AI Agents for OB-GYNPractice Automation in 2026

Almost every specialty guide about AI in medical practices makes the same silent assumption: that a patient visit is a transaction. Someone books, someone arrives, someone documents, someone bills. Automate any link in that chain and you have automated something real.

Obstetrics breaks that assumption twice. The visit is not a transaction, it is one rung on a ladder that has to be climbed in order across roughly forty weeks. And the claim is not attached to the visit — under the global obstetric package, twelve months of care can produce a single claim, filed after delivery. A practice administrator who has run a dermatology or an orthopedic front office and then walks into an OB-GYN practice discovers that most of the operational intuitions transfer and the two most important ones do not.

That is not a reason to avoid automation here. It is the reason automation here is unusually valuable, because the work that software is best at — counting, remembering, reconciling, and not getting bored — maps almost exactly onto what this specialty actually finds hard. What it is a reason for is building in a specific order, with a specific boundary, which is what this article lays out.

One more scoping note before the workflows. There is a widely cited finding that for every hour of direct clinical face time, physicians spend nearly two additional hours on EHR and desk work. It comes from Sinsky and colleagues in the Annals of Internal Medicine in 2016, and it is worth citing with its limits every single time: 57 physicians, 430 observed hours, four specialties — family medicine, internal medicine, cardiology and orthopedics — in four states, with 21 physicians self-reporting after-hours diaries showing one to two hours nightly, and the authors themselves noting the data came from self-selected, high-performing practices and may not generalize. Note what is not on that specialty list: obstetrics and gynecology. We could not locate an equivalent time-motion study for this specialty. If someone quotes you an OB-GYN-specific administrative burden figure, ask where it came from before you put it in a business case.

Start with the structure, because everything operational follows from it. Under the global obstetric package, the practice reports one code covering antepartum care, the delivery and postpartum care: 59400 for routine obstetric care with vaginal delivery, 59510 for cesarean, 59610 for vaginal delivery after previous cesarean, and 59618 for cesarean delivery after previous cesarean. The global codes are billed as one unit, with either a single date of service or a date span.

The specifics in this section come from UnitedHealthcare's Commercial and Individual Exchange Reimbursement Policy 2026R0064A, which follows CPT and ACOG, corroborated by the CMS National Correct Coding Initiative Policy Manual for 2026. Read the payer-variance warning in the limitations section before you treat any of it as universal — Medicaid programs frequently unbundle prenatal care entirely.

The package includes all routine prenatal visits until delivery — "approximately 13 for uncomplicated cases"in the policy's own words. That number is worth pausing on, because it is one of the very few clean, sourceable operational figures in the entire medical-AI market. It is not a vendor benchmark. It is stated in a national payer's published reimbursement policy, sourced to CPT and ACOG, and you can open the PDF and read it. Most numbers in this market do not survive that test, which is why this article prints so few of them.

Global obstetric package composition per UnitedHealthcare Reimbursement Policy 2026R0064A, with corroborating language from the CMS NCCI Policy Manual 2026. One payer's policy — verify yours.

Now the timing rule, which is the one that governs the entire revenue cycle and which almost no general-purpose practice software models correctly. High-risk and complication E/M visits "should not be reported until after the patient delivers," because — quoting the policy — "it is not until then that appropriate assessment for the number of antepartum visits can be made." Those visits must carry modifier 25 and a high-risk diagnosis. And there is a second condition on top, drawn from ACOG: if the extra visits were performed to monitor a potential problem and no problem actually developed, the additional visits are not reported at all.

Sit with what that asks of a practice. A visit that happens in March generates a billing decision in November, and the input to that decision is a clinical judgment about why the visit happened and how the pregnancy subsequently went. The information needed to make it correctly is recorded across nine months in a chart nobody re-reads end to end. This is a memory and reconciliation problem, and it is the single most defensible place for software in this specialty.

There is also a case the ledger must handle from day one: shared care. When the practice does not provide the full course — the patient transfers in, transfers out, or changes insurer mid-pregnancy — the antepartum-only codes apply: 59425 for four to six visits and 59426 for seven or more, reported as a single claim submission and excluding the confirmatory visit. Fewer than four visits are itemized as E/M. Any agent that assumes every pregnancy in the panel is a full-course pregnancy will be wrong about a meaningful share of the panel, and it will be wrong quietly.

Two more mechanics from the NCCI manual that a billing agent should encode as hard rules rather than suggestions. Maternity procedures carry a global period of MMM on the Medicare Physician Fee Schedule Database, and wound repair codes 12001–13153 shall not be reported to describe closure of a surgical incision for codes with an MMM global period. Third- and fourth-degree lacerations are handled instead by appending modifier 22 to the global or delivery-only code with supporting documentation — and that determination is clinical, so the agent flags the candidate and stops.

The prenatal series is a scheduling problem unlike any other in outpatient medicine. It is a cadence, not a calendar: monthly, tightening to biweekly, tightening to weekly toward term, interleaved with imaging appointments that sit on a different resource with a different authorization rail — and all of it running against an unschedulable event. Labor pulls the physician out of clinic without notice, and the cancellations it causes do not cost one appointment. They cost a rung on a ladder, and the ladder has a deadline that cannot move.

On top of that sits a gynecologic surgical block, a postpartum window inside the global package, and a well-woman recall cadence for the non-obstetric side of the panel. No other specialty in this cluster has a scheduling problem this structurally hard, which is exactly why an agent that treats each booking as independent produces a calendar that is technically valid and operationally wrong.

Frenchy Digital workflow shortlist for OB-GYN practices, 2026, ordered by defensibility.

Notice what is not on that list. There is no triage agent, no symptom checker, no autonomous messaging, no coverage determination. Those are the workflows vendors lead with, and they are the ones that either cross the FDA line, cross a state disclosure line, or put untrusted patient text directly into an agent's instruction path. The five above have a common profile: they compute over data the chart already holds, they hand a human a conclusion, and if they are wrong the cost is a wasted worklist rather than a clinical or legal event.

The imaging rail deserves its own note. Maternal and fetal ultrasound 76801–76828 sits outside the package, on separate authorization, often on a separate resource with its own capacity constraints and sometimes at a separate site. Two failure modes recur: a study performed on a lapsed authorization, and a study performed and then never billed because it was mentally filed under "the pregnancy" and the pregnancy is billed as one code at the end. Both are detectable from data the chart already holds, and both are pure paperwork-state tracking with no clinical judgment involved.

The gynecologic side of the practice runs on a different clock and different billing constructs, and it has to be reconciled against an obstetric calendar that can be disrupted at any hour. Surgical scheduling here means pre-authorization, facility or ASC coordination, pre-op clearance from other practices, and a date that must survive the surgeon's call exposure.

The billing constructs come from the Medicare Claims Processing Manual, Chapter 12, and two of them are worth encoding as pre-submission checks because the manual states them explicitly and both are easy to get wrong at claim time. The first is the modifier distinction: if an evaluation and management service occurs on the day of surgery, the physician bills using modifier -57, not -25, and -57 is not used with minor surgeries because the global period for minor surgeries does not include the day prior. The second is split care: where physicians agree on transfer of care during the global period, the surgeon reports the surgical code with modifier -54 and the post-operative provider reports the same code, same date of service, with modifier -55 — and the date on which care was relinquished or assumed must be shown on the claim in the remarks field.

Gynecologic surgery coordination workflows and their boundaries, Frenchy Digital, 2026.

There is one item in the split-care mechanism that is emphatically not automatable, and it is worth naming because the surrounding paperwork looks so automatable that people assume it comes along. Chapter 12 requires that both the surgeon and the physician providing post-operative care keep a copy of the written transfer agreementin the beneficiary's medical record. That is a signed clinical hand-off between two named humans. An agent can detect that it is missing, chase it, and refuse to let the claim proceed without it. An agent cannot be a party to it.

Prior authorization deserves a realistic framing here rather than an optimistic one. CMS-0057-F is frequently pitched to practices as though it were about to solve their authorization problem. Read the rule text: every obligation in 42 CFR §422.122 runs to "an MA organization." It binds payers, not practices. What is live now is the payer side of the exchange — a specific reason for denial regardless of the communication method beginning January 1, 2026, and public reporting of prior-authorization metrics. The payer FHIR prior-auth API is a January 1, 2027obligation, and even then a practice's relationship to it is as a potential consumer, not as a duty-holder. Plan your authorization workflow around document exchange and portals, because that is what you will be doing through 2026.

This is the single most important architectural constraint in medical AI, and it is the one most often left out of a sales conversation. The ONC/ASTP Certification Companion Guide for §170.315(g)(10), last updated 05-15-2026, states it plainly:

The § 170.315(g)(10) certification criterion requires Health IT Modules to support API-enabled 'read' services for single and multiple patients. … These services specifically exclude 'write' capabilities, where authenticated and authorized third-party applications would be able to create or modify EHI through a secure API.

— ONC/ASTP Certification Companion Guide, § 170.315(g)(10), updated 05-15-2026

Read that as an operational sorting rule. Anything your agent needs to know is available through a certification-backed pathway in any certified EHR. Anything your agent needs to do inside the EHR — book the slot, move the appointment, post the charge, file the note, update the problem list — happens one of exactly three ways: through a write API your EHR vendor chooses to offer you on commercial terms and after its own app review, through user-interface automation with the fragility and terms-of-service exposure that implies, or through a human clicking a button.

There is a security dividend in this constraint that is worth taking deliberately rather than accidentally. An agent that cannot write cannot be manipulated into writing. Read-only-by-default is the strongest single control available against the prompt-injection problem described below, and it costs nothing, because the certification landscape has already imposed it on you. Practices that treat the read-only boundary as a limitation to be engineered around usually end up with both a fragile integration and a wider blast radius. Practices that treat it as the default architecture — agent reads, human commits — end up with a system that degrades safely.

A related caution about the regulatory ground under all of this. ASTP/ONC published a proposed deregulatory rule in the Federal Register on December 29, 2025 that would, among other things, remove the clinical decision support certification criterion and reserve §170.315(a)(9). It remains a proposed rule; comments closed February 27, 2026 and we located no final rule. Do not plan around it. What survives regardless of how that rulemaking resolves is the operational point: an EHR write path is a commercial arrangement with your vendor, not a right you can assume.

Every practice type in this cluster handles sensitive data. This one handles data where the existence of the appointment can be the disclosure. Pregnancy, pregnancy loss, fertility treatment, contraception, sexually transmitted infection, sexual assault care and gynecologic oncology all share a property that a scheduling engineer will not think about unprompted: a message read by the wrong person in a shared household, or a portal notification on a shared device, can cause harm without any clinical content leaving the building at all.

We want to be precise about the legal ground here rather than dramatic, because the distinction matters for how you build. HIPAA has no general provision about the mere fact of association. The one place federal law does work that way is 42 CFR §2.12, the substance-use-disorder confidentiality rule, whose trigger is identification of a patient as having or having had a substance use disorder in records obtained by a federally assisted SUD program — not the clinical content. That distinction is directly relevant to obstetrics, where substance-use screening in pregnancy is routine and where a practice may be receiving records from, or operating, such a program.

For reproductive-health data more broadly, we could not verify from primary sources within this article's research scope whether any reproductive-health-specific federal privacy provision is currently in force beyond baseline HIPAA. So we are not going to tell you one exists, and we are not going to tell you one does not. What we will tell you is that the association risk is real regardless of how that question resolves, that it is a design decision the practice owns, and that the correct engineering posture is to assume the strictest reading and ask your counsel for the current position before you rely on anything looser.

Frenchy Digital privacy boundary for OB-GYN agent deployments, 2026. Regulatory citations are to the rules in force as of August 12, 2026.

Three of those rows deserve expansion because they are where implementations actually go wrong.

One further note for practices offering perinatal behavioral health, because the strictest statutory language in the country on AI in a clinical conversation sits there. Illinois PA 104-0054 §20(b) prohibits a licensed professional from allowing AI to make independent therapeutic decisions, to directly interact with clients in any form of therapeutic communication, to generate therapeutic recommendations or treatment plans without review and approval by the licensed professional, or to detect emotions or mental states. Its §15 requires that where a session is recorded or transcribed, the patient is informed in writing that AI will be used and of the specific purpose of the tool, and consents — and the statute explicitly excludes acceptance of broad terms of use from what counts as consent. The same act defines a safe zone: administrative support means tasks that do not involve communication, including appointment scheduling and reminders, billing and claims processing, and drafting general logistics communications that contain no therapeutic advice. That is a good description of the boundary this whole article is drawing, written into law by one state.

Put this table in the statement of work, not in the appendix. Every disagreement we have seen between a practice and a vendor six months into a deployment traces back to a row in it that was never made explicit at the start.

Frenchy Digital human-in-the-loop boundary for OB-GYN agent deployments, 2026.

The attestation row is the one that ends every argument, so it is worth quoting the source. The Medicare Program Integrity Manual, Chapter 3 §3.3.2.4, effective January 17, 2025, addresses AI by name:

The treating physician/non-physician practitioner's (NPP's) signature on a note indicates that the physician/NPP affirms the note adequately documents the care provided. We note this type of practitioner concurrence is also required when using Artificial Intelligence (AI) technology to capture the transcription of medical record entries.

— CMS Medicare Program Integrity Manual, Ch. 3 §3.3.2.4 (Rev. 13008, effective 01-17-25)

The same section closes the obvious workarounds. Reviewers shall not consider attestation statements from someone other than the author of the record entry in question — explicitly including the case where two individuals are in the same group. Contractors shall not consider attestation statements where there is no associated medical record entry. And an attestation cannot be used to backdate a plan of care. The model attestation CMS publishes is first-person, credentialed, and states that the signer understands falsification, omission or concealment may subject them to administrative, civil or criminal liability. There is no configuration in which software makes that statement.

The "never" rows are not risk aversion. They are the places where the consequence of being wrong is not a rework queue. Everything above them is where the value is, and there is a great deal of value above them.

This is the part of the article that matters most, and it is deliberately not a list of benefits. Each phase below has an owner, an entry criterion, an exit criterion and an instruction for what to do when the phase fails. A phase that has not met its exit criterion does not hand off to the next one, and nothing after Phase 0 starts until the write-path question has an answer in writing.

Frenchy Digital sequenced implementation path for OB-GYN administrative agents, 2026. Week ranges assume a single-site practice with one EHR and two dominant payers.

Some notes on why the sequence is ordered this way, because the ordering carries most of the value.

• Phase 0 exists to kill bad projects cheaply: Two to four weeks and a discovery fee is the correct price for discovering that your EHR vendor will not grant a write path, or that your Medicaid book unbundles prenatal care so completely that the global-package logic covers a minority of your volume. Both of those are common. Both are catastrophic to find in week fourteen.
• Phase 1 produces a number you generated, not a number a vendor gave you: Reproducing the biller's own visit count on thirty consecutive delivered charts is a real acceptance test with a real ground truth. It also gives you an honest estimate of how much of your panel is shared care, which determines how much of the rest of the programme is worth building.
• Phase 2 is where the money is, and it needs a named physician reviewer from day one: Deferred-billable capture touches the claim. If a physician reviewer is not named, funded and scheduled before the phase begins, the held lines pile up unreviewed and the phase quietly converts into an unbilled backlog, which is worse than not having built it.
• Phase 3 is deliberately after the billing work, not before it: Patient-facing messaging is the phase most likely to be pitched first because it demos well. It is also the phase with the highest privacy consequence and the least reversible failure mode. Put the compliance review of every template on the critical path, and treat a clinical token in a template as a stop-the-line event.
• Phase 4 reconciles the two calendars: Surgical coordination only works once the obstetric side is instrumented, because the constraint on the surgical calendar is the surgeon's obstetric call exposure. Building it first produces a scheduler that is confidently wrong about availability.
• Phase 5 is optional and most practices should skip it: If your EHR vendor grants a contracted write path and you can demonstrate a tested reversal procedure, automated writes remove real clicks. If either condition is missing, the human commit queue is not a consolation prize — it is the design that degrades safely, and it is what we recommend by default.

Every deployment breaks. The difference between a recoverable programme and an abandoned one is whether the failure was anticipated, instrumented and reversible. These are the specific failure modes for this specialty, in rough order of likelihood, each with the signal that tells you it is happening and the rollback that stops the bleeding.

Frenchy Digital failure-mode register for OB-GYN agent deployments, 2026.

Two of those rows are worth more than the others, and they are the two that are least like ordinary software failure.

The first is the message that discloses the visit type. It will not present as an incident report. It will present as one patient, once, telling the front desk something quiet. By then the template has been sending for weeks. That is why the control has to be preventive — an automated scan that fails any template containing a clinical token, run in continuous integration, plus a weekly sample audit of actually-sent messages for the first quarter — rather than detective. And it is why the rollback has to be a switch someone can flip in minutes without an engineer.

The second is prompt injection, which deserves its own treatment because the published evidence lands directly on this specialty. Lee and colleagues, in JAMA Network Open in December 2025, ran a controlled simulation across 12 clinical scenarios including pregnancy contraindications. Across 216 evaluations, attacks achieved 94.4% success at turn 4 and persisted in 69.4% of follow-ups, and extremely high-harm scenarios including FDA Category X pregnancy drugs succeeded in 91.7% of dialogues. The authors' conclusion was that commercial LLMs demonstrated substantial vulnerability to prompt-injection attacks that could generate clinically dangerous recommendations, and that even flagship models with advanced safety mechanisms showed high susceptibility.

And the finding that breaks the standard mitigation story: Clusmann and colleagues, in Nature Communications in February 2025, ran 594 attacks against four vision-language models and found all of them susceptible, including to sub-visual prompts embedded in medical imaging data that are, in their words, non-obvious to human observers. A human-in-the-loop who cannot see the injection cannot review it away. FDA's own 2026 material makes a compatible point from a different direction, describing automation bias as the propensity of humans to over-rely on a suggestion from an automated system, producing errors of commission and omission, and increasing where there is not sufficient time to consider other information.

We should also be honest about what has not been studied. We found no published study of prompt injection via patient portal messages, referral faxes or payer portals in a live practice. The published work covers patient dialogue and imaging. Extending it to a fax-to-OCR-to-agent pipeline is reasoning, not evidence — though the reasoning is not weak: an inbound fax is untrusted text entering an instruction path with no sender authentication at all.

The mitigations that are real are architectural, because they do not depend on detecting the attack. Least privilege at the API boundary. Read-only by default, which the certification landscape has already given you for free. A human signature as the only path from draft to record. Unique agent identity plus audit controls so a successful injection is reconstructable afterward. Out-of-band confirmation for anything irreversible — money moved, orders placed, messages sent to patients, records released. What is marketing, and should be named as such, is any vendor claim of an injection detection or guardrail accuracy rate: there is no independent benchmark for clinical prompt-injection defense that we could locate, and every figure in circulation is vendor-published.

Most regulatory writing in this market is either alarmist or promotional. The actual position, as of August 12, 2026, is narrower and more usable than either: administrative automation sits on statutory safe ground, clinical recommendation sits inside a four-criterion test with a revocable enforcement posture layered on top, and a handful of state statutes impose disclosure duties that vary by state and by whether the content is clinical or administrative.

Regulatory boundary summary for medical-practice AI agents, verified against primary sources on August 12, 2026. Not legal advice — verify against counsel and re-check dated items.

Four clarifications that practices consistently get wrong, and that vendors consistently exploit.

• The administrative exclusion is statutory, not a favor: 21 U.S.C. §360j(o)(1)(A) excludes from the device definition software intended for administrative support of a health care facility, and names the processing and maintenance of financial records, claims or billing information, appointment schedules, business analytics, information about patient populations, admissions, practice and inventory management, determination of health benefit eligibility and population health management. Every workflow in this article's shortlist sits inside that sentence. You are not relying on FDA's forbearance for any of it.
• Enforcement discretion is not approval: FDA's Clinical Decision Support Software guidance issued January 29, 2026 supersedes the January 6, 2026 version, which itself replaced the September 2022 guidance. It did not reclassify single-output software as non-device. Software providing a specific preventive, diagnostic or treatment output or directive still fails Criterion 3; FDA announced that it intends to exercise enforcement discretion where only one option is clinically appropriate and the function otherwise meets §520(o)(1)(E). Discretion can be withdrawn without notice-and-comment. Any vendor saying FDA approved, cleared or blessed their CDS is wrong.
• The HIPAA rules you are held to today are the existing ones: The HIPAA Security Rule NPRM published January 6, 2025 is still only proposed — the Unified Agenda entry for RIN 0945-AA22 places it in long-term actions with final action projected July 2027. So mandatory multi-factor authentication, asset inventory and encryption at rest are not obligations today. The technical safeguards you are actually bound by are 45 CFR §164.312, under which encryption remains addressable rather than required. Build to the stricter standard by choice; do not let a vendor tell you the law already requires it.
• The nondiscrimination duty is live and it is an inventory task: 45 CFR §92.210 imposes a general prohibition, an ongoing duty to make reasonable efforts to identify patient care decision support tools using input variables measuring race, color, national origin, sex, age or disability, and a duty to make reasonable efforts to mitigate the resulting risk. It applies whether or not the tool is AI. The compliance date is reported as May 1, 2025, though we note that as secondary. In an obstetric practice, sex is an input to essentially every risk instrument in use, which makes the identification duty concrete rather than theoretical. Whether §92.210 is subject to any injunction or pending rescission, we could not verify — we can only report that it appears in the current CFR.

On state AI law, the practical rule is that the trigger differs by state and the content type matters. California's AB 3030 turns on whether the communication pertains to patient clinical information — and the statute expressly says that term does not include administrative matters including appointment scheduling, billing, or other clerical or business matters. So an appointment confirmation needs no disclaimer; a message answering a clinical question does, unless a licensed human read and reviewed it first. Texas HB 149 §552.051(f) is broader on its face, reaching any AI system used in relation to health care service or treatment, with disclosure required no later than the date service is first provided and no human-review exemption written into that subsection. Texas SB 1188 is the most instructive statute in the country here because it grants permission and attaches conditions in the same breath: a practitioner may use AI for diagnostic purposes if acting within scope regardless of the use of artificial intelligence, if the use is not otherwise restricted, and if the practitioner reviews all records created with AI — plus disclosure to patients. Five words carry the entire doctrine: the license, and therefore the liability, does not move.

On Colorado, we have to flag a genuine inconsistency in our own sources rather than paper over it. SB 26-189 was signed May 14, 2026 and repeals and reenacts Colorado's AI statute. Our regulatory research places the substantive deployer and developer duties, and the applicability to consequential decisions, at January 1, 2027, with only narrow rulemaking-authority subsections effective on passage. Another source in our own research set argued for an earlier effective date. We are reporting the January 1, 2027 position because it is the one supported by the enrolled bill text our researchers parsed, and we are telling you the disagreement exists so you check it rather than trusting either of us. What is stable and matters more for a practice: the reenacted statute carries a carve-out for HIPAA covered entities and their business associates — but only to the extent the covered entity is doing business in Colorado, and for a health-care provider only where it operates from a location within Colorado, and never for a consequential decision related to employment or an employment opportunity. A covered entity must still provide patients a general notice of the use of advanced technologies; and a richer disclosure attaches where automated decision-making touches financial-assistance or discounted-care eligibility. Colorado's Department of Law filed proposed draft rules on August 11, 2026 with a comment period running into October — so this is moving, and a statement made today has a short shelf life.

Finally, the federal wildcard, hedged deliberately: federal preemption of state AI law is being actively pursued, and nothing has displaced the state statutes above. We are not going to cite an executive order number, a bill number or any provision, because we could not fetch the primary documents. If a vendor tells you state AI rules are about to be preempted and therefore you need not plan for them, ask them for the citation.

Each row corresponds to a specific documented constraint from earlier in this article. None of them are hypothetical, and each is answerable in a first meeting.

Frenchy Digital red-flag list for OB-GYN practice AI buyers, 2026.

On the accuracy row, we want to be explicit about the house position, because it is the thing that most distinguishes how we evaluate this market. We do not score vendors on accuracy. There is no independent benchmark for AI accuracy in medical scheduling, coding, ambient documentation, or voice agents — every figure circulating in this market was published by a company selling into it. Excluding accuracy from evaluation is not a limitation of our method. It is the method. What can be verified is documented EHR integration, whether a BAA is publicly offered, SOC 2 or HITRUST status, pricing transparency, whether a product is EHR-locked or standalone, and corporate stability from public record. Score those. Ask for accuracy methodology, and treat a headline number offered without one as a statement about the vendor rather than about the product.

Ask the vendor what their system must never do. A vendor with a real answer has thought about the boundary, has probably lost an argument with a compliance officer, and has shipped into a practice that survived it. A vendor who says the system can do anything you configure it to do has told you they will hand you the boundary problem at go-live.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope medical-practice agent work in 2026. They assume the integration assessment — including the EHR write-path question — and a measured before-state are in scope from day one, because those are what make the result defensible afterward.

Frenchy Digital cost bands for medical-practice AI agent engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, evaluation-set expansion, payer-rule maintenance, incident response and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

A budgeting note specific to this specialty. The payback argument here is unusually easy to state and unusually easy to measure, because the deliverable is a claim that goes out complete and on time. Measure days from delivery to claim submission, rework rate on obstetric claims, and the number of separately billable services performed and never billed, before and after. If those three do not move, the programme did not work, and you should be able to say so within a quarter. We would rather build something you can turn off than something you cannot evaluate.

The retainer line is not a maintenance formality either. Payer policies change, an EHR release moves a field, a state adds a disclosure duty, and a rule set that was correct in January is quietly wrong by October. A practice running an unmaintained billing agent is running an unmonitored one, and the failure mode is silent — the claims still go out.

This article has spent several sections holding vendors to an evidence standard. It would be indefensible not to apply the same standard to itself. Here is what we could not establish, stated plainly.

• Reproductive-health-specific federal privacy provisions: We could not verify from primary sources, within this article's research scope, whether any reproductive-health-specific federal privacy provision is currently in force beyond baseline HIPAA, or what its litigation status is. We have therefore made no claim in either direction and have built the privacy section on the rules we could verify — §164.502, §164.312, §160.103 and 42 CFR Part 2 — plus a design posture that assumes the strictest reading. Ask your counsel for the current position before relying on anything looser.
• Minor-consent and adolescent-confidentiality law: Gynecologic care for adolescents raises real questions about which contact record receives a message and which portal account can see what. State minor-consent law varies and we did not verify it for any state. Resolve your portal and contact-record configuration with counsel before enabling any outbound messaging on a mixed-age panel.
• The global-package rules here are one payer's: The composition of the obstetric package in this article is UnitedHealthcare's 2026 policy 2026R0064A, which follows CPT and ACOG, corroborated where possible against the CMS NCCI Policy Manual 2026. Medicaid programs frequently unbundle prenatal care entirely, and other commercial payers vary. Encode your own payers' policies; do not encode ours.
• No independent accuracy benchmark exists for any of these product categories: Not for scheduling, not for coding, not for ambient documentation, not for voice agents. Every accuracy figure in this market is vendor-published. We excluded accuracy from evaluation by design and we are telling you that rather than substituting a number of our own.
• The standard operational statistics in this market do not survive tracing: The widely quoted per-no-show cost figure and the annual national cost of missed appointments both trace to a single 2017 byline by a scheduling vendor's chief medical officer with no methodology. The phone-abandonment benchmarks used to sell voice agents have no traceable origin. The frequently cited claim-rework figures could not be chased to a published methodology. We do not print them, and a business case built on them has no floor.
• No OB-GYN-specific administrative time-motion study: The Sinsky 2016 study everyone cites covered family medicine, internal medicine, cardiology and orthopedics — not obstetrics and gynecology. We could not locate an equivalent for this specialty. If you need a burden baseline, measure your own practice; it takes two weeks and it is yours.
• Colorado's effective date is disputed within our own research: One line of our research places SB 26-189's substantive duties at January 1, 2027 based on the enrolled bill's own effective-date section; another argued for an earlier date. We report the January 1, 2027 position and flag the disagreement rather than presenting a clean answer we do not have. Colorado rulemaking is live as of August 11, 2026, so re-check before relying on any statement here.
• The read-only certification position rests on the Certification Companion Guide: We quote the ONC/ASTP Certification Companion Guide for §170.315(g)(10), updated 05-15-2026, rather than the regulation text itself, which we did not parse directly. The operational conclusion — that EHR write access is vendor-discretionary — is unaffected either way, because it follows from what vendors actually offer commercially.
• Prompt injection in practice channels is unstudied: The published evidence covers patient dialogue and medical imaging. We found no study of injection via patient portal messages, referral faxes or payer portals in a live practice. Our treatment of those channels is reasoning from the imaging and dialogue findings, and we have labelled it as such throughout.
• Ordinary software failure still applies: Document extraction degrades on poor scans and unfamiliar layouts. Classifiers drift as payer templates change. Integrations break on EHR upgrades. None of these are exotic AI failures, all of them need an owner and a monitoring budget, and that is what the retainer exists for.

None of this argues against building. It argues for building in the order above, instrumenting the before-state, and being straight with your own partners about which numbers in the business case were measured in your practice and which were borrowed from somewhere else.

The boundary holds regardless of any of it. Everything described here is an administrative and documentation system operating under human review. No agent in this article makes a clinical determination, signs or attests to a record, decides medical necessity, releases information, or acts on legal process. A human commits every consequential action, and the audit log proves which human it was.

The honest summary of this specialty is that the highest-value automation available to you is also the least glamorous thing in the vendor market. Nobody demos a visit counter. Nobody builds a keynote around a held-line queue that gets reviewed at delivery. But a practice that counts its antepartum visits correctly, captures the separately billable deviations contemporaneously, and submits a complete claim within days of delivery has fixed the thing that actually costs it money — and it has done so without touching a clinical decision, a signature, or a disclosure.

Start with Phase 0. Get the write-path answer in writing. Reproduce your biller's visit count on thirty charts. If those three things go well, the rest of the sequence is ordinary software delivery with an unusually careful boundary. If they do not, you have spent two to four weeks learning something true about your own operation, which is a better outcome than most AI programmes in this market achieve in a year.