---
title: "Healthcare Administration – The Physician AI Handbook"
source: "https://physicianaihandbook.com/implementation/healthcare-administration.html"
author: "Bryan Tegomoh, MD, MPH"
last_updated: "2026-11-13"
doi: "10.5281/zenodo.18251404"
license: CC BY 4.0
generated-by: handbook-markdown-export
---

# Healthcare Administration

AI for healthcare administration should be judged by whether it returns measured time or money to clinical work; in the two deployment studies of drafted patient messages reviewed here, neither detected a time saving. Administrative work in U.S. healthcare includes registration, insurance checks, prior authorization, billing, patient messages, scheduling, referrals, staffing, and credentialing. At one academic health system, billing and insurance-related work cost an estimated \$20.49 for a primary care visit and \$215.10 for an inpatient surgical procedure, even with a certified electronic health record ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)). Two early studies of large language model drafts for patient portal messages did not detect a reduction in the time clinicians spent per reply ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201); [Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)).

Learning Objectives

After completing this chapter, clinicians and health system leaders should be able to:

- Describe what billing and insurance-related work cost at one academic health system, and how much of it physicians did
- Interpret the two prospective studies of AI-drafted replies to patient portal messages, including what each measured and how the authors explain the absence of a time saving
- Name the practice-side prior authorization steps that AI tools target, and state what the evidence does and does not show
- Name the measures that show whether an administrative AI tool returned time to clinicians
- List the questions to ask before a practice or health system adopts an administrative AI tool

Chapter Summary (TL;DR)

**Cost baseline:** In a costing study at one academic health system, billing and insurance-related work took an estimated 13 minutes and \$20.49 for a primary care visit and 100 minutes and \$215.10 for an inpatient surgical procedure; the figures were modeled from interviews in 2016 and 2017 ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)).

**Patient messages:** In a 5-week prospective single-group quality improvement study of 162 clinicians at Stanford, the mean draft utilization rate was 20% of replies with a draft available, and audit-log reply time did not change; among the 73 clinicians who completed both surveys, task load and work exhaustion scores fell ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)). In a randomized waiting-list quality improvement study of 52 primary care physicians at UC San Diego, access to drafts was associated with 21.8% longer read time, no significant change in reply time, and 17.9% longer replies ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)).

**Prior authorization:** In a survey by a physician advocacy organization, the American Medical Association, 1,000 physicians who complete prior authorizations in a typical week reported that their practices complete an average of 40 per physician each week ([AMA, 2025 survey](https://www.ama-assn.org/system/files/prior-authorization-survey.pdf)). In a preprint benchmark in simulated systems, the best AI agent completed 38.3% of prior authorization tasks from start to finish ([Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint).

**Bottom line:** Measure an administrative tool with audit-log time, error rates, and what staff and patients experience, against a comparison group. Audit-log time is itself an approximation, and a drafting tool can lower reported task load without saving measured time.

**Scope:** The evidence below comes from six sources and is not a systematic review. Scheduling, referrals, staffing, and credentialing have a measurement guide but no cited studies.

### Evidence at a Glance

| Use | Study | Design | Setting | What was measured | Main result | Main caveat |
|----|----|----|----|----|----|----|
| Cost of billing work | [Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148) | Costing study (time-driven activity-based costing) | One academic health system | Time and cost of billing and insurance-related work per encounter | \$20.49 per primary care visit to \$215.10 per inpatient surgical procedure | Modeled from interviews, not from actual bills |
| Drafted portal replies | [Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201) | Prospective single-group quality improvement study | Stanford, two divisions, 162 clinicians | Draft use, time, surveys | Mean draft utilization 20% of available drafts; no change in time; lower task load and work exhaustion scores among survey respondents | No control group; 5 weeks |
| Drafted portal replies | [Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565) | Randomized waiting-list quality improvement study | UC San Diego, primary care, 52 physicians plus 70 controls | Read time, reply time, reply length | Read time +21.8%; reply time no significant change; replies 17.9% longer | Volunteers at one system; no patient outcomes |
| Prior authorization burden | [AMA, 2025 survey](https://www.ama-assn.org/system/files/prior-authorization-survey.pdf) | Survey by a physician advocacy organization; not peer reviewed | 1,000 U.S. physicians who complete prior authorizations weekly | Self-reported workload and harms | 40 prior authorizations per physician per week | Self-report; the AMA advocates prior authorization reform |
| Administrative agents | [Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint | Preprint benchmark | 135 tasks in simulated systems | Whole-task completion | 38.3% of prior authorization tasks for the best agent | Simulated portals and synthetic patients |

## Introduction

Administrative work is the part of healthcare that moves information and money around the visit: verifying coverage, obtaining approvals, answering patient messages, filling schedules, routing referrals, and keeping staff rosters and credentials current. Some of it lands on physicians: in the costing study below, physicians did 3 of the 13 billing minutes for a primary care visit and 15 of the 100 for an inpatient surgical procedure ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)). Vendors sell AI tools for many of these tasks. When administrative leaders make the purchase, the clinicians whose time the tool is meant to save should help decide how success is measured.

Related topics are covered in other chapters. Operating room scheduling, no-show prediction, patient flow, and medical coding are covered in [Hospital Operations and Management](../implementation/hospital-operations.html). Ambient scribes and the billing rules for AI-drafted notes are covered in [AI-Assisted Clinical Documentation](../practical/documentation.html#sec-documentation-billing). Payer use of algorithms in coverage decisions and the federal prior authorization rules are covered in [Healthcare Policy and AI Governance](../future/policy.html#sec-payer-algorithms).

| Term | Meaning |
|----|----|
| Billing and insurance-related work | Tasks done to get paid: registration, eligibility checks, prior authorization, coding, claims, and follow-up on denials |
| Time-driven activity-based costing | A costing method that maps each step of a process, estimates the staff time it takes, and multiplies by the cost of that time |
| Prior authorization | A health plan’s requirement for approval before a drug or service qualifies for payment |
| Patient portal message | A message a patient sends through the health system’s online portal, answered from the clinician’s electronic inbox |
| Audit log | The electronic health record’s time-stamped record of each user action, used to measure time spent on tasks |
| Waiting-list design | A study in which one randomized group starts the intervention now and another starts later, so the later group serves as a comparison in the meantime |
| Task load | Self-reported cognitive load of work, here a 4-item physician task load score adapted from the NASA Task Load Index (0 to 100; lower is less load) |
| Work exhaustion | A 4-item burnout measure from the Stanford Professional Fulfillment Index (0 to 4; lower is less exhaustion) |
| Net promoter score | The percentage of respondents who would recommend a tool (9 or 10 out of 10) minus the percentage who would not (0 to 6); ranges from −100 to 100 |

## The Administrative Cost Baseline

Before buying a tool to cut administrative cost, a health system needs a measured baseline of what each task costs and who does it. One academic health system in North Carolina measured this for billing and insurance-related work. Researchers interviewed 27 health system administrators and 34 physicians in 2016 and 2017, built a process map of an insurance claim’s path through the revenue cycle, and estimated the time and cost of each step. The work took an estimated 13 minutes and cost \$20.49 for a primary care visit, 32 minutes and \$61.54 for a discharged emergency department visit, 73 minutes and \$124.26 for a general medicine inpatient stay, and 100 minutes and \$215.10 for an inpatient surgical procedure. Professional billing costs ranged from 3.1% of professional revenue for an inpatient surgical procedure to 25.2% for a discharged emergency department visit, and were 14.5% for a primary care visit. Physicians’ own share, estimated from a survey of a convenience sample of physicians, was a median of 3 minutes per primary care or emergency visit and 15 minutes per surgical procedure, for tasks such as completing billing fields, answering coding queries, and obtaining prior authorization. At that rate, billing for one primary care physician cost an estimated \$99,581 a year, assuming 4,860 visits. Outside primary care, costs after the encounter were mostly attributable to medical coding ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)).

The authors found no significantly wasteful steps in a billing organization of more than 1,500 full-time-equivalent staff. They attributed the cost to a large degree to differing payment requirements across many payers and health plans, and concluded that administrative costs remained high even with a certified electronic health record. The estimates come from a model built on interviews rather than from actual bills, and the authors describe them as conservative. They exclude what payers spend to process claims. The system’s billing was centralized in one organization, and smaller or more fragmented organizations probably incur higher costs ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)).

At the national level, an NBER working paper, which has not been peer reviewed, estimates that wider AI adoption could save 5% to 10% of U.S. healthcare spending, roughly \$200 billion to \$360 billion a year in 2019 dollars. The estimate rests on use cases built on technologies available when it was written, which the authors judged attainable within five years; it excludes generative AI and one-time implementation costs. Administrative costs supply about 35% of the projected savings. The paper states that administrative costs are estimated to account for nearly 25% of U.S. healthcare spending and lists workflow integration of AI insights among the challenges of adoption. Three of the four authors are at McKinsey & Company ([Sahni et al., NBER, 2023](https://www.nber.org/system/files/working_papers/w30857/w30857.pdf)). The estimate is a projection, not a measured result, and it does not cover the drafting tools and agents reviewed in this chapter.

The costing study had no data from before the electronic health record, so it cannot show what the record changed. It does show that substantial billing costs remained two to three years after a certified record was adopted. An AI tool that speeds one step, such as drafting an appeal letter, leaves the other steps in place.

## Patient Portal Messages and Inbox Drafting

Replying to patient portal messages is clinical communication. It is covered here because health systems adopt drafting tools for the same reason they adopt other administrative tools, to reduce desk work. Epic built drafting into its inbox, and two academic health systems published prospective evaluations in 2024.

### A single-group study at Stanford

Stanford Health Care turned on drafts for all attending physicians, advanced practice practitioners, clinic nurses, and clinical pharmacists in primary care and in gastroenterology and hepatology for 35 days in July and August 2023. GPT-3.5 Turbo sorted each eligible message into one of four categories (general, results, medications, and paperwork), and GPT-4 wrote a draft using the message, selected structured data such as age and allergies, and the last clinic note. Messages in a language other than English, with an attachment, or from a proxy were excluded. Of 197 enrolled clinicians, 162 were analyzed, and 73 completed both surveys ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)).

On average, a clinician started from the draft for 20% of the replies that had a draft available. Audit-log reply time, read time, and write time did not change compared with the 35 days before the pilot. Mean reply action time was 259.7 seconds before the pilot and 267.9 seconds during it. Clinicians’ own impression differed: on a 5-point scale where lower scores mean more agreement, they rated the tool’s ability to save time at 2.23 before the pilot and 2.29 after, and 12 of 13 free-text comments about time were positive. The authors note that perceptions of time may differ from time captured in electronic health record metadata.

Among the 73 clinicians who completed both surveys (71 for work exhaustion), mean task load fell from 61.31 to 47.26 on a scale of 0 to 100 (paired difference, −13.87; 95% CI, −17.38 to −9.50), and mean work exhaustion fell from 1.95 to 1.62 on a scale of 0 to 4 (paired difference, −0.33; 95% CI, −0.50 to −0.17). Of 12,844 messages that received a reply, 9,621 (75%) had a draft available; 20% were excluded by design and 5% had no draft because of technical limits such as computing capacity. The overall net promoter score was 10 among 73 respondents. Among physicians and advanced practice practitioners it was 13 in primary care and −19 in gastroenterology and hepatology; scores ranged from −60 among 5 primary care nurses to 71 among 7 primary care clinical pharmacists. Of 143 coded free-text comments, 77 were positive and 57 negative; voice and tone drew the most negative comments (14), followed by content relevance and effect on workflow (9 each) and length (8), and 5 negative comments concerned accuracy ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)).

The study had no control group and ran for 5 weeks, so novelty and the knowledge of being studied may have improved the survey scores; the authors call for controlled trials at multiple sites. They also caution that electronic health record metadata only approximates time, and they reported no adverse safety signals over the 5 weeks. Every draft generated with GPT-4 has a cost, which the authors say must be justified with clinical practice data when multiplied across millions of messages; because clinicians started from about 1 in 5 available drafts, most generated drafts were never used as the starting point for a reply ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)).

### A randomized waiting-list study at UC San Diego

UC San Diego Health also tested drafting built into Epic, in primary care from June to August 2023. Fifty-two attending physicians who volunteered were randomized to start immediately (25) or 21 days later (27), and 70 physicians in the same departments who did not volunteer served as a comparison group over the same period. Drafts covered four message types: refills, results, paperwork, and general questions. By default, a reply started from a draft included a statement that part of the message was generated automatically and was reviewed and edited by the physician. The researchers had hypothesized that drafts would reduce time spent reading and replying, and sized the study to detect a 20% reduction ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)).

Across 10,679 replies, access to drafts was associated with a 21.8% increase in read time (95% CI, 5.2% to 41.0%), a −5.9% change in reply time that was not significant (95% CI, −16.6% to 6.2%), and replies that were 17.9% longer (95% CI, 10.1% to 26.2%). Median read time in the immediate group was 26 seconds at baseline and 31 seconds at 3 and 6 weeks; in the comparison group it was 21, 22, and 23 seconds. The confidence interval for reply time excludes a reduction as large as 20% but not a smaller one. The authors suggest the longer read time may reflect the need to read both the patient’s message and the draft, and might reflect effort spent scrutinizing drafts for fabricated content, which the authors call hallucinations. In one example the draft suggested an x-ray or physical therapy referral, clinical advice the authors describe as outside its scope; the physician kept only its first sentence. The authors write that the tool’s current performance suggests human input is still essential. In survey comments, some physicians valued the drafts as an empathic starting point, and others found them wordy or too ready to recommend an appointment. The authors suggest the added length may be personal touches that patients value, which the study did not test. The study covered one academic health system’s early experience in primary care and did not measure patient experience ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)).

### Reading the two studies together

Neither early deployment detected faster replies. At Stanford, task load and work exhaustion scores improved among survey respondents, with no control group. At UC San Diego, read time and reply length increased. Before adopting drafting, a health system should name its goal: less time, lower task load, or better replies. Neither study measured reply quality or patient experience, so a longer reply is not shown to be a better one. Each goal needs its own measure. Audit logs measure time, validated surveys measure burden, and review of a sample of sent replies measures accuracy and whether a draft gave clinical advice that the clinician did not intend. Both studies used 2023 models and one vendor’s integration. Later versions may perform differently and should be measured the same way.

## Prior Authorization From the Practice Side

Prior authorization requires a practice to obtain a health plan’s approval before a drug or service qualifies for payment. The American Medical Association surveyed 1,000 practicing U.S. physicians from a Medscape panel in December 2025, limited to those who provide at least 20 hours of patient care a week and complete prior authorizations in a typical week. Practices reported completing an average of 40 prior authorizations per physician each week, and physicians and their staff reported spending 13 hours a week on them; 40% said they had staff who worked only on prior authorization. Ninety-five percent reported care delays, 79% said prior authorization at least sometimes leads patients to abandon treatment, and 26% reported a serious adverse event. Physicians reported the phone as the most common method for medical-service requests, 24% said their electronic health record offered electronic prior authorization for prescription drugs, and 27% of those who complete drug requests said the drug prior authorization information in their record or e-prescribing system was rarely or never accurate. Only 21% said they always appeal an adverse decision; reasons for not appealing included not expecting success (59%), too little staff time (52%), and care that could not wait (49%). Sixty percent said they were concerned that AI increases, or will increase, prior authorization denial rates; the survey summary does not say whose use of AI the question meant ([AMA, 2025 survey](https://www.ama-assn.org/system/files/prior-authorization-survey.pdf)). The AMA advocates prior authorization reform, the survey is not peer reviewed, and the figures are physicians’ own reports rather than measured counts, time, or outcomes.

Practice-side tools target three steps: deciding whether a request is needed, assembling the documentation that maps the patient’s findings to the plan’s criteria, and drafting appeals. None of the sources reviewed here measured any of them in a live practice. A drafting tool addresses only one of the reasons physicians gave for not appealing, too little staff time.

In a preprint benchmark of 135 administrative tasks in a simulated electronic health record, two payer portals, and a fax system, the best agent, Claude Opus 4.6 run as a computer-use agent (one that works through screens as a person would), completed 38.3% of prior authorization tasks from start to finish. The agent with the highest step-level score passed 82.8% of subtasks but completed 26.7% of whole tasks, and the environments used synthetic patients, so these rates describe a simulation rather than live payer portals. The benchmark is reviewed in [Evaluating AI Agents in Healthcare](../implementation/ai-agent-evaluation.html) ([Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint).

Even if an agent submitted requests faster, the plan would still set the criteria, the review, and the response time. No source reviewed here measured whether practice-side AI changes approvals or time to decision. Federal rules for plan response times and electronic prior authorization interfaces, and the rule that a coverage decision must rest on the individual patient’s circumstances, are set out in [Healthcare Policy and AI Governance](../future/policy.html#sec-payer-algorithms). For a practice, the measures that matter are requests per week, staff hours, time to decision, denials, and overturned denials, compared with the period before the tool.

## Scheduling, Referrals, Staffing, and Credentialing

AI tools are also sold for phone lines and scheduling, referral intake, staff scheduling, and credentialing paperwork. No studies of these uses are cited in this version; the table below is a measurement guide, not a summary of evidence. For each use, ask for the measures below before and after go-live, from a controlled comparison where possible. The evidence on predicting missed appointments, and on acting on those predictions, is reviewed in [Hospital Operations and Management](../implementation/hospital-operations.html#sec-hospital-ops-no-shows).

| Use | What the tool does | Measures to track | What could get worse |
|----|----|----|----|
| Phone lines and scheduling | Answers calls, books and reschedules visits | Calls abandoned, time to answer, days to the next appointment, booking errors | Patients who cannot reach a person; urgent symptoms routed as routine |
| Referral intake | Reads faxed or electronic referrals and creates orders | Days from referral to scheduled visit, referrals lost or returned, data entry errors | A misread referral that delays care |
| Staff scheduling | Builds nurse and clinician rosters from forecast demand | Unfilled shifts, overtime, agency use, staff-reported fairness | Rosters that meet a forecast but not the actual census |
| Credentialing | Collects and checks documents for privileges and payer enrollment | Days to complete files, payer enrollment delays, errors found on audit | An expired license or sanction missed |

For each use, the comparison that counts is the same measure before and after, or against a unit that did not use the tool, over the same period. Questions about where patient data goes are covered in [Privacy, HIPAA, and Patient Data Security](../implementation/privacy.html).

## Evaluating an Administrative AI Tool

The questions below adapt [Evaluating AI Clinical Decision Support Systems](../implementation/evaluation.html) to administrative tools. The studies are named by first author: Tseng ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)), Garcia ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)), and Tai-Seale ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)).

| Question | A good answer | Example in this chapter |
|----|----|----|
| 1\. Whose time or what cost is the tool meant to reduce? | A named role, task, and baseline measure | Billing work costs were measured per encounter type (Tseng) |
| 2\. How will time be measured? | Audit logs with a stated definition of each time measure, and surveys reported separately | Clinicians rated the tool as time-saving while audit-log time did not change (Garcia) |
| 3\. Against what comparison? | A randomized, waiting-list, or concurrent control group | Waiting-list randomization with nonparticipant controls (Tai-Seale) |
| 4\. What could get worse? | Read time, error rates, patient complaints, or denials tracked alongside | Read time was 21.8% higher with access to drafts (95% CI, 5.2% to 41.0%) (Tai-Seale) |
| 5\. Who checks the output, and how is that checked? | A named reviewer and a sampled audit of what was sent or submitted | A draft offered clinical advice the physician removed (Tai-Seale) |
| 6\. What does each use cost? | Cost per draft generated and per draft used | Each draft had a cost; clinicians started from about 1 in 5 (Garcia) |
| 7\. What do patients see? | A disclosure policy and a way to reach a person | A default statement that part of the reply was generated automatically (Tai-Seale) |
| 8\. Which patients or messages get nothing? | A list of exclusions and the share of work they represent | Non-English, proxy, and attachment messages were excluded; 1 in 4 replied messages had no draft (Garcia) |
| 9\. How long is the pilot, and what result would stop it? | A period long enough to outlast novelty, and a threshold set in advance | A 5-week pilot with no control group (Garcia); a study sized to detect a 20% time reduction (Tai-Seale) |
| 10\. Where does patient data go? | An integration that meets HIPAA requirements under a signed agreement | The Stanford tool was described as HIPAA-compliant and integrated in the electronic health record (Garcia) |

## Check Your Understanding

Scenario 1: Turning On Drafted Replies

A health system plans to turn on AI-drafted replies to patient portal messages for all primary care clinicians, citing a pilot in which clinicians said the tool saved them time.

**Question:** What should the medical director ask before scaling it?

**Answer:** Ask whether time was measured in audit logs or only by survey. In a randomized waiting-list study, drafts were associated with longer read time, no significant change in reply time, and longer replies ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)). In a single-group study, burden scores improved while measured time did not change ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)). Decide whether the goal is time, burden, or reply quality, and measure that goal against a comparison group. In that study, clinicians rated the tool as time-saving although audit-log time did not change. Audit a sample of sent replies for clinical advice that the clinician did not intend, and track results by role and specialty, since net promoter scores in the Stanford study ran from −60 among primary care nurses to 50 among gastroenterology and hepatology nurses.

Scenario 2: A Prior Authorization Agent

A vendor offers an AI agent that fills in payer portal forms for prior authorization and promises to cut staff hours in half.

**Question:** What evidence should the practice ask for?

**Answer:** Ask for whole-task completion rates on live portals, not step-level accuracy. In a preprint benchmark in simulated systems, the best agent completed 38.3% of prior authorization tasks from start to finish, and the agent with the highest step-level score passed 82.8% of subtasks but completed 26.7% of whole tasks ([Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint). Ask who checks each submission before it is sent, how errors are caught, and what happens to denials and time to decision after go-live. Measure staff hours against the practice’s own baseline.

Scenario 3: A Vendor’s Savings Claim

A vendor says its billing tool will pay for itself in a primary care practice.

**Question:** How can the practice test the claim before a pilot?

**Answer:** Start from a measured baseline. At one academic health system, billing and insurance-related work for a primary care visit cost an estimated \$20.49, of which \$6.36 was physician time ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)). A tool cannot save more than the steps it removes cost. Ask which steps the tool removes, what those steps cost locally, and what the tool costs per visit.

## Key Takeaways

Essential Points

1.  **Know the baseline:** At one academic health system, billing and insurance-related work cost an estimated \$20.49 per primary care visit to \$215.10 per inpatient surgical procedure, which the authors attributed in large part to differing payment requirements across payers ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)).
2.  **Neither inbox study detected a time saving:** Neither of two prospective studies of AI drafts for patient portal messages detected shorter reply time; one found longer read time and longer replies ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)), and the other found lower task load and work exhaustion scores without a control group ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)).
3.  **Prior authorization burden comes from plan rules as well as paperwork:** Physicians in a survey by an advocacy organization reported 40 requests per physician per week ([AMA, 2025 survey](https://www.ama-assn.org/system/files/prior-authorization-survey.pdf)); no source reviewed here shows that practice-side AI reduces that work.
4.  **In one preprint benchmark, the best agent completed 36.3% of whole tasks:** The tasks ran in simulated systems with synthetic patients ([Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint).
5.  **Measure time with audit logs:** Clinicians’ impressions of time saved and audit-log time disagreed at Stanford ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)); report both, with a comparison group.

Questions About Healthcare Administration

### How much does billing and insurance-related work cost per patient visit?

At one academic health system with a certified electronic health record, billing and insurance-related work was estimated to take 13 minutes and cost \$20.49 for a primary care visit, and 100 minutes and \$215.10 for an inpatient surgical procedure. Physicians’ share was a median of 3 minutes per primary care visit ([Tseng et al., 2018](https://doi.org/10.1001/jama.2017.19148)). The estimates were modeled from interviews at one centralized billing organization, and the authors describe them as conservative.

### Do AI-drafted replies save physicians time on patient messages?

The two early prospective studies did not detect a time saving. At Stanford, the mean draft utilization rate across clinicians was 20% of replies with a draft available, and audit-log reply, read, and write times did not change, although task load and work exhaustion scores improved among survey respondents ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)). At UC San Diego, in a randomized waiting-list study, drafts were associated with 21.8% longer read time, no significant change in reply time, and 17.9% longer replies ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)). Both studies used 2023 models at single academic health systems.

### Can AI reduce prior authorization work for practices?

No study reviewed here shows that it does. Physicians in a survey by the American Medical Association, which advocates prior authorization reform, reported completing 40 prior authorizations per physician each week, with 13 hours of physician and staff time ([AMA, 2025 survey](https://www.ama-assn.org/system/files/prior-authorization-survey.pdf)). In a preprint benchmark in simulated systems, the best AI agent completed 38.3% of prior authorization tasks from start to finish ([Bedi et al., 2026](https://arxiv.org/abs/2604.09937), preprint).

### How should a health system evaluate an administrative AI tool?

Name the task and whose time or cost it should reduce, measure that baseline first, and compare against a randomized, waiting-list, or concurrent control group. Measure time with audit logs and report surveys separately, because the two can disagree ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)). Track what could get worse, such as read time, errors, denials, or patients unable to reach a person, and audit a sample of what the tool sent or submitted.

### Are patients told when a reply was drafted by AI?

At UC San Diego, a reply started from a draft carried, by default, a statement that part of the message was generated automatically and was reviewed and edited by the physician ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)). Neither inbox study measured how patients responded. Disclosure policy is a local decision.

### Is it safe to send AI-drafted replies?

Only after a clinician has read and edited them. In one UC San Diego example, a draft suggested an x-ray or a physical therapy referral, which the authors called outside its scope, and the physician kept only the first sentence ([Tai-Seale et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.6565)). The Stanford authors reported no adverse safety signals over 5 weeks; safety was not among that study’s stated outcomes ([Garcia et al., 2024](https://doi.org/10.1001/jamanetworkopen.2024.3201)).

## Conclusion

The evidence reviewed here is thin: one costing study that sets a baseline, two prospective single-system quality improvement studies of drafted patient messages that did not detect time savings, one advocacy-organization survey of prior authorization burden, and one preprint benchmark of agents in simulated systems. Physicians and administrators can ask of each tool whose time it saves, how that time is measured, against what comparison, and what gets worse.

Continue Reading

- [Hospital Operations and Management](../implementation/hospital-operations.html): operating room scheduling, no-shows, patient flow, and medical coding evidence
- [AI in Medical Coding and Billing](../implementation/medical-coding-billing.html): automated coding research, documentation queries, and audit risk
- [AI-Assisted Clinical Documentation](../practical/documentation.html): ambient scribes and billing rules for AI-drafted notes
- [Healthcare Policy and AI Governance](../future/policy.html#sec-payer-algorithms): payer algorithms and federal prior authorization rules
- [Evaluating AI Agents in Healthcare](../implementation/ai-agent-evaluation.html): how to test agents that work through portals and EHR screens
- [Evaluating AI Clinical Decision Support Systems](../implementation/evaluation.html): evidence design and local validation

![](../images/bryan-tegomoh-author.jpg)

[Bryan Tegomoh, MD, MPH](https://bryantegomoh.com/)

Physician-scientist and epidemiologist in health AI and biosecurity.

[Follow @BryanTegomoh](https://x.com/intent/follow?screen_name=BryanTegomoh) [Follow on LinkedIn](https://www.linkedin.com/comm/mynetwork/discovery-see-all?usecase=PEOPLE_FOLLOWS&followMember=bryan-tegomoh-md-mph) [Support](../support.html)

The Physician AI Handbook is free to read and cite, with no third-party advertising. If this chapter informed your work, [you can support the handbook](../support.html).