Dataset / Healthcare admin

Healthcare admin

Available

Clinic work in EHRs, payer portals, and fax inboxes, recorded live and delivered without patient data.

Specialties: 8+
Systems: 9 EHRs and portals
Human sign-off: Every task
Patient data: None

Types of data

Environments

Simulated EHRs, payer portals, and fax inboxes filled with synthetic patients. Every task has a goal, a starting state, a step limit, and a grader.

For: Reinforcement learning with checkable rewards
Format: Hosted app, task specs in JSON, one grader per task

Trajectories

Complete runs of real workflows: each page, each action, and why. Successes and failures, with the step where a run first went wrong marked.

For: Fine-tuning and step-level rewards
Format: One JSON file per run, with a page snapshot and screenshot per step

Human feedback

What clinic staff approved, edited, or sent back, with the edit and the reason. It shows where an agent should stop and ask.

For: Preference training and knowing when to ask
Format: One JSON line per review: the work, the decision, the edit, the reason

Evals

Referrals, prior authorizations, and appeals that are never published, refreshed every quarter.

For: Clean measurement no model has trained on
Format: Run by us against your model; you get a scored report

One record

A trajectory is one JSON file, with a page snapshot and a screenshot for every step. The patient is always synthetic.

{
  "task": "File a prior authorization for an MRI of the lumbar spine",
  "dataset": "healthcare_admin",
  "workflow": "prior_authorization",
  "patient": "synthetic",
  "steps": [
    {
      "n": 7,
      "page": "Payer portal · Authorization request",
      "snapshot": "steps/0007.json",
      "screenshot": "steps/0007.webp",
      "action": "fill({ label: \"CPT code\" }, \"72148\")",
      "reason": "The order is for CPT 72148, MRI of the lumbar spine without contrast.",
      "progress": "advanced"
    }
  ],
  "sign_off": {
    "result": "approved",
    "edited": false
  },
  "grade": {
    "checks_passed": 9,
    "checks": 9
  }
}

What it covers

Workflows

Referral intake
Fax triage
Eligibility checks
Prior authorization
Denial appeals
Scheduling

Specialties

GI and endoscopy
Cardiology
Retina and ophthalmology
Behavioral health
Primary care
Vein and vascular
Physical therapy
Pain and spine

Workflows seen in

Epic
eClinicalWorks
athenahealth
NextGen
Elation
Tebra
Office Ally
Oracle Health
Availity

How a record is made

DeskShifts runs Eva, an AI employee, inside clinics' own EHRs, payer portals, and fax inboxes. Eva sorts faxes, enters referrals, checks eligibility, files prior authorizations, and appeals denials, and a person at the clinic signs off on her work.

Real work

Eva does a task in the clinic's own system, the way a new hire would, through the browser.
Recorded

Every page, every action, and the reason for it is kept, along with what the clinic approved.
Rebuilt without patients

Names, dates, IDs, and notes are replaced with synthetic patients, and the task is replayed on a simulated portal. Nothing from the clinic's records leaves the clinic.
Graded

The replay is checked against the outcome the clinic approved, so every record has a verified answer.

Privacy

No patient data leaves the clinic. Every record is rebuilt on synthetic patients before it is delivered.
De-identified to the HIPAA standard. Each release is reviewed by an independent expert before it ships.
Clinics opt in by contract, and can withdraw their workflows from future releases.
Simulated portals are look-alikes, built for the workflow, not copies of any vendor's software.

Want a specific type or workflow?

Tell us the workflow, the specialty, the type of data, and how much you need.