.
- Before buying AI that drafts decisions about people, ask who can override it, how fast, and what records prove it.
A human in the loop is only a control if the human can change the answer
Many organisations say a person makes the final call on what their AI produces. That claim is easy to make. It is harder to prove.
A Wired investigation of HCA Healthcare's AI scheduling tool puts that claim under pressure. The facts are contested, and both sides are on record. The lesson holds either way: oversight is a design feature, not a sentence in a policy.
What nurses allege, and what HCA says
Wired reports that HCA co-developed a scheduling tool called Timpani with Palantir. It began rolling out in 2023 and now runs at roughly 130 of HCA's 190 locations.
Wired's account draws on six nurses. One of them, Amber Retzloff, works in critical care in Florida and is a union leader. Their shared complaint is schedules that leave units short of nurses, or short of seasoned ones, with Sundays named as a weak spot. They also say the tool often disregards the preferences they state.
Retzloff says she asked for 50 specific shifts over four months. By her count, the software handed her different shifts more than half the time. She also describes recent shifts where her four colleagues were all junior. As the only veteran, she says she put off care for the sickest patients while coaching the newer nurses.
HCA's spokesperson, Harlow Sumerford, puts the final say with nursing leaders rather than the software. He calls the idea that Timpani is built to cut staffing at patients' expense a misrepresentation. HCA says it keeps improving the tool using nurses' feedback.
Scope matters here. HCA has about 100,000 nurses, and six spoke to Wired. HCA's marketing also features nurses praising the tool as easy to use. These are allegations from some nurses, not a finding about all of them.
How the system works
Timpani is built on Palantir Foundry, a platform that stores and analyses data from across a company's operations. The software predicts how many patients will arrive and how many nurses each unit will need. It then drafts the nursing schedules itself.
That design shifts the manager's role. Managers once built schedules by hand and generally consulted nurses well ahead of overriding a request. Now the software drafts first, and people react to it.
Nurses say the route for reacting has narrowed. Some local managers could edit schedules at first. Now nurses report sending problem shifts to a central team at HCA headquarters in Nashville. They describe an appeal with no direct conversation and no chance to explain.
That is the gap between the stated control and the experienced one. HCA says leaders decide. Nurses describe a remote appeal. Only logs and records could show which account matches daily practice.
Two numbers that describe the same system differently
HCA's figure is that Timpani books nurses on 1 percent of their requested days off, known as red days, on average. Lee Barker, a nurse in Missouri, says the number sounds small. At Barker's facility, though, such assignments did not happen before Timpani. Barker says nearly every nurse there has since been given one.
An average describes the system overall. It does not show who absorbs the misses, or how hard they are to fix. Barker and Retzloff say a shift swap after a denied appeal works only about half the time.
HCA measures success differently. Executive Michael Schlosser wrote in July that over 98 percent of initial schedules include a mix of skill and experience. Nurses describe units where that mix was missing.
The audit question
One lawsuit goes to the heart of oversight. Angelique Russell, a former HCA data science manager, sued in July. She alleges she was fired for raising concerns about Timpani. Her suit also alleges HCA routinely erased the data the tool uses to build schedules. She says that prevented auditing of how well it worked.
HCA has yet to respond formally in court, and Russell declined to comment. These are allegations, not findings.
The wider point does not depend on who wins. If inputs and outputs are not kept, a claim like 'managers make the final decision' cannot be tested by anyone.
A separate case: thin review at Rayus Radiology
Wired also cites internal documents from Rayus Radiology, a US imaging chain. This is a different problem. Rayus began getting complaints this year about a Foundry-based tool that fills in scheduling fields from doctors' orders. In some cases it linked requests to the wrong patient profiles or listed scans that did not match the orders.
Staff who schedule at Rayus have their own theory about the cause. They point to forms that depart from the usual layout, handwriting that is hard to read, and too little checking by people. That is a suspicion, not a finding. Rayus spokesperson Kit Crancer says the broad characterisations do not fairly reflect its safeguards. In his words, the company uses technology to back up human judgment and oversight, not to take its place.
Palantir declined to comment on the record. It has described its software as presenting information, with customers responsible for their data, policies and decisions.
What leaders should ask before the next rollout
Any organisation that lets software draft decisions about people can put these questions to its team and its vendor.
First: who can overrule the output, and how long does it take? Count the steps and the days, not the policy statement.
Second: do we keep the data the system used and the schedule it produced? Can an independent party re-run the decision?
Third: do we track results beyond efficiency? Include swap success, sick calls, turnover and injuries, and report them by unit, not only as an average.
Fourth: when something goes wrong, who is responsible under our contract? Palantir has said customers are responsible for their data, policies and decisions. Does your contract say the same, and does your team know it?
A control you cannot test is a promise. Oversight that leaves no records can be claimed, but it cannot be checked.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Wired.
