AI Agents for Veterinary Practicesin 2026

Start with the fact that governs everything else in this article, because almost nobody selling into your practice will lead with it. There is no FDA premarket approval requirement for medical devices intended for animal use. A veterinary AI product can reach the market with no validation study, no disclosure of what it was trained on, and no accuracy reporting whatsoever. And no state veterinary board rule specifically authorizes or prohibits the use of AI in veterinary diagnosis.

Hold that next to human medicine. A comparable diagnostic product for people runs through an FDA clearance or approval pathway before it can be sold. Certified health IT developers publish standardized transparency attributes for predictive features. Several states now require a clinician to disclose to the patient that AI was used in diagnosis or treatment. None of those mechanisms exist on the veterinary side. Not weaker versions — none.

This is documented rather than inferred. The Journal of Veterinary Internal Medicine and the AVMA's own review of the ethical and legal implications both describe the same gap.

The regulatory asymmetry between human and veterinary AI products, and where the evaluation burden lands as a result.

This is not an argument against buying. Some of these products are genuinely good, and a few vendors hold themselves to a higher standard than any rule requires. It is an argument against buying the way most practices buy software — on a demo, a testimonial, and a monthly price. The absence of a regulatory floor means the questions you ask are the only quality control in the system.

It also means one thing does not change: liability. The licensed veterinarian remains professionally responsible for the outcome. That part is identical to human medicine. What is missing is everything that ordinarily sits upstream of the clinician to reduce the chance that a bad product ever got in front of them.

When nobody upstream is checking, the questions you ask in procurement are not diligence theater. They are the entire quality system.

— Frenchy Digital buying principle

Two documents are worth reading in full before you buy anything, not because they bind you but because they are the only serious published positions in the profession.

The first is the American Association of Veterinary State Boards whitepaper, Regulatory Considerations of the Use of Artificial Intelligence in Veterinary Medicine, published in 2025. It recommends four things: transparency about the datasets a model was trained on, human-in-the-loop verification of AI output, safeguards for client data privacy, and informed consent where the risk to the patient is meaningful. It names the risks explicitly — hallucinations, training-data bias, automation bias, and the absence of premarket approval.

The second is the California Veterinary Medical Association's formal AI policy. Its core position is a single sentence worth memorizing: the veterinarian is the authority for making decisions, establishing treatment protocols, and making diagnoses and recommendations. AI supports that judgment. No current technology replaces it.

Neither is a rule. Neither carries an enforcement mechanism. But an association whitepaper that four state-board members would recognize is a much better procurement checklist than anything a vendor will hand you, and it has the advantage of not being written by someone trying to sell you something.

The AAVSB's 2025 recommendations converted into a buying checklist — Frenchy Digital, 2026.

One note on the AAVSB document's date: it was published in 2025, and published references to the specific month conflict with one another. The content is what matters, and the content has not changed.

If you have sat through a veterinary AI pitch in the last two years, you have seen this number: the United States needs up to 55,000 additional veterinarians by 2030. It comes from a Mars Veterinary Health analysis published in August 2023. It is presented as settled. It is not.

The AVMA commissioned Brakke Consulting to test it. The resulting analysis, published in October 2024, forecast to 2035 and concluded that graduates from existing colleges are likely enough to meet demand across that horizon. Its analyst's verbatim conclusion is worth quoting exactly, because paraphrases of it have been used in both directions.

The projections in this analysis do not justify a conclusion of overall excess capacity or capacity shortage by 2030 or 2035.

— John Volk, Brakke Consulting, via AVMA (October 2024)

The methodological criticism is specific: the earlier work relied on simple linear projection without modelling how supply and demand respond to each other. That is a fair critique of a forecasting method, not an accusation of bad faith — and it does not by itself prove the opposite case either.

A third study, commissioned by the AAVMC, lands between the two: 70,092 new veterinarians needed through 2032 against 52,926 projected graduates, a shortfall of 17,166. A real gap — and roughly a third the size of the figure in the vendor deck.

Three published positions on veterinary workforce supply, plus the labor-market data that has to fit alongside them.

Now the numbers that make this concrete. Per AVMA data, veterinarian unemployment is 0.7% and average real starting salary is $129,000 — a tight market on any reading. And yet about 7% of 2025 graduates received no offers at all, while visit volume has fallen four consecutive years.

Brakke also flagged thirteen prospective new veterinary colleges seeking accreditation, representing roughly a 40% capacity increase over ten years — though their graduates would account for only about 4% of practicing veterinarians by 2035. Supply is moving. Slowly, and behind the debate.

What this means for a practice owner is narrow and useful. If you cannot fill a role, you have a local hiring problem inside a national market that is not obviously short of people. Compensation, schedule design, and the quality of the working environment move that needle. Software does not. Where software helps is by reducing how much administrative work each person carries — which is a real benefit, honestly stated, and a different claim from solving a national shortage.

The reason this article exists in 2026 rather than 2023 is economic. Practice volume has been declining for four consecutive years, and the shape of the decline tells you what to buy.

Veterinary visit and revenue trends, 2022–2025, with the 2026 forecast. Visit data per Vetsource; forecast per Frontiers in Veterinary Science, October 2025.

Read the last two rows together, because that is the whole story. Transaction volume fell 4.7% nationally in 2025 while revenue still rose 2.6%. Practices are extracting more per visit. That works until it does not, and the 3.8% drop in wellness visits is the early warning — wellness is the category clients defer first and the one that feeds everything downstream.

Cornell economists put a frame around it. Writing in Frontiers in Veterinary Science in October 2025, they found the veterinary economy entered a recessionary phase in late 2024, with negative growth expected to persist through mid-2026 and a possible recovery later in the horizon. The traditional belief that veterinary medicine is recession-proof does not survive the data.

The confusing part is that the surrounding market looks healthy. The APPA reported $158 billion in US pet industry spending in 2025, up 3.7%, with $165 billion projected for 2026, and dog ownership expanding from 51% to 53% of US households. More pets, more money in the category, fewer visits to the clinic. Spending is moving somewhere other than your exam room.

It also changes the payback math. If you cannot count on volume growth, the return has to come from cost avoided or revenue recovered on visits you would otherwise have lost. Both are measurable. Measure them before you sign, so you have a baseline to compare against — the single most common reason practices cannot tell whether a tool paid for itself is that nobody wrote down the before-state.

Strip away the category talk and one segment of veterinary AI is genuinely mature, competitively priced, and easy to evaluate: ambient documentation. A scribe listens to the consultation and produces a draft note the veterinarian edits and signs. There are roughly twelve credible tools, prices are public, and most offer trials.

Veterinary ambient scribe pricing as published by vendors, 2026. Confirm current pricing directly — this segment moves.

The pricing model matters more than the headline number. A per-DVM price like ScribbleVet's $150 per doctor per month on annual billing scales linearly with your team. A flat price like HappyDoc's roughly $119 to $149 per month with unlimited users does not. For a solo practitioner the per-DVM option can be cheaper; for a six-doctor hospital the flat option is a different order of magnitude. Run the arithmetic at your own doctor count over 24 months before comparing anything else.

Two structural notes. ScribbleVet was acquired by Instinct Science in January 2026 — worth raising in a renewal conversation, because acquisitions change roadmaps and contract terms. And Digitail's Tails AI is native to a practice-management platform that publishes public API documentation, which means the integration question that dominates every other scribe evaluation simply does not arise.

One boundary that does not move: the draft is not the record. A veterinarian edits and signs, every time. That is not caution for its own sake — it is the AAVSB's human-in-the-loop verification recommendation applied to the one workflow where it is easiest to let slide, because the drafts are usually good and reviewing them starts to feel like a formality. That feeling is automation bias, which the AAVSB names by that exact term.

Here is the constraint that decides more veterinary AI outcomes than model quality ever will: whether your practice-management system will let anything write into it.

Every agent worth having eventually needs to read from and write to your PIMS. Reading is usually solvable. Writing is where projects stall, budgets grow, and vendors get vague. The system you already run has largely decided how expensive your next three years of automation will be — and you probably chose it for reasons that had nothing to do with this.

Practice-management platform integration posture as it affects third-party AI. Verify against your own version and hosting arrangement.

Digitail is the most open environment in this market: public API documentation and more than fifty integrations. If you run it, third-party integration is a configuration exercise. ezyVet, owned by IDEXX, is the customizable cloud option — real work, but a supported path. Shepherd is cloud-native and integration-friendly. AVImark supports external tools, with the usual caveat about verifying your specific version.

Cornerstoneis the closed one. Its API was not built for third-party AI integration, and the AI capability available within it is largely IDEXX's own — VetConnect PLUS and Panorama. This is not a defect so much as a strategy, and it is a coherent one. But it changes your options materially. If you run Cornerstone, expect most third-party agents to operate alongside the system rather than writing into it, expect exports and manual steps in the workflow, and expect the integration line on any quote to be the largest one.

One legitimate answer is that there is no API write path and the workflow ends in copy-paste. That can still be worth buying — a scribe that saves eight minutes of typing is worth something even if the last step is manual. Just price it as a manual step, count the seconds honestly, and do not let it be discovered after the contract is signed.

Radiograph interpretation is the highest-profile veterinary AI category and the one where the absence of a regulatory floor bites hardest, because it is the category where being wrong has clinical consequences and where almost every published number originates with the vendor selling the product.

Veterinary imaging AI vendor claims. All figures in the second column are vendor-supplied, reported via trade sources, and are scale rather than accuracy measures unless noted.

Be precise about what those numbers are. SignalPET reporting 2,300-plus clinics and more than 50,000 films per week is an adoption figure. Vetology reporting a 6.6 million image bank and 89-plus classifiers is a scale figure. Neither says how often the software is right, and neither is independently audited. They are useful — a product running at that volume has survived contact with real clinics — but they are not performance data, and a deck that presents them as though they were is doing something you should notice.

Antech and IDEXX's RapidRead take a different approach, pairing AI output with board-certified radiologists. That is the human-in-the-loop architecture the AAVSB recommends, implemented at the product level rather than left to the practice. Worth asking exactly what the radiologist reviews, on what timeline, and whether that changes at higher volumes.

If you take one thing from this section: in imaging, ask for published metrics and a described test set, and treat the answer as diagnostic of the vendor. Whether a vendor has the data is a technical question. Whether they will show it to you is a character question, and in an unregulated market the second one is the more useful signal.

Before you design any client-facing agent, understand the one hard federal constraint in this space. The FDA's definition of the veterinarian-client-patient relationship requires a physical examination of the animal or a medically appropriate visit to the premises where the animal is kept. It cannot be satisfied by telemedicine alone. The enforcement discretion extended during COVID has been withdrawn.

VCPR constraints affecting client-facing veterinary automation. Federal position per AVMA guidance on telehealth and the VCPR; state positions vary and change.

Telemedicine maymaintain an existing VCPR. That is the useful opening: rechecks, post-operative monitoring, chronic-disease follow-up, and medication questions for a patient already seen in person. Several states permit a telemedicine-established VCPR under their own practice acts — Arizona, Idaho, New Jersey, Vermont and Virginia among them, with Arizona's 2025 law capping prescriptions at 14 days with one refill.

What this rules out is narrower than people assume, and it is worth being specific. An agent cannot establish a VCPR, cannot conduct a virtual consultation that stands in for an exam, cannot advise on whether a patient needs to be seen, and cannot participate in prescribing. What an agent can do is everything on the administrative side of that line: booking the in-person visit that establishes the relationship, collecting history before the appointment, assembling the record for a follow-up, handling the reminder sequence, and routing an urgent call to a human immediately rather than assessing it.

That distinction — administrative versus clinical — is the same line the CVMA policy draws and the same line the AAVSB's human-in-the-loop recommendation implies. It is also, conveniently, where most of the actual time savings are.

A short section, and a deliberately careful one. Consolidation matters commercially, but there is no primary source for how much of the market corporate groups hold. Circulating figures range widely across broker and marketing sites with nothing underneath them, so no percentage appears in this article. What can be stated is who the consolidators are and why the distinction changes a purchase.

• Mars Veterinary Health: Operates VCA, Banfield, and BluePearl. The largest consolidator, and the only major one that is not private-equity backed. Also the publisher of the 55,000-veterinarian projection discussed above — worth knowing when you read it.
• NVA: More than 1,400 hospitals. Acquired by Ethos Veterinary Health from JAB Holding on July 31, 2025.
• Thrive Pet Healthcare: Roughly 380 clinics, backed by TSG Consumer Partners.
• Mission Pet Health: Formed in late 2024 from the combination of Southern Veterinary Partners and Mission Veterinary Partners, operating across 41 states.

The commercial consequence is a difference in how software gets bought. Corporate groups buy practice-management systems centrally and standardize across hospitals. One decision, one contract, one configuration, rolled out. Independents buy per clinic, on their own timeline, against their own P&L. Two entirely different sales motions, and two different buying experiences.

If you are an independent, this is why some products feel like they were not designed for you: they were designed for a central IT function that does not exist in your building. Ask specifically what onboarding looks like for a single-location practice without dedicated IT, and who you call at 7am when it breaks.

If you are evaluating on behalf of a group, the calculus inverts. Central standardization is your leverage on price and your risk on fit — a configuration that works in a high-volume urban hospital may be wrong for a rural mixed-practice location. Per-hospital override capability is worth negotiating for explicitly, before the contract, because retrofitting it later is a product change rather than a setting.

Everything below is administrative automation under human review. None of it makes a clinical decision, and none of it removes a veterinarian from a judgment call. That constraint is not a hedge — it is what makes these workflows defensible in a profession where the only published policies say the veterinarian is the decision authority.

Defensible veterinary agent workflows and their required human checkpoints — Frenchy Digital, 2026.

Three of these deserve emphasis given the economics described earlier. After-hours call handling matters because a missed call from a client with a sick animal is a lost visit and often a lost client, and in a market where transactions are falling 4.7% a year the visits you already had are the cheapest ones to keep. The rule is absolute: capture and route, never advise. If a caller describes something urgent, the agent connects them to a human immediately rather than assessing severity.

Wellness reminder recovery targets the 3.8% decline directly. Most practices have a reminder system that fires on schedule and stops. An agent can work the harder cases — the client who has drifted eight months past due, whose last three reminders bounced, whose dog is on a chronic medication that has not been refilled. Personalized, grounded in actual visit history, drafted for a human to approve. That is recovered revenue on patients you already have.

Records requests and referrals are pure administrative drag: assembling history, chasing the specialist, tracking whether anything came back. Nobody enjoys it, it does not require clinical judgment to assemble, and it is one of the clearest hour-for-hour returns in a practice. The letter is still signed by a veterinarian.

One security note that applies to any agent reading content from outside your practice — client emails, portal messages, forwarded records, referral letters. Prompt injection is an unsolved problem. Text inside an inbound document can read as an instruction to the model. There is no complete defense today, so the discipline is blast-radius reduction: allowlist what the agent can do, deny by default any action a human did not request, and never give an agent that reads untrusted content the ability to write to a record or send a message without review. The OWASP Top 10 for LLM Applications is the right checklist to hold a vendor against.

Since no regulator does a first pass, you are it. Ten questions, sent in writing before the demo rather than after the pilot. The answers separate vendors who built for veterinary practices from vendors who built for a pitch meeting.

Frenchy Digital veterinary AI due-diligence question set, 2026.

Two of these carry more weight than the rest. The first is the training-data question, because it is the AAVSB's leading recommendation and because a model trained overwhelmingly on one species, one breed profile, or one equipment type will underperform on a caseload unlike its training set — and you will not find that out from a demo. The second is the write-path question, because it determines the true cost of the project and is the claim most often overstated.

Add one artifact request to the list: ask for a redacted export of a single day of audit logs from an existing customer's deployment. It answers more than a security questionnaire, because a vendor who cannot produce it does not have the logging, whatever the questionnaire said. In a market with no premarket review, the ability to reconstruct what the system did is the only evidence you will ever have.

Finally, run a real trial with a real baseline. Two weeks, one workflow, one measured before-state. Impressions from a demo are worth very little; ten timed consultations are worth a great deal. This is the part practices skip, and it is the part that would catch nearly every bad purchase.

None of these are hypothetical. Each one has surfaced in a real evaluation, and several of them are common enough to be predictable.

The Frenchy Digital red-flag list for veterinary AI buyers, 2026.

The first row is the one to internalize, because it is the most common and the most revealing. A vendor claiming FDA approval for a veterinary AI product is claiming something that does not exist — there is no premarket approval requirement for animal-use devices, so there is no approval to have received. Sometimes it is loose language about a human-medicine product line. Sometimes it is not. Either way it tells you how carefully this company speaks about regulatory matters.

A vendor who will not put training-data description, integration write path, data retention, and audit-log export into a contract before you sign will not put them into the product after you sign either.

— Frenchy Digital buyer’s principle

Most practices should buy an ambient scribe rather than build one. That market is competitive, priced between $40 and $450 a month, and there is no advantage in reinventing it. Custom work makes sense when the workflow is specific to how your practice runs, when the integration is the hard part, or when you are a group standardizing across locations and the off-the-shelf configuration does not fit.

Frenchy Digital cost bands for veterinary 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 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 that changes decisions. The integration work is largely a fixed cost paid once and reused by everything after it. The first agent carries the PIMS connection, the review queue, the logging, and the training. The fourth inherits all of it. This is also why the discovery band exists as a separate engagement — on a closed PIMS the honest recommendation is sometimes to buy an off-the-shelf tool and stop, and it is cheaper for everyone to find that out in week three than in month five.

On payback: with visits declining, build the case on hours returned and revenue recovered rather than on additional appointments. Staff time per visit, missed-call rate, lapsed-reminder recovery, and days from consultation to signed note are all measurable before and after. If a vendor or an agency will not help you define those measures up front, that is its own answer.

A careful evaluation makes a purchase defensible. It does not make it effective, and there are real reasons a veterinary AI deployment underperforms. If you are building a business case, build it on this.

• The evidence base is thin, and that is a fact about the market rather than a criticism of it: With no premarket approval requirement and no disclosure obligation, there is very little independent performance data on veterinary AI products. Vetology publishing metrics across 300,000 test cases in early 2026 made news precisely because it was the first. Do not mistake absence of published failures for evidence of success.
• Vendor scale figures are not accuracy figures: Clinic counts, films per week, image-bank size, and classifier counts describe adoption and scale. None of them tell you how often the software is right on a case like yours, and none are independently audited.
• Training-set mismatch is invisible until it is not: A model trained largely on one species mix, breed profile, or equipment type will behave differently on your caseload. Automation bias makes this harder to catch, because a plausible-sounding output is easier to accept than to question.
• Integration is where projects die: The binding constraint in this industry is practice-management API openness, not model capability. On a closed system the integration cost can exceed the software cost, and it is the line item most often underestimated at signing.
• Prompt injection is unsolved: Any agent reading untrusted external content — client emails, forwarded records, referral letters — is exposed. The mitigation is blast-radius reduction, not prevention: restricted tools, deny by default, human review before anything is sent or written.
• Software does not create demand: Visits have fallen four consecutive years and Cornell economists expect negative growth through mid-2026. An agent can reduce cost per visit and recover visits you were losing. It cannot make clients spend more on veterinary care.
• Staff adoption is the quiet failure: A tool nobody was trained on becomes a tool nobody uses, and the subscription renews anyway. Budget training time explicitly, name an owner in the practice, and check adoption at 30 and 90 days rather than assuming it.
• The liability does not move: The licensed veterinarian remains responsible for the outcome regardless of what the software suggested. No vendor contract changes that, and in a market with no regulatory floor, no clearance sits behind the product to share the weight.

None of this argues against buying. It argues for buying narrowly, measuring honestly, and keeping a veterinarian between the software and the patient. The practices that get value from these tools are the ones that instrumented the before-state and picked one workflow at a time.

And the boundary holds regardless: these are administrative and documentation systems operating under human review. They do not diagnose, they do not treat, and no amount of architecture makes it appropriate for one to practice veterinary medicine.