HCA’s Timpani Dispute Turns AI Nurse Scheduling Into a Test of the Audit Trail

On October 2, WIRED reported that nurses at HCA Healthcare, the largest hospital chain in the US, say Timpani, an AI scheduling tool the company built with Palantir, frequently ignores their preferences and routinely leaves shifts short of nurses, or short of experienced ones. HCA says nursing leaders, not the tool, make the final scheduling decisions. Its own published figures describe a tool that cuts managers’ scheduling time and rarely leaves a shift staffed entirely by beginners.

Both sides have numbers, and the numbers do not collide, because they measure different things at different levels. This piece lines up what HCA has published, what nurses and their unions describe, and a federal complaint filed in July by a former HCA data science manager. That complaint alleges that the inputs behind Timpani’s daily staffing predictions were overwritten on a rolling basis, which would leave neither side able to check its claims against the original decisions. Everything in the complaint is an allegation, and HCA has not yet answered it in court.

This reflects the public record as of October 2, 2026. We did not have access to Timpani and did not test it, so every statement about how it behaves is attributed to the person or document that made it.

What Timpani is and how widely it runs

HCA describes itself as operating 189 hospitals and about 2,500 other sites of care in 19 US states and the UK, with more than 300,000 employees, 100,000 of them nurses. Michael Schlosser, its chief transformation officer, describes Timpani in a paper first published in the June 2026 issue of Management in Healthcare and republished on HCA’s site on July 21. In it, the platform is “live at more than 130 hospitals, supporting over 1,200 nursing departments.”

The tool does two jobs. WIRED’s Paresh Dave reports that Timpani was built on Palantir’s Foundry platform and “uses algorithmic forecasts of patient volumes and staffing needs to automatically build nursing schedules.” The federal complaint described below calls it “an AI-enabled staffing and census forecasting system developed by HCA in collaboration with Palantir” and says nurse staffing levels are determined in part by its daily census predictions. So one layer predicts how many patients a unit will have, and another turns that into schedules that match nurses’ skills and preferences. The dispute over schedules and the allegation over records sit at different layers, and that matters later.

The vendor’s role is part of the argument too. Palantir reported $1.94 billion in second-quarter revenue, up 93 percent from a year earlier. According to WIRED, a Palantir spokesperson declined to comment by name, and the company has described its software as presenting information in a more helpful manner and emphasized that customers are responsible for the data, policies, and decisionmaking. We covered a separate Palantir procurement dispute in August.

The numbers HCA has published

Schlosser’s paper is the fullest public account of what HCA measures. Four figures matter for the dispute, and each leaves something out.

A floor is not a ratio

The skill-mix figure deserves a closer read. The paper says Timpani sorts nurses into “beginner, proficient or expert” and reports that more than 98 percent of shifts include “a mix of skill and experience levels,” with under 2 percent “staffed entirely by beginners.” The only failure it names is an all-beginner shift. Amber Retzloff, the Florida critical care nurse and National Nurses United leader whom WIRED quotes, describes recent shifts where all four of her colleagues were junior and she was the only senior nurse, and says she had to delay care for the sickest patients to guide the novices through milder cases. A shift like that would not trip a test whose only failure is an all-beginner roster. That does not make her account right or HCA’s figure wrong. It means the published figure is a floor, and a floor cannot show how thin the experienced end of a shift is.

Who the benefit is measured for

The paper is open about where the benefit lands. During beta testing, it says, “it became clear that the true unit of change was not frontline staff alone, but nurse managers and directors,” and its lead efficiency number is manager time. The Timpani section reports no figure for nurse time spent on shift swaps and appeals, which the nurses in WIRED’s story say has grown. It also reports none for the outcomes its own framework lists as worth tracking, such as “Positive changes in team workflows and satisfaction.”

The two accounts also start from different baselines. The paper says schedules were previously “built largely around employee self-scheduling,” which left gaps that had to be filled with “premium resources” or contract labor, while the nurses WIRED quotes remember managers building schedules by hand and consulting them before overriding a request. Which baseline a unit came from changes what a fair before-and-after comparison looks like, and neither account says how many units started from which.

The 1 percent figure comes from HCA through WIRED rather than from the paper. WIRED’s sentence is that Timpani “schedules nurses for 1 percent of their requested days off, or what are known as red days, according to HCA.” Lee Barker, an HCA nurse in Missouri, told WIRED that the number might look low but that being scheduled on a red day was previously unheard of at their facility.

What nurses and their unions describe

WIRED spoke to Retzloff and to five nurses at other locations. They allege that Timpani routinely schedules too few nurses or too few experienced ones, particularly on Sundays, and frequently ignores stated preferences for spacing out shifts, working nights and weekends, or taking specific days off. Retzloff says she requested 50 specific 12-hour shifts over four months and, by her count, was assigned different shifts more than half the time, often on back-to-back-to-back days. Before Timpani, WIRED reports, managers handled scheduling manually and generally consulted nurses well in advance when overriding a request.

On appeals, WIRED reports that some local managers could initially edit Timpani’s schedules, but nurses must now flag problem shifts to a centralized team at HCA headquarters in Nashville, which reviews appeals without direct communication or room for explanation, according to nurses. If an appeal is denied, nurses try to swap shifts, and Barker and Retzloff claim the chance of success is only about 50 percent. A union of about 500 nurses at the psychiatric center where Barker works has filed a grievance arguing HCA is failing to sufficiently honor scheduling preferences, and WIRED says the matter is likely headed to arbitration. National Nurses United, which represents about 10,000 of HCA’s 100,000 nurses, says HCA has rejected its requests for information about the technology.

The union’s August 27 release, which reported demonstrations in eight cities, describes Timpani as “removing hospitals’ local managers from staffing decisions and replacing them with an anonymous office in Nashville.” The union is an interested party, and its campaign against Palantir is broader than scheduling: the same release cites the company’s work for the Defense and Homeland Security departments. WIRED notes that Retzloff and Barker are union leaders. The same caution applies to HCA, whose paper was written by its chief transformation officer and whose acknowledgements say the research was supported in whole or in part by HCA.

HCA spokesperson Harlow Sumerford told WIRED that nursing leaders, not Timpani, make final scheduling decisions, and called assertions that the tool “is designed to reduce staffing at the expense of patient care” misrepresentative. He said HCA’s goal is “to have the right caregivers, with the right skills and experience, available to meet the needs of our patients while also considering our colleagues’ scheduling preferences,” and that the company continues to improve Timpani based on nurses’ feedback.

A 1 percent rate per request is not a 1 percent chance per nurse

Barker told WIRED that less than a year after Timpani arrived, essentially every nurse at the facility had been assigned to work a day they were supposed to be off. Set that beside HCA’s 1 percent and the two look incompatible. They may not be, because HCA’s figure is a rate per requested day off, while a nurse experiences the chance of at least one violation across every day they request. If each request had an independent 1 percent chance of being violated, the share of nurses with at least one violation after a given number of requests would be one minus 0.99 raised to that number:

This is arithmetic, not evidence about HCA: the number of red-day requests per nurse is not public. A nurse working three 12-hour shifts a week has about 209 days off a year (365 minus 156), so even a nurse who requested all of them would have an 87.8 percent chance of at least one violation at a 1 percent rate. Clustering of violations in particular units or schedule periods would lower these shares, which widens the gap with “essentially every nurse” rather than closing it. Taken at face value, Barker’s account therefore points to a local rate well above the company average, a broader category than formal red-day requests, a different way of counting than HCA’s, or “essentially every” being looser than it sounds. The published average cannot tell us which. What would is the share of nurses with at least one violated red day per quarter, reported by facility.

A records allegation, not a scheduling one

What the complaint says

On July 15, Angelique Russell, who held the title Manager, Data Science in HCA’s Digital Technology and Innovation department, sued HCA in the US District Court for the Middle District of Tennessee (Russell v. HCA Healthcare, Inc., No. 3:26-cv-00983). She alleges she was fired on April 3 in retaliation for reporting that Timpani “could neither be audited nor reproduced.” The claims are retaliatory discharge under the Tennessee Public Protection Act, common-law retaliatory discharge, and promissory estoppel.

The technical core is one paragraph. The complaint says the “feature inputs used to generate each day’s staffing and census predictions were overwritten on a rolling basis and became permanently unavailable once Google BigQuery’s Time Machine retention window expired,” so HCA could not reproduce historical predictions or determine, after the fact, what census inputs had driven its staffing recommendations on a given day. It also alleges that Timpani’s lead data scientist acknowledged that not retaining historical feature data was intentional during development, and that a director acknowledged that snapshot copies of the relevant tables had been kept for audit purposes but were later “discontinued due to cost.” Those are the complaint’s descriptions of what HCA personnel said. WIRED reports that HCA has yet to formally respond in court.

The one incident the complaint describes in detail is at HCA Florida Citrus Hospital, where it says retroactive manual corrections to census data “modified prior feature values and caused cascading inaccurate census under-predictions that impaired a nurse manager’s ability to complete staffing schedules.” The affected unit, it says, was “at risk of being scheduled with fewer nurses than the patient population required.” The complaint frames this as risk, and we did not find an allegation of a specific patient injury in it.

WIRED paraphrases the allegation as HCA “routinely deleting data the tool uses to create schedules.” The complaint’s language is narrower and more technical. It concerns the feature inputs to daily census and staffing predictions, and it does not allege that nurses’ stated preferences or finished schedules were deleted.

The timeline it alleges

Where the case stands

The CourtListener docket shows an August 24 order setting HCA’s answer for September 25. Later entries dated September 25 and September 30 are labeled “Extension of Time to File Answer,” and entries dated September 28 and October 1 are labeled “Order on Motion for Extension of Time to Answer.” The docket we checked on October 2 shows no answer. An initial case management conference is set for October 7 before Magistrate Judge Luke A. Evans.

What HCA’s own policy and paper promise

HCA’s Responsible AI Policy, EC.031, is public and took effect on October 1, 2024. It says AI solutions will be “subject to periodic audit for the duration of the solution lifespan until deprecation.” Its requirements include “Accountable and consistent monitoring of solution performance and Outcome generation” and documenting “key assumptions and decisions, including any applicable or otherwise appropriate version control,” and it tells colleagues to “Promptly notify the Director of Responsible AI” about “anomalies, a decline or material deviation in accuracy of Outputs.” The policy text we read sets no retention period for model inputs, and it also requires that data usage, storage, disclosure and destruction “incorporate controls to protect individuals’ privacy and rights (e.g., data minimization practices),” so retention is a trade-off against privacy rather than a free choice. Whether a periodic audit of a prediction system can work without the inputs behind each prediction is the question the complaint puts to the court.

Schlosser’s paper offers other health systems a checklist that includes “Model lifecycle infrastructure (versioning, monitoring, audit logs),” “Post-deployment audits and safety checks,” “Human-in-the-loop design with override capabilities,” and “Performance measures beyond accuracy (eg safety, equity, usability).” These are HCA’s own standards. They are useful here because they name the artifacts that would settle the dispute.

The complaint also invokes federal rules. The HIPAA Security Rule’s audit-controls standard, 45 CFR 164.312(b), requires covered entities to implement mechanisms that “record and examine activity in information systems that contain or use electronic protected health information.” The Medicare conditions of participation for hospitals, 42 CFR 482.23(b), require “adequate numbers of licensed registered nurses, licensed practical (vocational) nurses, and other personnel to provide nursing care to all patients as needed.” Whether and how either applies to a forecasting system’s inputs is a legal question for the court, and we express no view on it.

What a defensible record looks like

Whatever the court decides, the engineering question is general: when a model’s output sets who works when, what must be kept to reconstruct a decision later?

Time travel is a recovery window, not an archive

Google’s BigQuery documentation says time travel covers “the past seven days by default,” and the window can be set from a minimum of two days to a maximum of seven. Deleted data then enters a fail-safe period of seven more days that users cannot query, and “Once the fail-safe period has passed, Cloud Customer Care can’t recover any of your deleted data.” The complaint’s “Time Machine” presumably refers to this feature, which Google’s documentation calls time travel. If time travel is the only history, the horizon for reconstructing a given day’s inputs is at most a week.

The documented alternative is a table snapshot, a read-only copy of a table at a point in time that can carry an expiration. Google says “BigQuery only stores bytes that are different between a snapshot and its base table, so a table snapshot typically uses less storage than a full copy of the table.” The marginal cost of a snapshot therefore depends on how much of the table changes between snapshots. The complaint attributes the end of Timpani’s snapshots to cost, and whether that was a reasonable trade for a table of that size is something the public record cannot answer.

Keep the decision, not just the data

These are our suggestions, not drawn from HCA’s documents. For every scheduling or forecasting run, keep:

• the input snapshot, or a content hash plus a pointer to an immutable copy;
• the code version, model version and parameters;
• the output as produced, before any human edit;
• every override, with who made it, when, and the stated reason;
• corrections as new versions rather than in-place edits, since the Citrus Hospital allegation is that retroactive corrections modified prior feature values.

Publish the distribution, not only the average

Again our suggestions, not HCA’s metrics or a standard. Each question below needs a measure at the level where people experience the system.

If you run a system like this

• Report the share of people affected, not only the per-event rate. A rate per request cannot be turned into a statement about how many people are touched.
• Say whether a figure is a floor or a ratio. “No all-beginner shifts” and “a balanced mix” are different promises.
• Make the retention window match the audit promise. If policy says solutions are audited for their whole lifespan, the inputs and overrides must outlive the platform’s recovery features, or the policy should say plainly what is kept and what is minimized for privacy.
• Record who changed what. A human in the loop is only as strong as the log of what that person saw and changed. We looked at the same gap in the Pentagon’s hallucinated cargo manifest incident.

Who holds the record

Three parties each point to another. Palantir, according to WIRED, says customers are responsible for the data, policies, and decisionmaking. HCA says nursing leaders make the final decisions. Nurses say appeals now go to a central team in Nashville. A decision that three parties can each attribute to another is one that none of them can reconstruct without the records, which is why the retention allegation matters more than any single average. It is also why an organization grading its own system, as in OpenAI’s research acceleration report, ends up choosing the figures that frame the argument. The first problem Russell alleges she reported, training a model on means that included its validation and holdout groups, is a form of data leakage; our target leakage linter tutorial covers a sibling form, columns derived from the target.

What to watch

HCA’s answer to the complaint, and whether it addresses the retention allegation, is the first signal, followed by the October 7 case management conference and the arbitration over the grievance at the psychiatric center. Beyond the courtroom, watch whether HCA publishes per-nurse or per-facility figures, and whether the union’s information requests are answered. Timpani is live at more than 130 hospitals, so what HCA can produce when asked to reconstruct a specific day is the test that matters.