BRENNEN

Industries / Health / Healthcare / Nurse staffing

Contracts with hospitals

Contracts with hospitals are the agreements under which agencies supply nurses: the hourly bill rate a hospital pays, the fees and terms around it, and the software and managed programs through which orders are placed and invoices paid.

6

Milestones achieved, from 2013 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2013

    Agencies buy the ordering software

    A vendor management system (VMS) is software through which a hospital sends its orders for contract nurses to many agencies and pays their invoices; a managed services program (MSP) hands one firm the running of all of a hospital's contract staffing. AMN Healthcare bought the VMS ShiftWise on 20 November 2013 and Medefis on 7 January 2015. By AMN's own figures, about 1% of its revenue flowed through MSP relationships in 2008, about a third in 2014 and 48% in 2025, when $3.3 billion of client spending passed through its MSPs and VMS.

    sec.govsec.gov

    Credit

    AMN Healthcare

  2. 2015

    England caps agency prices

    From 1 September 2015 the two regulators of English NHS trusts, Monitor and the NHS Trust Development Authority, set each trust a ceiling on agency nursing spend and required agencies to be hired through approved frameworks. From 23 November 2015 they capped the hourly rates trusts may pay agency staff, aiming to bring them down to 55% above NHS pay rates by April 2016. The National Audit Office found trusts' agency spending fell from a peak of £3.7 billion in 2015-16 to £3.0 billion in 2016-17, faster for nurses than for doctors.

    LessonBy the government's count the NHS was again spending £3 billion a year on agency staff in 2023/24, so caps alone did not end the reliance.

    gov.uknao.org.ukgov.uk

    Credit

    Monitor; NHS Trust Development Authority

  3. 2021

    Texas hires a surge workforce

    In the COVID-19 winter of 2020–21 the Texas Department of State Health Services paid private vendors to send nurses and other staff to hospitals at premium rates. By the Dallas Morning News's count the state was set to spend more than $5.1 billion on this, nearly two-thirds of its $8.1 billion of federal Coronavirus Relief Fund money; at the peak, in early February 2021, three vendors had nearly 14,000 workers deployed, mostly from other states.

    dallasnews.com

    Credit

    Texas Department of State Health Services

  4. 2022

    Congress asks for a probe

    On 24 January 2022 Representatives Peter Welch and Morgan Griffith, in a letter signed by nearly 200 members of Congress from both parties, asked the White House COVID-19 coordinator to have federal competition and consumer protection agencies examine nurse staffing agencies. They cited reports that agencies charged two, three or more times pre-pandemic rates and kept 40% or more. The American Staffing Association answered that about 75% of what agencies charge covers nurses' pay, taxes, insurance and benefits. PolitiFact reported in February 2022 that the letter's authors had no reply from the White House and that no bill to cap travel nurse pay had been proposed.

    morgangriffith.house.govpolitifact.com

    Credit

    Peter Welch; Morgan Griffith

  5. 2022

    Crisis rates rise, then fall

    A bill rate is the hourly price a hospital pays an agency, out of which the agency pays the nurse. AMN Healthcare's revenue rose from $2.22 billion in 2019 to $5.24 billion in 2022 and fell to $2.73 billion in 2025; the company says demand and bill rates "rose sharply and then declined" through the pandemic. Its gross profit stayed near a third of revenue throughout: 33.5% in 2019, 32.7% in 2022 and 28.3% in 2025 (worked out here from its filed figures). By SIA's survey of travel nurse firms, the average travel nurse bill rate fell from $106.63 an hour in 2023 to $89.78 in 2024.

    data.sec.govdata.sec.govsec.govstaffingindustry.com

    Credit

    AMN Healthcare (annual reports); SIA

  6. 2025

    Merger runs out of time

    Aya Healthcare agreed on 3 December 2024, in a deal announced the next day, to buy Cross Country Healthcare for about $615 million in cash. The US Federal Trade Commission asked the two firms for more information on 20 February 2025. The 43-day federal government shutdown then pushed the end of its review to 30 December 2025, past the deal's deadline, and Aya ended the agreement on 4 December 2025, paying Cross Country a $20 million termination fee.

    LessonA combination of two large nurse agencies drew a full antitrust review that outlasted the deal's own deadline.

    sec.govsec.gov

    Credit

    Aya Healthcare; Cross Country Healthcare; US Federal Trade Commission

Watch for

  1. NHS to end agency spending

    The UK government says NHS spending on agency staff fell by almost £1 billion in 2024/25 from £3 billion in 2023/24, after trusts were told to cut agency spending by 30% and bank spending by at least 10%. It pledged to eliminate agency spending altogether and to consider new law if progress stalled by autumn 2025.

    Achieved whenNHS England's audited accounts or a National Audit Office report confirm the fall for 2024/25 and 2025/26.

    gov.uk

    Claimed by

    Department of Health and Social Care and NHS England, June 2025

Open

  1. 01

    Agency invoice error test

    A public set of made-up rate cards and at least 1,000 made-up timesheet and invoice lines for contract nurses, with errors planted at known places: a shift billed at the wrong specialty or shift rate, overtime counted on the wrong base, a holiday rate on the wrong day, orientation hours billed where the contract makes them unbillable, a cancelled shift not credited, the same shift billed twice, and a bill naming a different nurse from the one on the timesheet. Score AI models and a rules-based checker on finding each error.

    The brief
    Credit · unclaimed
  2. 02

    Requests turned into orders

    A public set of 1,000 made-up requests for contract nurses as hospital managers send them, by email, text message, spreadsheet row or phone-call transcript ("need 2 ICU nurses, nights, 13 weeks from 5 January, CRRT experience"), each with an answer key in a fixed order format: unit type, specialty, number of nurses, shift pattern, start date, length, required certifications such as ACLS, minimum experience and any term that changes the bill rate. Code the unit type with the standard list that the National Database of Nursing Quality Indicators (NDNQI) uses and LOINC, the medical coding standard, publishes, so that labels such as "Tele", "PCU" and "Stepdown" land on one code. Score AI models on producing the correct order and on asking a question when a request is unclear instead of guessing.

    The brief
    Credit · unclaimed

Open 01 · the brief

Agency invoice error test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of made-up rate cards and at least 1,000 made-up timesheet and invoice lines for contract nurses, with errors planted at known places: a shift billed at the wrong specialty or shift rate, overtime counted on the wrong base, a holiday rate on the wrong day, orientation hours billed where the contract makes them unbillable, a cancelled shift not credited, the same shift billed twice, and a bill naming a different nurse from the one on the timesheet. Score AI models and a rules-based checker on finding each error.

Why it is still open

Invoice checks still rest on people. New York City's Comptroller found in December 2021 that the city's school health office paid $53.6 million, about 41% of $130 million spent on contract nursing from July 2016 to March 2019, without timesheets showing the work was done, and sometimes paid for a nurse other than the one who worked. Qualivis, a staffing firm, offers hospitals an invoice review done by its own accounting staff. No public test of invoice checking for contract nurses was found.

comptroller.nyc.govtha.org

Built with what exists

Python to generate rate cards, schedules and invoices from rules, rendered as PDF and CSV; Claude, Gemini and Mistral Large 4 (see the Note "A trillion open parameters") to read them; a rules-based checker in Python as the baseline; answer keys in CSV.

  1. First step

    Write one rate card and 100 invoice lines with 15 planted errors, and run two models on them.

  2. Done looks like

    1,000 invoice lines with at least 150 planted errors across 7 types; for each of at least 3 models and the baseline, the share of each error type found, the false-alarm rate, and the dollars of planted overbilling caught, published with the data.

Open 02 · the brief

Requests turned into orders

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of 1,000 made-up requests for contract nurses as hospital managers send them, by email, text message, spreadsheet row or phone-call transcript ("need 2 ICU nurses, nights, 13 weeks from 5 January, CRRT experience"), each with an answer key in a fixed order format: unit type, specialty, number of nurses, shift pattern, start date, length, required certifications such as ACLS, minimum experience and any term that changes the bill rate. Code the unit type with the standard list that the National Database of Nursing Quality Indicators (NDNQI) uses and LOINC, the medical coding standard, publishes, so that labels such as "Tele", "PCU" and "Stepdown" land on one code. Score AI models on producing the correct order and on asking a question when a request is unclear instead of guessing.

Why it is still open

Requests still arrive in every form. TXM Healthcare, a UK staffing firm, gets 7,000 to 8,000 emails a week, some with photos of handwritten notes, and uses Globus.ai to sort them, but neither publishes an accuracy figure. AMN Healthcare says its newest VMS uses automation without giving accuracy figures, and a 2023 paper on a travel nurse recommender read job feeds into a database that was not released. No public test of turning staffing requests into orders, and no published mapping of order labels to the standard unit list, was found.

globus.aisec.govarxiv.orgloinc.org

Built with what exists

Requests written by two people with hospital staffing-office experience, varied by an AI model and checked by hand; LOINC's answer list LL599-2 of unit types; Claude, Gemini and Meta's Muse Ember; ElevenLabs Scribe v3 to transcribe the call recordings (see the Note "Fewer word errors"); a JSON order schema and a scoring script.

  1. First step

    Write 50 requests, 10 of them deliberately unclear, agree their answer keys, and score one model.

  2. Done looks like

    1,000 requests with answer keys on which two people agree for at least 95% of fields; for each of at least 3 models, field-level accuracy, the share of orders entirely correct, and the share of unclear requests where the model asked rather than guessed; the label-to-unit-type table published with the data.