AI in healthcare delivery

Segments covered: provider organizations, meaning physician groups, ambulatory practices, and the administrative and operational side of hospitals. Life sciences, health insurers, and social assistance are outside this guide's lens, and so are clinical diagnosis and treatment decisions. Linked workflows: meeting follow-up and speech systems for visit documentation, document extraction for referrals and prior authorization, support replies for patient messages, knowledge retrieval, and forecasting. Linked roles: operations leaders, finance leaders, engineering leaders.

A provider organization sells clinical work, but part of its labor goes into the record and the paperwork around it: the visit note, the referral packet, the prior-authorization request, the patient message, the payer rule to look up. That administrative work is where this guide looks for AI. Deciding what is wrong with a patient and what to do belongs to licensed clinicians and is not this guide's subject; where software crosses into clinical decisions, the guide says where the regulatory line sits and stops.

Sector figures come from the Centers for Medicare & Medicaid Services (CMS), the Census Bureau, and the Bureau of Labor Statistics, cited with period, population, and limits. Regulatory statements go no further than the cited text; beyond that, they are questions for counsel. Recommendations on sequencing and controls are proposed practice, not findings. The physician groups in the economics and sequencing sections are invented. No provider was interviewed and no deployment was observed.

How the business makes money

Provider organizations are paid for services (visits, procedures, admissions, tests) by Medicare, Medicaid, private insurers, and patients, with some revenue under capitated or value-based arrangements that pay per patient rather than per service. Two consequences follow. Revenue depends on documentation, because a service not documented and coded to the payer's rules may not be paid. And capacity is set by clinician time, but time freed from paperwork becomes revenue only if demand, schedule slots, and payer rules allow more visits.

For scale, CMS reports US national health expenditure of $5.3 trillion in 2024, up 7.2% and 18.0% of GDP. Hospital spending was $1,634.7 billion and physician and clinical services $1,109.7 billion. By source, Medicare paid 21 percent of the total, Medicaid 18 percent, private health insurance 31 percent, and patients out of pocket 11 percent. CMS National Health Expenditure Fact Sheet, historical 2024 data, page updated June 24, 2026 These are national spending accounts, not provider revenue or margins. By our arithmetic, hospital and physician and clinical services together were $2,744.4 billion, about half of the total.

Prices come from government fee schedules or negotiated contracts; a provider cannot raise a fee because its notes got faster. A practice paid per visit gains from added visits; one paid per patient gains from lower cost per patient. We found no official provider-margin statistic comparable to the Census figures in other industry guides, so each organization should use its own accounts.

The Bureau of Labor Statistics estimates health care employment at 18,553,600 in September 2026, including 9,266,400 in ambulatory health care services and 5,797,800 in hospitals (seasonally adjusted, preliminary). BLS Current Employment Statistics, series CES6562000101; ambulatory, CES6562100001; hospitals, CES6562200001 These are job counts and say nothing about how work is changing.

Segments differ. A ten-clinician practice and a hospital system have different payer contracts, IT capacity, and compliance staff, and a hospital adds bed management, supply chain, and revenue-cycle departments. This guide uses an ambulatory group as its main example because its documentation, messaging, and referral work is easiest to see whole.

Where AI could touch the value chain

No Field Guide chapter was written for healthcare. Where a row names a chapter, it is a pattern that transfers, with the differences set out below the table; "Not yet covered" means no chapter exists.

Visit documentation. The speech systems chapter covers transcription and consent, and the meeting follow-up chapter shows how to turn a conversation into output a person approves. Three things differ in a clinic. The output is a legal medical record that supports care by other clinicians, billing, and legal review, so an omitted allergy or an invented finding matters far more than a missed action item. The treating clinician reviews and signs it, and that sign-off is the control; a design that makes careless signing easy defeats it. And the conversation includes the patient, so consent to recording and audio retention are questions for the privacy officer and counsel, not vendor defaults.

Referrals and prior authorization. For faxed and PDF packets, the document extraction chapter applies directly: separate reading from deciding, keep uncertainty visible, and measure completed packets rather than fields. What differs is that a missed clinical detail can send a patient to the wrong slot or delay a submission, so a staff member owns each scheduling or submission step. The provider's side of prior authorization is in scope; the payer's coverage decision is not.

Patient messages. The support replies chapter pattern, drafting from approved content for a person to approve, transfers to administrative messages: appointment changes, forms, billing. Messages with clinical content, such as a new symptom or a question about a result, go to clinical staff rather than being answered from a knowledge base, and the routing rule needs clinical owners and testing against missed urgent messages. The failure that matters is a symptom that sat in an administrative queue.

Knowledge and payer rules. The knowledge retrieval chapter shows how to build a search that cites its sources and respects access. Payer rules change, so every answer needs its source and effective date, and stale documents need an owner.

Forecasting. The forecasting chapter method, comparing against a simple baseline with service held fixed, transfers to supplies and staffing demand. A staffing shortfall affects patient safety and labor agreements, so the forecast advises a scheduler and does not set staffing.

Not yet covered. Coding and billing errors carry compliance and payment consequences, and a chapter would need billing-compliance review this guide cannot supply. Clinical decision support and diagnosis are clinical judgments where device and nondiscrimination rules may apply. These omissions are scope choices, not judgments that AI has no role.

Constraints that shape every project

Patient privacy. Under the HIPAA rules, a business associate is a person or company that handles protected health information on a covered entity's behalf, other than as its workforce, for functions such as data analysis, quality assurance, billing, or practice management, or provides consulting or administrative services involving it; subcontractors are included. A covered entity may let a business associate handle that information only with satisfactory assurance, documented in a written contract or other written arrangement, that it will be appropriately safeguarded. 45 CFR 160.103; 45 CFR 164.502(e) Whether an AI vendor, its model provider, or its host is a business associate or subcontractor, and what the agreement must say, are questions for the privacy officer and counsel before any patient data reaches a trial.

Minimum necessary. Covered entities and business associates must make reasonable efforts to limit protected health information used, disclosed, or requested to the minimum necessary for the purpose, with listed exceptions including disclosures to a provider for treatment. 45 CFR 164.502(b) That raises a design question for the privacy officer: what part of the chart does each tool need to see? HHS's guidance pages on these topics and on online tracking could not be retrieved; this guide cites the regulation text only.

The record and the signature. A drafted note signed by a clinician is the clinician's note. The human review playbook explains how to define what a reviewer checks, budget the time, and catch reviewers who stop checking. Record and audio retention follow state law and organizational policy, which this guide does not state.

Where software becomes a regulated device. The Federal Food, Drug, and Cosmetic Act excludes from the device definition software for administrative support of a health care facility, such as billing, scheduling, business analytics, and practice management. It also excludes clinical decision support that supports a health care professional's decision and lets that professional independently review the basis for a recommendation, unless the software analyzes a medical image or a signal from a diagnostic device or signal-acquisition system. 21 U.S.C. 360j(o) FDA's current guidance on that line was issued January 29, 2026, replacing earlier versions including the September 2022 guidance. FDA, Clinical Decision Support Software guidance FDA's list of authorized AI-enabled devices is, by its own description, not comprehensive. FDA AI-enabled device list This guide's workflows sit on the administrative side by design; whether any function of a product is a device is a question for regulatory counsel.

Nondiscrimination in clinical decision tools. A Section 1557 rule bars covered entities from discriminating on the basis of race, color, national origin, sex, age, or disability through patient care decision support tools (any automated or non-automated tool supporting clinical decision-making), and requires reasonable efforts to identify tools using those characteristics as inputs and to mitigate the risk. 45 CFR 92.210 The identification and mitigation duties took effect within 300 days of July 5, 2024. A June 2026 HHS notice reports a court vacatur of parts of the 2024 rule concerning gender identity; the sections it lists do not include 92.210, and HHS states the rest remain in force. Federal Register, June 2, 2026 A message-routing tool that influences who gets clinical attention first is close enough to ask counsel whether the rule applies.

EHR certification rules. Certified EHR software must support listed "source attribute" information about decision support interventions, such as developer, funding, and whether inputs like race, sex, or date of birth are used. 45 CFR 170.315(b)(11) A December 2025 proposed rule would remove the source-attribute requirements; as of October 20, 2026 we found no final rule in the Federal Register. HTI-5 proposed rule These obligations fall on EHR developers, but they shape what a provider can learn about tools inside its EHR.

Payers' use of algorithms. For Medicare Advantage, CMS told plans in February 2024 that an algorithm may assist a coverage decision, but the decision must rest on the individual patient's circumstances, including medical history, physician recommendations, and clinical notes; a prediction alone cannot be the basis to deny an inpatient admission or end post-acute care. CMS FAQ on CMS-4201-F, February 6, 2024 (copy hosted by the American Hospital Association); 42 CFR 422.101(c) That binds plans, not providers, but the clinical notes a provider submits are what the plan must consider.

Systems and integration. A tool whose output staff must copy and paste into the EHR shifts work rather than removing it. Ask the EHR vendor and the engineering lead how drafts enter the record and what happens when the tool is down.

The sector's economics lens

For your finance lead

The decisive question is what released clinician minutes turn into. Saved time produces cash only through a few routes, each needing a decision someone owns.

The illustration uses an invented 20-clinician ambulatory group. Every input is invented, not a benchmark: 30 minutes saved per clinician per clinic day, 220 clinic days a year, and a tool priced at $250 per clinician per month.

The routes are alternatives for the same minutes and must not be added.

Route A is real but not cash. Shorter evening charting may affect satisfaction and retention. Do not enter it as a saving unless it can be tied to a cost that changes, such as fewer departures to replace.

Route B depends on demand and payment. Added visits need waiting patients, rooms, support staff, and supplies, and collections depend on payer mix and denials. The $424,000 is revenue less the tool, not profit. Under capitation, added visits may bring no added payment.

Route C needs a staffing decision. Scribe cost stops only when positions go unfilled or contracts end.

Revenue-cycle effects. More complete notes or referral packets might reduce denials or speed payment; faster payment is working capital, not annual income. None has attributable measurement here, and none is valued.

Double counting. If the $110 per visit already reflects fewer denials, do not count the denial effect again. The 30 minutes must be net of review: if clinicians spend ten minutes a day checking drafts, the release is twenty. The economics chapter and the finance leaders guide explain capacity versus cash; how costs are classified for financial reporting is a question for the organization's accountants.

What is measured about adoption and outcomes

For your analyst

Adoption. The Census Bureau's Business Trends and Outlook Survey (BTOS) asks employer businesses every two weeks: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" In the collection period of September 21 to October 4, 2026, covering September 7 to 20, 2026, the estimate for health care and social assistance (NAICS 62) was 24.8% answering yes, with a standard error of 1.01 points, against 23.8% for all sectors. The three preceding cycles were 24.0%, 24.7%, and 25.5%. For the same period, 29.3% of businesses in the sector said they expected to use AI in the next six months. BTOS sector data, published October 8, 2026

Read these figures narrowly:

• NAICS 62 includes social assistance, such as child care and family services, alongside ambulatory care, hospitals, and nursing and residential care, so it is broader than this guide's lens.
• They are weighted shares of employer firms, not of employment or revenue; a solo practice and a hospital system count equally. BTOS methodology, version 5
• Multi-unit companies operating in several sectors or states are reported separately, so large health systems may be missing from the estimate.
• The question was reworded in November 2025 from use "in producing goods or services" to use "in any business function," so earlier estimates are not comparable. Census Bureau, May 2026
• "Any business function" does not separate clinical from administrative use or say anything about results.

Which functions. A BTOS AI supplement, reference period November 2025 to January 2026, asked about use over six months by function. Census working paper CES-WP-26-25 For NAICS 62, as a share of all sector firms: sales and marketing 12.6%, information technology 12.6%, management and administration 9.4%, customer service 8.1%, and provision of services 6.4%. Among firms whose employees used generative AI, 86.8% used it for writing or editing documents and communications, 39.7% for interpreting or summarizing documents, and 37.5% for paperwork or filing. Among AI-using firms, 96.1% reported no change in total employment from AI. Among firms not planning to use AI, 54.4% said it was not applicable to their business and 28.1% cited privacy or security concerns. BTOS AI Supplement Table 2026 These are self-reports from a different period than the biweekly figure, covering the whole sector.

Outcomes. This guide reports no measurement of AI effects on documentation time, visit volume, denials, message response, or cost in provider organizations. Published studies and vendor case studies of ambient documentation exist, but this guide has not reviewed them, and vendor-reported results would be labeled as such. Reviewing that literature is the first item of evidence still needed.

Who is involved

The physician lead and operations leader must agree on what a released minute is for before a pilot; otherwise route A happens by default. Privacy and compliance belong before vendor selection, not at signature. A vendor cannot decide who signs a note or which messages go to clinicians. Clinical, nursing, and compliance roles have no role guide yet.

Expertise specific to this industry. A general AI builder lacks the healthcare knowledge this guide's constraints depend on: whether a vendor, its model provider, or its host is a business associate or subcontractor and what the agreement must say; what minimum-necessary access each tool needs; how a draft enters the EHR and what happens at downtime; and the clinician sign-off that makes a drafted note the clinician's own. This guide places that expertise in-house, with the privacy officer, the physician lead, and EHR and IT staff, before vendor selection, with regulatory counsel as adviser where a function nears the device line or Section 1557. See Staffing an AI pilot and Choosing and working with an AI partner.

Sequencing: a first move and what follows

The following is a fictional, proposed sequence. Brightwater Primary Care is an invented physician-owned group with six clinics and 34 clinicians, paid mostly per visit, with one small capitated contract. Its referral queue runs two weeks behind, and clinicians complain about evening charting.

First: measure the work. Before choosing a vendor, Brightwater samples two weeks of after-hours EHR time, the referral backlog's causes, and portal messages by type. In the invented findings, after-hours charting is concentrated among a third of clinicians, the referral backlog comes from incomplete packets, and scheduling and forms make up a large part of messages. Counsel confirms what agreements are needed before any patient data reaches a vendor.

Second: a bounded documentation pilot. Brightwater pilots drafted notes with eight volunteer clinicians, patient consent under a counsel-approved policy, and every note signed by the treating clinician. Beforehand, the medical director and operations leader agree on the route: four clinicians add one visit slot a day once drafts are reliable; the others finish charts in clinic hours. A physician reviewer audits sampled notes against recordings for omissions and invented findings, following the pilot design playbook.

Third, in parallel and independent: referral intake. Staff use document extraction to pull referral fields into a review screen and flag missing items; a scheduler approves every appointment. This could come first if the referral backlog is the larger problem.

Later, if both work: administrative message drafting. Using the support replies pattern, Brightwater drafts replies to scheduling and forms messages only; a routing rule owned by the nursing lead sends anything about symptoms, medications, or results to clinical staff.

Counterexample: when to wait. Suppose Brightwater had found instead that its EHR could not accept drafts, its clinicians already finished charting in clinic hours, and it had no privacy officer and unsigned agreements with existing vendors. A documentation tool would then add copy-and-paste work to solve a problem the practice does not have, under unreviewed data arrangements. The right call is to fix vendor agreements and referral forms and revisit when integration exists.

Questions for operators, investors, and vendors

• Where does clinician and staff time go today: notes, inbox, referrals, prior authorization, or phone? Who measured it?
• For documentation, what will released time become, and who owns that change in the schedule or staffing?
• Will the vendor sign a business associate agreement? Which subcontractors handle patient data, and what is retained, for how long, and for what uses, including model training?
• How are errors in drafted notes found before signing, and how is clinician review measured over time?
• Does any function suggest diagnoses, orders, codes, or triage priority? Has regulatory counsel reviewed it?
• For message drafting, how are clinical messages detected and routed, and what was the miss rate in testing?
• For a vendor claim, what settings, baseline, and period were used, and who other than the vendor measured it?

When to wait

Defer when the EHR cannot accept drafts in the right place, nobody has decided what released clinician time is for, privacy and counsel have not reviewed vendor agreements and data flows, or a product's useful functions cannot be separated from diagnostic or treatment suggestions. Defer message drafting until clinical routing rules have owners and have been tested. Simpler investments may do more: cleaning up referral forms and fax workflows, using EHR templates and smart phrases, adding scheduling capacity or a scribe for the clinicians with the heaviest charting load, or fixing the top denial reasons with ordinary rules. Those are legitimate outcomes of the analysis, not failures to adopt AI.

Evidence still needed and related reading

The largest gap is outcome evidence: this guide has not reviewed the peer-reviewed literature on ambient documentation, message drafting, or prior-authorization automation, and reports no effect sizes. Operator accounts are missing for how released clinician time was used, how note review was audited, and how clinical messages were routed. No comparable source for provider margins was found. HHS guidance on business associates, minimum necessary, and online tracking could not be retrieved; read it before relying on this section. Recheck the HTI-5 rule and Section 1557 litigation before each update.

Continue with speech systems, meeting follow-up, document extraction, human review, operations leaders, and the economics chapter.

Hiring in New York: open roles at NYC AI companies building for healthcare, read daily from the companies' own job boards. These companies sell AI into the industry; they show which skills vendors are hiring for, not what firms in the industry are hiring.