The Zero-Click Clinic:An AI Agent Blueprint for 2026

A zero-click clinic is a practice where the default administrative path completes without a human click, and humans are deliberately routed to the exceptions and to every clinical, billing-attestation, legal or safety decision.

Read that definition twice, because two words carry the entire idea and both of them get lost in the marketing. The first is default. The zero-click clinic is not a clinic where nothing is clicked; it is a clinic where the ordinary, unremarkable, high-volume path — the eligibility check that resolves cleanly, the fax that belongs to an obvious patient, the recall list that computes itself — reaches its end without anyone touching it. The second is deliberately. The exceptions are not what fell through. They are what you routed to a human on purpose, because a human is the right answer.

This matters because primary care and family medicine sit at the exact point where the volume is highest and the margin per encounter is thinnest. A family medicine practice runs the widest administrative surface of any specialty in ambulatory care: it schedules everything, refers to everyone, receives documents from every direction, chases records it did not generate, manages chronic-disease recall across a whole panel, and absorbs the prior-authorization burden of every referral it makes. Nothing about that is clinically interesting, and all of it is expensive.

The scale of the desk work is not new information. Sinsky and colleagues, writing in Annals of Internal Medicine in 2016, observed that for every hour of direct clinical face time, physicians spent nearly two additional hours on EHR and desk work. That study should be cited with its limits every time: 57 physicians, 430 observed hours, four specialties, four states, with the authors themselves noting the data came from self-selected, high-performing practices and may not be generalizable. It is a decade old and it is not a current measurement. It is quoted here for one reason only — because it is one of the very few numbers in this field that can be chased to a named, dated, fetchable source, which is more than can be said for almost anything a vendor will show you.

On the authorization side, the AMA's 2025 Prior Authorization Physician Survey, fielded among 1,000 practicing physicians and released on May 13, 2026, reports that physicians complete an average of 40 prior authorizations per week, that prior authorization consumes an average of 13 hours of physician and staff time each week, that 40 percent of physicians employ staff dedicated exclusively to prior-authorization tasks, that 94 percent say it contributes to burnout, and that 74 percent report denials have increased over the past five years. That is a self-reported physician survey and should be described as one. It is still a real instrument with a real sample and a published release, which distinguishes it from most of what circulates.

We refuse statistics nobody can source, and we say so in the article. Naming a number as untraceable is more useful to an operator than repeating it. You will see us do this repeatedly below — with the vendor accuracy rates, with the per-no-show costs, and with the return-on-investment figure that turns out to have no origin at all.

— Frenchy Digital, house rule for the healthcare cluster

This article is the pillar of a ten-article implementation series. The nine specialty guides that follow it — dermatology, physical therapy, eye care, OB-GYN and the rest — assume this definition and do not contradict it. If you read only one section here, read the boundary table.

The single most common reason a practice cannot tell whether its AI deployment worked is that it never wrote down what things were like before. This is not a philosophical point. It is the difference between a renewal conversation you win and one you lose, and it takes two weeks.

Seven measures. Each has a method, an owner and a reason. Take them before you sit through a demo, because a demo changes what you think your problems are.

The Phase 0 baseline set for a primary care practice — Frenchy Digital engagement standard, 2026.

Two notes on method. First, take the measures with the crudest instrument that works. A tally sheet taped to the front desk for two weeks produces a more defensible number than a dashboard you configure for a month and never validate. Second, take them again at day 90 using exactly the same method. Changing the instrument between the before and the after is the most common way a genuinely successful project ends up unprovable.

There is one industry-scale figure worth knowing, and it comes with a disclosure. The Index report published by DataSpring — the organisation formerly known as CAQH, which converted from nonprofit to for-profit in January 2026 and rebranded on June 8, 2026, and is owned by shareholder companies affiliated with health plans — puts the remaining industry savings opportunity from full electronic transactions at $21 billion, on 2024 transaction data released in February 2026. That is an industry benchmark, not an independent one, and the ownership matters. We cite it with the disclosure attached, which is the only responsible way to use it.

This is the single most important architectural fact in medical practice automation, and most buyers discover it after the contract is signed.

Certified API access under 45 CFR §170.315(g)(10) is read and search only. ASTP/ONC's own Certification Companion Guide, updated May 15, 2026, states it without ambiguity:

Read services include those that allow authenticated and authorized third-party applications to view EHI through a secure API. 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.

— ASTP/ONC, Certification Companion Guide for §170.315(g)(10)

What that guarantees you, in any certified EHR, is the ability to read demographics, problems, medications, allergies, results, notes and coverage; to search; and to export in bulk. What it does not guarantee is everything that makes automation feel like automation: booking the appointment, writing the note, filing the document, updating the insurance record, placing the order, posting the payment. Each of those exists only if your EHR vendor chose to build it, chose to expose it, and chose to let your agent vendor use it. It is a commercial arrangement, not a certification right.

Do not overstate the constraint either. Epic voluntarily publishes some write-capable FHIR APIs, documenting create support on resources including DocumentReference, Observation, QuestionnaireResponse, AllergyIntolerance, Condition, Communication and BodyStructure, and update support on several of the same. The documented write surface is narrow and resource-specific, but it is not empty. The honest description is that write access exists in patches, at each vendor's discretion, and you must find out which patches you have before you design a workflow that depends on them.

Where the vendor cannot get an API write, the write happens by robotic process automation — a robot typing into the EHR's user interface on a service account. Notable is, to its credit, the one vendor in this category that says so plainly on its own integrations page, describing its method as using APIs, RPA, HL7 and more to get data into the right fields. That candour is a reason to trust the vendor and a reason to ask harder questions, both at once. RPA breaks on UI changes, is usually invisible in the EHR audit log as a system action, and attributes actions to whatever human's credentials it runs under — which collides directly with the requirement for unique user identification at 45 CFR §164.312(a)(2)(i), a required specification rather than an addressable one.

Two things follow, and they shape everything downstream. First: an agent can always decide and draft; it cannot always commit. Design for that asymmetry rather than against it. Second: read-only-by-default is not merely a compliance posture, it is the strongest available security control and it is free. An agent that cannot write cannot be injected into writing.

The trajectory is moving, slowly. ASTP/ONC's HTI-5 proposed rule, published in the Federal Register on December 29, 2025, retains §170.315(g)(10), states an aim to move beyond read-only interactions in future FHIR API requirements, and proposes to remove the third-party-seeking-modification-use condition from the Infeasibility exception — the condition EHR developers have used to limit write access. The comment period ran into late February 2026 and we located no final rule as of today. Write access is on the federal agenda. It is not a right today.

Sequence is the decision that most determines whether a zero-click programme succeeds, and it is the decision most often made backwards. Practices automate the workflow that annoys them most, which is almost always the phones. Phones are the worst possible starting point.

The correct ordering principle is reversibility first, visibility last. Start where the work is high volume and low variance, where everything the agent needs is already in the chart, where the output is a worklist rather than an irreversible action, and where failure means a human doing today's job manually. Finish with the workflows that patients see and that commit irreversible actions.

Recommended automation sequence for a primary care practice — Frenchy Digital, 2026.

Notice how much of this list runs on read-only data. That is deliberate. Every workflow in positions one through four can be built and proven before you have resolved the write-path question with your EHR vendor, which means you can be delivering value in week seven while the commercial conversation about write access is still running. Practices that sequence the other way spend their first two months in a procurement negotiation with nothing live.

This is the plan we run. Six phases, each with a named owner, an explicit entry criterion, an explicit exit criterion, and a written answer to the question every plan avoids: what do you do when this phase fails?

Ninety days is enough to get two workflows live and measured in a practice of five to twenty providers. It is not enough to rebuild the front office, and any plan that promises otherwise is describing a demo.

The 90-day zero-click implementation path — Frenchy Digital delivery standard, 2026.

Three details in that table are doing more work than they appear to. The first is shadow mode in Phase 2. Running the agent in parallel with the human for two full weeks, and reviewing disagreements by name rather than in aggregate, is the cheapest quality signal available and the one most often skipped under schedule pressure. The disagreements are where you learn that your workflow definition, not the model, is the problem.

The second is rehearsing the rollback in daylight before Phase 3 exits. A rollback that has never been practised is a document, not a control. Turn the workflow off deliberately on a quiet Tuesday, watch the manual path absorb it, time how long the switch takes, and write the number down. You will use it.

The third is that the write-path decision sits in Phase 4, not Phase 0. This is counterintuitive and it is intentional. Asking your EHR vendor the write question is easy; getting a written answer takes weeks and is easier to obtain once you are a live customer of an integration rather than a prospect. Do the read-only workflows first and let the commercial conversation run in parallel.

The boundary is not a philosophy. It is a table, it is signed by your medical director, and it is enforced in the software. Here is ours, with the authority for each line, because a boundary without a citation is an opinion that erodes under commercial pressure.

The human-in-the-loop boundary for administrative AI agents in a US medical practice, as of August 12, 2026.

Three lines deserve expansion, because they are the ones vendors push hardest against.

The administrative exclusion is statutory and it is broad. 21 U.S.C. §360j(o)(1)(A) excludes from the device definition software intended for administrative support of a health care facility, including the processing and maintenance of financial records, claims or billing information, appointment schedules, business analytics, information about patient populations, admissions, practice and inventory management, analysis of historical claims data to predict future utilization or cost-effectiveness, determination of health benefit eligibility, population health management, and laboratory workflow. Scheduling, eligibility, claims handling, referral logistics, inventory, population lists — an agent doing those things is outside the device definition by statute, not by FDA grace. This is the sentence that makes the zero-click clinic legally coherent.

The moment an agent recommends, four criteria apply. §360j(o)(1)(E) makes a recommending function non-device only if it does not acquire, process or analyze a medical image, an in-vitro diagnostic signal or a pattern from a signal acquisition system, and if it displays or analyzes medical information, supports or provides recommendations to a health care professional, and enables that professional to independently review the basis for those recommendations. The Clinical Decision Support Software guidance issued January 29, 2026 — which supersedes a January 6, 2026 version, which in turn replaced the September 2022 guidance — did not change that. Software providing a specific directive output still fails Criterion 3. What FDA announced is enforcement discretion over that failure. The guidance's own wording at page 10 turns on there being only one clinically appropriate option; at FDA's March 11, 2026 town hall the same policy was described with the word recommendation. Same policy, different noun — quote whichever document you are actually citing, and do not put the town hall's noun in the guidance's mouth. Write it as enforcement discretion, never as approval or clearance, because discretion is revocable without notice-and-comment and an exclusion is not.

The signature never moves. This is the cleanest rule in the entire regulatory stack, and CMS names AI explicitly. The Medicare Program Integrity Manual, chapter 3, §3.3.2.4, effective January 17, 2025, states that the treating practitioner's signature on a note indicates that the practitioner affirms the note adequately documents the care provided, and adds that this type of practitioner concurrence is also required when using Artificial Intelligence technology to capture the transcription of medical record entries. The same section makes the model attestation first-person and credentialed, with an express acknowledgement of administrative, civil or criminal liability, and forbids reviewers from considering an attestation statement from anyone other than the author of the record entry. There is no configuration in which an agent attests. An AI agent may draft, transcribe, suggest and assemble; the signature is the act that converts a draft into a billable, legally operative record, and it is personal, credentialed and non-delegable.

Before the table, the methodology — because the methodology is the reason to trust the table.

Top 10 AI agent platforms for medical practices, ranked on verifiable public attributes only. Checked 2026-08-12. Accuracy excluded from scoring by design.

The second half of the assessment is corporate. A platform that runs your front office is a continuity dependency, and the clearest cautionary tale in this category is not a product that failed but a product that was withdrawn for portfolio reasons after an acquisition. Read the ownership column as carefully as the security column.

Corporate record for the ranked platforms, from public announcements and dockets. Checked 2026-08-12.

Explaining the cuts matters as much as explaining the picks. Seven credible companies are frequently listed alongside the ten above and do not belong in a practice-facing table, each for a reason a buyer can verify.

Vendors excluded from the practice-facing Top 10, with the verifiable reason for each cut.

One more structural point that the Cohere row makes better than any argument could: the payer is automating too.A practice's prior-authorization agent is increasingly negotiating with another vendor's agent on the other side of the table. That does not make automation pointless — it makes the assembly-and-evidence half of the workflow more valuable, and the guess-what-the-payer-wants half less so.

In 2026 the EHR vendors moved into the agent layer, and that changed the buying question. It is no longer which agent platform. It is what does my EHR already ship, when, and at what price — and what is genuinely left over for a third party? Ask that first, because the answer may save you a six-figure procurement.

Native EHR agent offerings as documented on public record, checked 2026-08-12. Several primary sources in this table were blocked to our fetcher; treat the reported items as reported.

Epic's Phil Lindemann, quoted in MedCity News's March 2026 coverage of Agent Factory, described the ambition plainly: "Now with Penny, Art and Emmie, you can say, 'I want to redesign an entire process that involves scheduling, referrals and patient experience.'" The same article carries a useful counterweight from Canvas Medical's chief executive Adam Farren, who — in MedCity News's characterisation, not in his own words — applauds Epic's approach but does not think hospitals are ready to take advantage of the platform yet. What he actually said, verbatim: "Show what's possible, and the details matter, and often a narrow job accomplished with an agent is much more impactful than a more complex workflow with broader surface area for error."

That is the same conclusion this article reaches from a different direction, and it is worth sitting with: even the EHR-native path fails without workflow discipline. A narrow job done reliably beats a broad workflow with more surface area for error, whoever ships it.

Every deployment breaks. The difference between a practice that keeps its agents and one that rips them out is whether the break was detected by a metric or by a patient. Here are the eight failure modes we see in primary care, each with the signal that catches it early and the rollback that contains it.

Failure modes, detection signals and rollbacks for AI agents in a primary care practice — Frenchy Digital operations standard, 2026.

The fourth row deserves its own paragraph, because it is the one where the honest answer is uncomfortable. Prompt injection is not solved. Any agent reading untrusted external input — patient messages, faxes, referral packets, payer portal pages — is exposed, and the correct framing is blast-radius reduction, never prevention.

The best-sourced evidence is peer-reviewed and it is not comfortable reading. Lee and colleagues, in JAMA Network Open in December 2025, ran a controlled simulation across 12 clinical scenarios and reported that across 216 evaluations, attacks achieved 94.4 percent success at turn four and persisted in 69.4 percent of follow-ups, with extremely high-harm scenarios including FDA Category X pregnancy drugs succeeding in 91.7 percent of dialogues. Its limits must be stated: it is a controlled simulation rather than field data, the main experiment used lightweight models with a small proof-of-concept on flagship models, and three co-authors disclose company roles. Separately, Clusmann and colleagues in Nature Communications in 2025 showed that all four vision-language models tested were susceptible, and — the finding that breaks the standard mitigation story — that sub-visual prompts embedded in medical imaging data are non-obvious to human observers.

Read those two together and the conclusion is unavoidable: a human-in-the-loop who cannot see the injection cannot review it away. We found no published study of prompt injection via patient portal messages, referral faxes or payer portals in a live practice, and that absence is itself worth stating. What works is architectural and does not depend on detecting the attack: least privilege at the API boundary, read-only by default, the human signature as the only commit path, unique agent identity with audit controls so a successful injection is reconstructable afterwards, and out-of-band confirmation for anything irreversible. Any vendor quoting you an injection-detection accuracy rate is quoting a number it produced itself; no independent benchmark for clinical prompt-injection defence exists that we could locate.

Federal law draws its sharpest line not between AI and not-AI, but between administrative data movement and clinical recommendation. The state layer in 2026 adds a second axis: disclosure obligations that bind the practice, and decision-authority restrictions that bind payers. Know which is which, because vendors routinely sell you a payer obligation as though it were yours.

The disclosure and authority overlay for AI agents in a US medical practice, as of August 12, 2026. Verify against your own states before deployment.

Two federal points that consistently get misstated. First, the HIPAA Security Rule you are held to is the existing one. The notice of proposed rulemaking published January 6, 2025 remains a proposal; the Unified Agenda entry for RIN 0945-AA22 places it in long-term actions with final action projected for July 2027. So encryption at rest is still addressable rather than required, unique user identification is still required, and audit controls under §164.312(b) are still the standard your agent must satisfy. Build to the proposal anyway — none of it is bad engineering — but do not tell your board a proposed requirement is a current one.

Second, minimum necessary applies to your agent's context window.45 CFR §164.502(b)(1) requires reasonable efforts to limit PHI to the minimum necessary for the intended purpose, and the treatment exception at §164.502(b)(2) does not cover retrieval assembled for a billing, scheduling, prior-authorization or quality-reporting agent. Dumping the whole chart into the context window because it is easier is the specific practice this rule reaches. And on business associates: an API-based model provider processing PHI is a business associate, so is the agent-framework vendor in front of it, and so is the model provider's own cloud host as a subcontractor business associate under §160.103(3)(iii). A single BAA with the application vendor does not close the chain.

Here is the part vendors skip and the part that determines whether your programme survives its first staffing change. The zero-click clinic does not remove people. It changes what the people do, and the new work is harder than the old work.

Role changes in a zero-click primary care practice — Frenchy Digital engagement observations, 2026.

Two honest warnings. The first is that exception work is not residual work. When you automate the eligibility checks that resolve cleanly, what remains for your staff is a residue of cases that are all genuinely hard — no easy ones left to break up the day. We deliberately print no ratio for that split: nobody has measured it independently, and your own Phase 0 baseline will give you a better number than any benchmark on offer. That is a real change in cognitive load, and it is the reason the first month after go-live often feels worse than the month before it. Say so to your team in advance, and staff the queue with someone who is good at hard cases rather than whoever has capacity.

The second is that the daily queue review is a new job that nobody assigns.Somebody has to look at the exception queue every day, look at the weekly sample, and re-take the measures every month. In practices where this is nobody's explicit responsibility, the deployment does not fail loudly — it degrades quietly over about a quarter, and by the time anyone notices, the manual workarounds have re-established themselves and the software is a line item nobody defends at renewal.

We are not going to give you a return-on-investment figure, and you should be suspicious of anyone who does. The most-repeated ROI number in this market — a 468 percent return on an appointment-scheduling deployment — surfaces only in SEO content farms and has no origin at all. We chased it and found nothing behind it. It is a genuinely useful example of a statistic with literally no source.

What we can give you is the arithmetic to run against your own Phase 0 baseline. Payback in a primary care practice comes from four places, in descending order of defensibility.

• Overtime and temp-agency spend: The most defensible line and the easiest to measure, because it appears in payroll rather than in a model. Take the trailing twelve months, and at day 90 compare the same months of activity. This is the number that convinces a partner group.
• Redeployed capacity, valued honestly: Hours returned to staff are only worth money if those hours go somewhere that generates value or absorbs work you would otherwise have hired for. Value them at the loaded hourly cost of the role, not at a physician's rate, and only count hours you can name a destination for.
• Avoided rework at the front of the revenue cycle: Eligibility exceptions caught before the visit rather than after it. Measure the count from your baseline and the downstream denial rate for that cause. Note that we deliberately do not use the widely circulated per-rework cost figures — they are unsourceable, and your own denial data is better than a benchmark anyway.
• Throughput, only if you actually want it: Freed clinician or room time converts to revenue only if there is unmet demand and the clinicians want to see more patients. Both assumptions are frequently false. Test them before you build a case on them.

What we will not count, and neither should you: an economy-wide administrative-cost estimate presented as your practice's opportunity. The trillion-dollar figure that appears in vendor funding materials covers all administration across the entire health economy, most of it inside insurers and hospitals. It is not an addressable market for a six-provider clinic, and using it as one is how a business case becomes fiction.

Frenchy Digital cost bands for medical practice AI 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 expansion, incident response and a quarterly review. Every engagement carries a 30-day post-launch warranty, and you receive a written, fixed-price phased proposal within 5 business days of the discovery call. Full source-code and IP ownership transfers to your practice at delivery.

Twelve signals. None of them is proof of a bad product, and each of them should change the questions you ask next.

Vendor selection red flags for medical practice AI agent platforms — Frenchy Digital due-diligence standard, 2026.

The inverse is also worth writing down, because it is short. The vendors worth your time answer the write-path question plainly, name the EHRs they integrate with, publish a security attestation with a type and a date on it, will sign a BAA before the pilot, will let you export your audit logs, and will tell you which of your workflows depend on RPA. Six answers. If a vendor gives you all six in the first two meetings, you are talking to a serious company.

This section is the point of the article. Everything above is only as good as what sits behind it, and here is what does not.

• No independent evaluation of any product in this category exists: We searched on August 12, 2026 for randomized, peer-reviewed or third-party evaluations of multi-step administrative agent platforms in medical practices and found none. What exists in the literature concerns ambient clinical documentation, which is a different category. This is a negative finding, not a fetched fact, and it may change.
• No BAA disclosure from any vendor: Not one vendor in the ranked table publicly states that it offers a business associate agreement. We could not verify that any of them does or does not. Ask in writing.
• No prices, anywhere: We located no published starting price from any vendor in this category, native or third-party. Every pricing statement above reads contact sales because that is genuinely all that is public.
• Security posture we could not confirm: We could not verify SOC 2 Type II or HITRUST from a primary source for Hyro, Assort, Candid, Qventus or LeanTaaS, and Tennr's and Adonis's badges do not state a type. Several of these are discoverability findings rather than proof of absence — one vendor in this set publishes its attestation on a blog post rather than a security page, so site-search the domain before you conclude anything.
• Named integrations we could not corroborate: Tennr, Candid, Innovaccer, Qventus, LeanTaaS and Infinitus name no EHR at all on the pages we could read. Adonis's homepage logo strip is unlabelled and readings of it differ, so we printed the conservative floor. Re-fetch before relying on any count.
• athenahealth's no-additional-cost claim: athenaAmbient is reported to be delivered to athenahealth customers through routine updates at no additional cost. The primary press release was blocked to our fetcher and we could not verify it. For a small practice weighing a third-party platform this is the single most consequential fact in the article, and you must confirm it with your own representative.
• Epic's commercial terms: We could not verify Agent Factory's general availability date, any price for any Epic agent, or Epic's integration and Showroom fees from a primary source. Integration-cost figures circulating in agency blogs have no Epic source behind them and we refuse to print them.
• Colorado's interim status: SB 26-189 was signed May 14, 2026 and by its own Section 5 takes effect January 1, 2027, applying to consequential decisions made on or after that date, with a narrow set of rulemaking and appropriation provisions effective on passage. Our sources conflict on the operative status of the predecessor statute during the second half of 2026 and on whether the pause on its enforcement is an attorney-general commitment or a court-ordered stay. We state the January 1, 2027 date because it is quoted verbatim from the enrolled bill; treat the interim position as unsettled and take Colorado-specific advice.
• The 2026 state payer-side wave: A number of 2026 statutes restricting payers' use of AI in utilisation review are reported by law-firm trackers. We did not verify a single one of those bill numbers against primary text, and at least one identifier in the circulating list cannot be looked up as written. We name none of them here.
• Federal preemption: Federal preemption of state AI law is being actively pursued, but nothing has displaced the state statutes described above. We did not fetch the underlying executive order or bill and we cite no number, scope or provision for either.
• Litigation is alleged, not decided: The CommonSpirit and Fallon matters involving Innovaccer are pending. The underlying complaint is paywalled, we could not read it, and Innovaccer's response was not obtained. Treat both as allegations.
• Vendor status decays fast: Two companies in this category were acquired or wound down between 2023 and 2026, and one had its standalone product discontinued for portfolio reasons rather than performance ones. Anything in the ranked table could be stale within a quarter. The date on the table is not decoration.

None of this argues against building. It argues for building the measurement alongside the agent, sequencing by reversibility, holding the boundary where the law puts it, and being honest inside your own organisation about which numbers are yours and which came from someone selling you something. The practices that get durable value from AI agents are the ones that instrumented the before-state and never let a vendor define the scoreboard.