Top 10 AI Agents forVeterinary Practices in 2026

Veterinary medicine has exactly one independent benchmark of commercial AI, it was published in a peer-reviewed journal three weeks before this article, and not one vendor quotes it. Ma D, Faulkner JE, Stander N, Raisis A and Joslyn S published an external validation study in the Journal of the American Veterinary Medical Association, volume 264, issue 8, pages 1022 to 1029, with a print date of 1 August 2026 (DOI 10.2460/javma.25.10.0691). They collected canine abdominal radiographs from general practice whose diagnoses had been confirmed by surgery, necropsy, CT, ultrasound, cytology or treatment response, and submitted them to six commercial AI platforms between September and December 2024.

Fifty-three cases were selected. After platform rejections, 307 evaluations were available for analysis. The results, in the authors' own words, were that platform performance was variable and mostly low to moderate. Classification of radiographic findings showed sensitivity of 28% to 78%, F1 score of 28% to 51% and positive predictive value of 25% to 54%, which the abstract attributes to frequent missed diagnoses. And then the sentence that should stop a practice owner mid-scroll: small intestinal obstruction, a critical finding, was often not identified, with a sensitivity of 23% to 69%.

Diagnostic performance varied between the 6 AI platforms tested and was overall low to moderate for this small sample. Even the best-performing algorithm had notable limitations, and none appeared suitable for clinical use in their current form.

— Ma D, Faulkner JE, Stander N, Raisis A, Joslyn S — JAVMA 264(8):1022–1029, August 2026

Now the honest limits, stated in the same breath as the finding, because that is what separates this from a listicle. This is a pilot study. The sample is 53 cases, only two of which were non-pathological, and the authors say plainly in the clinical relevance section that further independent external validations on a larger scale and performance gains are needed before AI platforms can be safely integrated into clinical practice. It covers radiology only. And the identity of the six platforms could not be confirmed from the paper itself within this review, so we do not attach vendor names to those results and neither should anyone else. Naming a vendor as a poor performer on the strength of a third-party blog's guess at the platform list would be the single worst error available in this article.

Everything else in this category is measured by the people selling it. There is no independent evaluation of veterinary AI scribes, phone agents, scheduling agents or practice-management assistants anywhere we could find as of 23 August 2026. So when a scribe vendor tells you it saves thirty hours a month, or a phone agent vendor tells you what percentage of calls it handles, you are reading a number with no methodology, no denominator and no third party anywhere near it.

The second structural fact is that nobody is checking. FDA's own page on how it regulates animal devices, whose footer read content current as of 04/06/2026 when we opened it, states that the agency does not require submission of a 510(k), PMA, or any pre-market approval for devices intended for animal use. It adds that manufacturers who exclusively make animal devices need not register their establishments or list their devices, and are exempt from post-marketing reporting. In human medicine an AI product making a diagnostic claim usually lands in a device pathway with a clearance number you can look up. In veterinary medicine there is no premarket review at all, so there are no clearances to lean on. That absence is not an oversight failure; it is the design of the statute. But it means the checking is yours to do.

So this ranking does something different from the ones you have already read. It does not score accuracy, deflection, containment or hours saved, because every such number in this market is published by the seller about itself. It scores what you can verify without leaving your desk. If you want the cross-industry version of the same method, our review of AI agents across ten industries in 2026 applies it everywhere, and the pattern holds: the read surface is wide, the write surface is narrow, and the numbers that sell the software are the ones nobody has audited.

Every cell in the table below was read off a vendor's own page, a regulator's own page or a public filing on 23 August 2026, and every one of them is re-checkable at a URL in the sources list. Where a fact was not published, the cell says not publicly disclosed. We never estimated a price, never upgraded an unspecified certification, and never filled a blank with a plausible guess.

In In the Matter of Presto Automation Inc., Securities Act Release No. 11352, Exchange Act Release No. 102177, Admin. Proc. File No. 3-22413, entered 14 January 2025, the SEC addressed a company that had told investors its drive-thru voice AI delivered over 94% accuracy even in noisy environments and 95% to 99% automated order completion. The Commission found that the product lacked the capability to take orders on their own and required substantial human involvement, with human order takers located abroad, primarily in the Philippines and India, who processed the vast majority of drive-thru orders. An internal message quoted in the order put it more bluntly than any regulator would: the company was telling investors the product was running 95%-plus accuracy without disclosing that AI was doing none of the work.

Four points of precision, because they matter and are routinely mangled. Presto consented without admitting or denying the findings. The remedy was a cease-and-desist order with no civil penalty— it was not a fine. The findings are against Presto only; the order's "Supplier A" is Hi Auto, against whom the SEC made no findings whatsoever. And Presto Automation Inc. is not Presto Phoenix Inc. That case is the reason a published performance percentage, on its own, is worth nothing to a buyer. It is also why the veterinary radiology finding above matters so much: it is the one place in this market where somebody outside the vendors measured something.

Here is what we refused to print, named rather than quietly omitted, because a ranking that hides its exclusions is asking you to trust it rather than check it.

One more refusal deserves its own paragraph, because it is handled badly almost everywhere and the subject deserves better. Veterinary mental health is a serious, well-documented area of research and it is not a sales argument. Two real CDC-authored studies exist. Tomasi and colleagues, writing in JAVMA 254(1):104–112 in January 2019, examined death records for 11,620 veterinarians and found that among veterinarian decedents the proportionate mortality ratiosfor suicide were 2.1 for men and 3.5 for women, significantly higher than for the general US population. A proportionate mortality ratio is the share of deaths attributable to a cause compared with expectation. It is not a rate and not a risk ratio, and the authors explain why they used one: the non-decedent veterinarian population for that period was unknown. Writing "veterinarians are 3.5 times more likely to die by suicide" misstates the study. We will not do it, and neither should a vendor deck. The honest operational point is narrow: administrative load is one documented practice-related stressor among many, and no study shows that buying an AI product changes any mental-health outcome. If this is live for you or a colleague, the 988 Suicide and Crisis Lifeline is available by call or text, and the AVMA maintains wellbeing resources.

Read the second column first. Of the ten products below, not one documents an agent that acts alone on anything clinical. The honest description of this category in 2026 is scribes that draft, classifiers that flag, and schedulers that automate — with one vendor shipping a two-way confirmation flow it calls agentic, which confirms bookings and escalates everything else to a human. That is not a criticism. It is the correct architecture given a legal regime in which diagnosis and prescribing require a licensed veterinarian inside a relationship software cannot hold. Be suspicious of any vendor claiming more.

There is a reason for the gap, and it is not vendor negligence. In human healthcare, HIPAA forces disclosure. In veterinary medicine, as the next section shows from the regulation's own text, HIPAA does not apply at all — so nothing forces it. The buyer has to demand it contractually, and almost nobody does.

Each entry below states what the product does, what is verifiable, what is not disclosed, who it fits, who it does not, and who owns it. Where a vendor makes a performance claim, we attribute it to the vendor in the sentence rather than repeating it as fact.

The most consequential company in veterinary software is not in the ranking above, and the reason is worth more to you than another table row. IDEXX Laboratories owns three practice management systems — Cornerstone, Neo and ezyVet, the last acquired in June 2021 — plus Vello, the client-engagement layer launched on 15 February 2024. Vello is the IDEXX product that comes closest to appearing in an AI listicle, and its documented features are automated reminders and confirmations, online scheduling, a no-download mobile experience and two-way texting. That is classic automation, not an AI agent, and it is locked to IDEXX's own PIMS. Ranking it would have required us to do the thing this article exists to refuse.

IDEXX's real role here is larger than a product entry, and it runs through the next section: it owns the systems the other ten must integrate with, and it operates the partner gate that decides who gets to.

The rest of the exclusions, each with its reason:

• Covetrus — AVImark, ImproMed, eVetPractice: A PIMS and distribution business rather than an AI agent, and its ownership is mid-change. Cencora's own SEC-filed exhibit states that Cencora and Covetrus entered into a definitive agreement under which MWI Animal Health and Covetrus will merge, creating a combined animal-health platform. Announced and pending — not merged. If an agent you are buying writes into AVImark, you should know the PIMS owner is changing hands. Covetrus Connect, its permissions-based integration platform built on VetData, is the access path, and there is no public self-service API reference for AVImark that we could locate.
• ScribbleVet: Acquired by Instinct Science, announced 16 January 2026. Evaluate it as an Instinct product or not at all.
• Vetstoria, Kontak and WhiskerCloud: All three are PetDesk under Petvisor, renamed PetDesk Direct Booking, PetDesk Phones and PetDesk Marketing. Three dead names, one live company.
• TeleVet: Rebranded as Otto. Not a separate product.
• Hippo Manager: Acquired by Shepherd Veterinary Software in a deal reported March 2025. Not an independent product.
• Radimal, Picoxia and X Caliber Vet AI: These names surface in coverage of the independent radiology study, but we could not verify their 2026 corporate status, and we could not confirm the study's platform list from the paper itself. A vendor we cannot confirm still trades in 2026 does not go in a table, and a vendor we cannot confirm was in a study does not get attached to its results.
• CoVet, VetRec, HappyDoc, ScribeNote, PawfectNotes, Chckup, Airvet and NaVetor: All surfaced repeatedly in AI-scribe comparison content and none was verified in this review. Worth knowing why the comparison content is not enough: most of the veterinary AI scribe pricing pages that rank for these searches are published by vendors in the same category or by affiliate sites earning on referrals. We did not build a table from them and neither should you.

Naming the exclusions is not padding. A stale roster is the single likeliest failure mode in a ranking article, and in this vertical six of the names still circulating describe companies that no longer exist independently. If you are comparing this list against another one you found, check the acquisitions first — it will tell you in thirty seconds whether anyone re-checked anything.

HIPAA does not apply to your animal patients, and you can prove it from the regulation's own definitions in about four minutes. The chain lives in 45 CFR 160.103 and every link is quotable. Protected health information means individually identifiable health information. Individually identifiable health information relates to the past, present or future physical or mental health or condition of an individual. Individual means the person who is the subject of protected health information. And then the parenthetical that settles it:

Person means a natural person (meaning a human being who is born alive), trust or estate, partnership, corporation, professional association or corporation, or other entity, public or private.

— 45 CFR 160.103, definition of "person" (eCFR, read 2026-08-23)

A dog is not a human being who is born alive. So a dog is not a person, so it is not an individual, so its medical record is not individually identifiable health information, so it is not PHI. The same section defines health care as care, services or supplies related to the health of an individual — so veterinary care is not health care as HIPAA uses the term, and a veterinary practice is not a covered entity for its animal patients. There is no PHI, no covered entity and no business associate, so a BAA is legally meaningless for it.

Which makes the "HIPAA-compliant" badge on a veterinary landing page a category error: there is no protected health information in an animal record to be compliant about. Often it is not the vendor's fault — several of these companies also sell into human dentistry, optometry or medicine, where the language is correct and a BAA genuinely binds. But for you it is a trust signal that means nothing, and worse, it displaces the four questions that do mean something.

So what actually protects client and patient data? Four things, in descending order of usefulness to you. State veterinary practice acts and board rules are the primary regime — they govern record content, retention, confidentiality and release, with the familiar carve-outs for zoonotic and bite reporting, suspected cruelty, and subpoena or court order. Section 5 of the FTC Act, 15 U.S.C. §45, reaches a vendor that misrepresents its security, its AI capability or its data use — the same deception theory that underpins the SEC action described earlier. Contract is your main enforceable lever, and in the absence of HIPAA it is the only one you fully control. And state consumer-privacy statutes of general application reach your client'spersonal information — name, address, payment data — rather than the animal's clinical record, with most single-site practices falling below the revenue and volume thresholds anyway.

The second constraint is the one that decides whether an AI phone agent is a scheduling tool or an unlicensed practice of veterinary medicine — and both of the framings you have read elsewhere are wrong. There are two different veterinarian-client-patient relationships and they do different jobs.

The federal VCPR at 21 CFR 530.3(i) requires that the veterinarian has assumed responsibility for medical judgements, that there is sufficient knowledge of the animal to initiate at least a general or preliminary diagnosis, and that the veterinarian is readily available for follow-up. It then adds the sentence that does the work: such a relationship can exist only when the veterinarian has recently seen and is personally acquainted with the keeping and care of the animal by virtue of examination of the animal, or by medically appropriate and timely visits to the premises where the animal is kept. That cannot be met by telemedicine alone. But note what this definition governs: extralabel drug use under AMDUCA, and through 21 CFR 558.6 veterinary feed directives. It governs prescribing.

The state question is separate, and the Association of American Veterinary State Boards says so in its own model regulations, which we read in full on 23 August 2026 (approved by its Delegate Assembly in September 2023 — note the approval date, not the upload path):

The in-person VCPR as defined by the CFR under Title 21-Food and Drugs is not a federal requirement for the practice of veterinary medicine. Therefore, Jurisdictions may or may not require it and may modify it when requiring it to practice.

— AAVSB, Model Regulations: Establishment and Maintenance of a Veterinarian-Client-Patient-Relationship (approved September 2023)

So the correct two-part statement is this. The federal VCPR requires an in-person examination or premises visit, but it gates prescribing rather than the general practice of veterinary medicine. Whether an in-person VCPR is required to examine, diagnose and treat is a state question, and AAVSB reports that since the pandemic state legislatures have been moving in both directions — some confirming the in-person requirement, some explicitly removing it. An article telling you that federal law forbids remote diagnosis is wrong. An article telling you that nine named states now allow a virtual VCPR is also wrong to rely on, because that list is compiled and republished almost entirely by telemedicine vendors with a commercial interest in the count being high, and it changes. Check your own board. The AAVSB commentary also reports that in 2025 its Delegate Assembly passed Resolution 2025-3, indicating that its member boards do not support establishing a VCPR solely by virtual means, and resolving that the practice of veterinary medicine is deemed to occur where the patient is located.

One point of status precision that gets mangled constantly: FDA's COVID-era enforcement discretion is withdrawn, not expired and not still in effect. Guidance for Industry #269, issued March 2020, had temporarily suspended enforcement of the in-person examination and premises-visit element. FDA withdrew it, with the notice published in the Federal Register on 21 December 2022. Separately, the AVMA's Model Veterinary Practice Act was revised and adopted by its House of Delegates in July 2025 — but a model act is not law anywhere until a state adopts it, and it should never be described as a requirement.

The third constraint is the one nobody in this market talks about: nothing guarantees you API access to your own practice's records. In human medicine there is at least a floor — a certification programme, an information-blocking rule, and a certified read API every EHR must expose. Veterinary has none of it. No certification regime, no information-blocking rule, no conditions of participation, no baseline right of access. The floor is whatever your PIMS vendor's partner programme decides. That is a stronger constraint than the one facing a medical practice, not a weaker one — a comparison worth making directly against our ranking for medical practices, where certified read-only access at least sets a legal minimum.

ezyVet, owned by IDEXX, publishes the clearest developer documentation in the market, and reading it tells you exactly how this works. Access is partner-gated: a developer applies, is approved, and only then receives a partner ID and client credentials. Authentication is OAuth 2.0 client credentials, with access tokens carrying a 12-hour TTL. Read and write endpoints exist across contacts, invoices, animals, appointments and diagnostics. And there are published operating ceilings a busy practice should size against: a per-endpoint throttle most commonly set to 60 calls per minute, and a global rate limit set by default to 180 calls per database per partner, with a 429 response over the limit. Most tellingly, the commercial application form asks which endpoints the developer requires and whether they will be used to read or write— so write scope is granted per integration at the vendor's discretion, not by default. The form also asks how many clinic customers the applicant already shares with ezyVet, which is a commercial qualification and a real chicken-and-egg problem for a new entrant.

IDEXX also operates a Data Services layer for Cornerstone, described in its documentation as a standardised JSON layer letting third-party applications read and write with compatible practice systems over a single connection, with two sync modes: query-based or real-time, and batch, where batch data are refreshed from practice locations on an interval measured in tens of hours rather than minutes. We read that from the documentation site rather than from a signed integration agreement, so treat it as a question to ask rather than a fact to plan against — but ask it, because an agent on a batch tier is working from data that is potentially days old, and that is a failure mode for anything touching appointments, inventory or reminders. Covetrus runs the equivalent gate through Covetrus Connect, a permissions-based integration platform built on VetData, and we located no public self-service API reference for AVImark at all.

And behind all three constraints sits ownership. Ask who owns the software before you ask what it can do — in this market the two largest diagnostics companies also own the practice management systems, the client-communication layer and the partner programme that decides which competitors get API access.

Consider a four-doctor small-animal practice with six credentialed technicians, open six days a week. Two figures do the work, both from the US Bureau of Labor Statistics Occupational Employment and Wage Statistics release for May 2025. A veterinarian's annual median wage is $130,100against an occupation employing 83,900 people. A veterinary technologist or technician's annual median wage is $47,380 against an occupation employing 129,140. Veterinary assistants and laboratory animal caretakers, 126,580 of them, sit at an annual median of $38,150.

Convert those to an hourly basis at 2,080 hours a year, which is the crude but transparent way to do it: roughly $62.55 an hour for a veterinarian and $22.78 for a technician. That ratio — a veterinarian's hour is worth about 2.7 technician hours — is the single most useful number in this article, and no competing listicle contains it, because most veterinary AI is sold on the technician side of the practice while being priced against the veterinarian's hourly value.

Notice what this scenario did not need: no industry statistic, no missed-call benchmark, no shortage projection and no vendor ROI model. Two BLS wage figures and one published price were enough to produce a defensible business case in either direction. If a vendor needs a proprietary statistic to make its case, that is information about the case.

In this vertical the failures are almost never model failures — they are integration failures, consent failures and scope failures. Each row below pairs the failure with the signal that detects it early and the design decision that limits the damage. Build the detection before you build the automation; a failure you cannot see is a failure you will hear about from a client.

An agent may prepare, gather, draft and route. A licensed veterinarian must commit anything that enters the medical record, changes a clinical decision, or reaches a client as a statement about the animal. That single sentence is the whole compliance posture, and it follows directly from the VCPR analysis above rather than from caution for its own sake. The table below is the version to paste into a vendor evaluation and ask them to agree to in writing.

Two boundaries deserve emphasis because they are where vendors push. The first is triage: an agent applying written urgency criteria a veterinarian authored — this species, this age, this presenting sign, this slot — is executing a protocol. An agent deciding for itself how sick an animal sounds is practising. The difference is whether the criteria exist on paper before the call. The second is prescribing. The federal VCPR gates extralabel drug use and cannot be established remotely, so an agent must never generate, refill or queue a prescription, and controlled substances are categorically out of scope. Routing a client's refill request to the prescribing veterinarian is fine and useful. Everything past that is a licensed act.

For most single-site practices the right first move is not a custom build — it is buying one product with a published price and measuring it against your own baseline. Custom work earns its place when the workflow is genuinely yours, when the integration your practice needs does not exist, or when you are running enough locations that the per-seat maths stops working. These are Frenchy Digital's bands. They are ours, not an industry benchmark, and we publish them for the same reason we scored pricing transparency: a number you can see is worth more than a number you have to ask for.

Senior-led delivery runs $150 to $225 per hour and retainers run $2,500 to $9,500 per month. Every engagement carries a 30-day post-launch warranty, full source-code and IP ownership transfers to you, and you receive a fixed-price phased proposal within 5 business days of the discovery call. We are a senior-led Black-owned agency in Los Angeles; you can reach us on +1 (424) 272-5601 or book directly at calendly.com/frenchydigital/discovery-call. If you want the detail on how we scope and build this kind of work, it is set out on our AI agent creation service page.

One sequencing note that saves practices real money. Do the integration check before the build decision. A discovery engagement that establishes what your PIMS vendor will actually grant — real-time or batch, read or write, which endpoints, at what rate ceiling — frequently changes the answer from build to buy, or from one vendor to another. In a market with no interoperability mandate, the integration question is not a technical detail at the end of the project. It is the first question, and getting it wrong is the most expensive mistake available.

Every flag below came out of this research rather than out of general advice, and each one is checkable in a demo or a single email. None of them is disqualifying on its own. All of them are questions you are entitled to a crisp answer to.

1. 1.A HIPAA badge on a veterinary landing page: Ask which of your records they believe is protected health information. HIPAA's own definitions limit it to a natural person, meaning a human being who is born alive. Then ask for the SOC 2 Type II report, the DPA, the retention schedule and the subprocessor list instead.
2. 2.A certification that belongs to the cloud host: Read the sentence the certification appears in. If it describes the hosting provider's assessments, the vendor's own certification status is not publicly disclosed, whatever the logo suggests. This is the most common compliance error available in this market and it takes thirty seconds to catch.
3. 3.A product name that implies clinical authority: Diagnose, Assess, Triage. Almost always the underlying feature is lawful decision support. Ask which outputs are recommendations to a clinician and which are ever shown to a client, and put the answer in the contract rather than trusting the product name.
4. 4.An accuracy, deflection or time-saved percentage with no study behind it: The one time a federal regulator audited a vendor's published AI performance metrics, it found a product that required substantial human involvement behind numbers of 94% and higher. Ask for the methodology, the denominator and the date. If there is no study, the number is a marketing asset.
5. 5.No published price: Twelve of the thirteen companies reviewed here require a sales conversation before you see a figure. That is normal in this market, which means it is not a scandal — but it does mean you should ask for the full price card in writing, including the AI add-ons, before the pilot rather than after.
6. 6.No reachable security or trust page: Several vendors' security URLs returned 404 on the day we checked. A certificate may well exist behind a sales NDA. Ask early, get the report before the pilot, and put the certification and its scope in the contract.
7. 7.Integration described as "we connect to your PIMS": That sentence covers a supported API with write scope, a nightly batch file, and screen-scraping. They are three different products with three different failure modes. Make them name which one, for your PIMS, for your practice.
8. 8.The vendor is owned by your diagnostics supplier: Not disqualifying, and often it buys you genuinely deeper integration. But ask directly about the integration roadmap for competing laboratories, and what happens to your data and workflow if you switch lab providers.
9. 9."Agentic" applied to a two-way SMS template: Ask precisely what the system does without a human present, and what it does when it encounters something outside its script. If the answer is that it creates a task for staff, that is a good design and a bad use of the word.
10. 10.A workforce-shortage statistic on the first slide: The most-quoted shortage projection in this profession is published by a company that owns the largest veterinary employer and a diagnostics business. Two other credible bodies publish irreconcilable numbers, one of which is no dire shortage. A vendor leading with it is telling you how their case is built.
11. 11.Anything that touches a prescription refill: The highest-risk automation in this vertical. Routing a request to the veterinarian is fine. Generating, approving or queueing a refill is a licensed act inside a relationship the federal rule says cannot be established remotely, and controlled substances are categorically out.
12. 12.A pitch that needs the human out of the loop to work: If removing veterinarian sign-off is what makes the ROI model close, the model is wrong. The commit step takes seconds; the labour saving lives in everything before it.

Here is what this article does not know, stated plainly, because a ranking that refuses vendor claims and then hides its own gaps has learned nothing from the exercise.

• Every compliance and pricing cell is a timestamp, not a permanent fact: All of them were read off vendor pages on 23 August 2026. Vendors change pages without notice. Each cell is re-checkable at the URL in the sources list, and you should re-check before you sign rather than trusting a table you found in an article — including this one.
• We could not confirm the six platforms in the independent radiology study: The platform names circulate in third-party coverage of the paper. We did not confirm them from the paper's own methods section, so we attach no vendor name to those results anywhere in this article. If you are evaluating a radiology AI product, ask the vendor directly whether it participated and what it makes of the findings.
• No SOC 2 Type II, ISO 27001 or published DPA was confirmed on any vendor's own trust page: That is a finding about disclosure, not proof of absence. Several vendors may hold certificates behind a sales NDA. What we can report is what is published, and in this category very little is — which is itself the practical consequence of HIPAA not applying.
• No data residency or retention commitment was located for any vendor: For products processing exam-room audio, and for at least one vendor headquartered outside the United States, that is a real gap and a fair contract demand.
• Ownership is unresolved for two companies in the table: One vendor's parent-group status and another's ownership rest on aggregator profiles rather than primary records, so we wrote the uncertainty rather than the claim. Two more corporate changes are pending rather than closed and are described that way throughout.
• We did not verify state-by-state telemedicine rules: The nine-state virtual-VCPR list that circulates is vendor-compiled and volatile, and we do not print it. We also did not read California's telemedicine amendments in their enrolled form, so no state-specific prescribing limit appears here. Check your own board and your own practice act before deploying anything client-facing.
• We did not verify current DEA controlled-substance requirements in this pass: Those rules changed repeatedly and mostly concern human medicine, so we write nothing about them from memory. The safe design rule stands on the federal VCPR alone: an agent must never generate, refill or queue a controlled-substance prescription.
• There is no FDA guidance regulating veterinary clinical AI: We located none as of 23 August 2026. An FDA CVM staff review article published in the American Journal of Veterinary Research in March 2025 describes the Center's internal adoption of AI for its own regulatory work — antimicrobial-resistance genomics, safety surveillance and IT modernisation. It is not guidance, it creates no obligation for vendors, and citing it as FDA's rules for veterinary AI would misdescribe a review article as regulation.
• No veterinary-specific AI statute was located in any US jurisdiction: General-purpose state AI laws may reach some of these products depending on how they are used and where. We did not survey them, and we will not list state requirements from memory.
• The only independent evaluation in this vertical is radiological, and it is a pilot: Fifty-three cases, only two of them non-pathological, and the authors ask for larger validations in their own clinical relevance section. It is the strongest evidence this market has and it is still a pilot study. Both halves of that sentence are true and we print both.

None of this argues against buying. It argues for buying the way you would buy a new analyser or a new anaesthetic machine: measure the baseline first, run one workflow, keep the licensed human on the commit step, put the compliance answers in the contract rather than the sales email, and re-check the vendor's published facts at renewal. If you want the operational build-out rather than the vendor comparison, we work through the same workflows in more depth in our guide to AI agents for veterinary practices.

And the boundary holds throughout. An agent reads, gathers, drafts, checks and proposes. A licensed veterinarian commits anything that enters the record, changes a clinical decision, or reaches a client as a statement about their animal. That is not caution for its own sake — it is the only architecture that survives a prompt injection, a board inspection and a bad Monday, and it happens to be the one that actually ships.