Adam Wysocki on the 2026 Digitail and AAHA Veterinary AI Study

In 2024, we at Digitail released a study on artificial intelligence in veterinary medicine. A lot has changed since then — so this summer, we conducted a new industry-wide survey to dig even deeper. Together with the American Animal Hospital Association (AAHA), we evaluated how veterinary professionals perceive and use AI in practice, and tracked what’s changed over the past two years. This year’s study also introduces a new measure of AI maturity, assessing how deliberately practices adopt, govern, and integrate AI.

You can download the full report at this link, and review the key insights from the study.

Adam Wysocki, founder of VetSoftwareHub, an independent veterinary software directory, and creator of the PIMS Selection Navigator, contributed the Discussion section of this study as an independent expert.

A bad system will beat a good person every time.
W. Edwards Deming, 1993

Two years ago, this survey asked whether the profession would try AI. The 2024 answer was a cautious maybe. Fewer than two in five were using it at work, sentiment leaned optimistic but not by much, and veterinary professionals under the age of 30 were the most skeptical group. The 2026 wave asked a different question and got a different kind of answer. Use is no longer the story. Management is.

That shift is the focus of this discussion. Adoption more than doubled, and most of the people using AI now use it every day. Every role, every age band, and every practice type sit in a narrow range. The gaps that structured the first paper have closed. What hasn’t closed is the gap between using a tool and running a system. Nearly half of AI-using practices have no plan for it and nobody responsible for it. Governance is the weakest dimension on the maturity index. Two thirds bolted AI onto existing work rather than redesigning the work around it. And the practices that did the unglamorous work of organizing around AI reported roughly six times as many business outcomes as the ones that didn’t.

The profession has finished the easy part. The rest of this discussion is about what that means and what should happen next.

From curiosity to habit

In 2024, familiarity predicted almost everything that mattered. Optimism, personal use, professional use, intent to adopt, belief in a competitive advantage. That finding was useful and, in hindsight, incomplete. Familiarity is what you measure when a technology is still optional. Once four in five people are using it, the more interesting question is whether the organization around them can keep up.

The 2026 data show how completely the optional phase ended. Occasional use all but disappeared. Non-adopters stopped hedging. Two years ago, nearly half of them were undecided. Now most intend to adopt, a firm minority has decided against, and even among non-users the largest group believes practices using AI will gain an advantage. Four in five respondents expect AI to be standard in clinic workflows within five years.

The demographic story reversed itself. In 2024 the under-30 group was the most skeptical cohort in the profession and among the least likely to use AI at work. In 2026 every age band sits at the same level of optimism, and adoption doesn’t differ by age at all. The under-30s closed the largest gap. Practice setting doesn’t matter either. General practice, emergency, mobile, shelter, and relief all sit in a narrow band. The profession didn’t adopt AI from the top down or from the young outward. It adopted it everywhere at once.

That’s the first implication for the industry. Arguments that treat AI as a generational preference, a corporate fashion, or a specialty hospital luxury are out of date. The remaining non-users aren’t a lagging demographic so much as a mix of the still hesitant and a small group, concentrated among older respondents, who’ve made a real decision to sit this out. Outreach to that group should be practical rather than evangelical. Show the work, show the guardrails, and leave the decision alone.

A whole-practice tool, still pointed at the paperwork

The second change is who uses it. In 2024 AI was a clinician and manager tool, and the spread between roles was wide. In 2026 it has nearly closed. Receptionists and CSRs moved furthest, and technicians and assistants closed similar gaps. AI left the exam room and reached the front desk, which is exactly where a technology that drafts messages and writes notes would be expected to land.

What people use it for has been slower to change. Documentation remains the dominant application by a wide margin. Client messaging, outreach, and marketing follow. Radiographic interpretation, clinical decision support, lab interpretation, and cytology sit well behind. Breadth of use grew only modestly. General practice, which is most of the profession, reports the highest use of documentation tools and the lowest use of clinical decision support and remote monitoring. The pattern isn’t mysterious. Practices automated the work that was already drowning them.

Satisfaction follows the same line. Communication tools score highest. Diagnostic tools score lowest. Radiographic interpretation ranks last on every measure and carries the highest dissatisfaction of any application in the survey. That finding should be read carefully. It doesn’t prove that veterinary imaging AI is a failed category. It does show that the profession’s lived experience of those products, in 2026, hasn’t matched the marketing. Accuracy remains the top worry in both waves. When a tool’s job is to draft an appointment reminder, a clumsy sentence is an inconvenience. When its job is to read an image, a miss is a clinical event. The market hasn’t yet earned the trust that the administrative tools have.

This is the cautious-adoption thesis from 2024, still intact and now supported by usage rather than speculation. The profession treated AI first as a productivity layer on top of existing medicine, not as a new way of practicing it. That was rational. It’s also incomplete. The next increment of value won’t come from writing the same SOAP note faster. It’ll come from connecting documentation, imaging, lab data, and client communication into a single patient record that can recommend a next step. Very few practices are there. Only about one in ten general and mobile practices have a system that recommends the next step for an individual patient based on history. Most have automated the reminder and stopped. Rule-based recalls aren’t AI, and practices that run them score no higher on maturity than practices that wait for the phone to ring. Automating yesterday’s outreach isn’t the same as practicing proactive care.

Adoption ran ahead of management

The Veterinary AI Maturity Model, adapted from the University of Chicago’s HAIRA framework for healthcare AI governance, is the new instrument in this wave. It scores six dimensions on a four-level scale. Strategy, governance, vendor evaluation, workflow integration, adoption depth, and capability. The shape of the scores matters more than any single number.

Practices score highest on feeling equipped and on how widely the team uses AI. They score lowest on governance, with workflow integration and vendor evaluation just above it. Strategy sits in the middle, which sounds respectable until the underlying item is unpacked. Nearly half of AI users with visibility of practice decisions have no AI plan and nobody responsible for it. Nearly a third have none of the four basic supports. No guidelines, no training, no named lead, no formal evaluation. Half of practices choose tools informally, comparing a couple of options or taking whatever is bundled or free. Only one in five formally evaluates tools for data handling and security and then reviews performance after adoption.

This isn’t a story about reluctant staff. Leadership support and staff openness are both high, and budget and infrastructure aren’t the binding constraint either. Training is. One in three say training on new technology is inadequate. The will is present. The operating system isn’t.

Integration makes the same point in a more concrete way. More than half of AI users with a PIMS are only partly connected or not connected at all and still move information by hand. The daily cost shows up in the friction data. Reviewing and correcting output is the most common friction point, and it’s the right kind. That’s clinical judgment being applied to a draft. Copying information between tools, re-entering the same data, and switching logins are the wrong kind. Connecting the tools changes which friction a practice feels. It doesn’t, by itself, remove friction. Practices that integrate without redesigning the work trade copy-paste for inconsistent outputs across products.

Two thirds of practices added AI without changing how they work or made only small adjustments. Fewer than one in ten regularly rethink workflows when they adopt a new tool. That number should bother anyone who has watched a practice install new software and then keep the old paper process next to it. Tools don’t change the day. Changed work changes the day. The maturity model treats full connection and redesigned workflow as a single top level for a reason. One without the other is an expensive way to stand still.

Maturity is what predicts results

The clearest pattern in the study is the link between the maturity index and reported business outcomes. Practices at the bottom of the scale report fewer than one outcome each, and seven in ten say AI has produced no measurable impact. Practices at the top report six times as many. By Level 3 the share seeing no impact is down to about one in ten, and it stays there. The relationship survives after accounting for how many AI applications a practice uses. Buying more tools isn’t a substitute for running them on purpose.

The outcomes themselves are mostly operational. More time for patient care or higher throughput is the most common, followed by client retention, new digital services, and better operational decisions. The free-text answers add a detail the checklist didn’t ask for. The respondents who wrote something in didn’t describe using the recovered hours to book more appointments. They described leaving on time. Doctors finishing notes before they walked out the door, less paperwork burnout, higher morale. The 2024 discussion hypothesized that AI could improve staff well-being by taking repetitive work off the team. The 2026 write-ins are that hypothesis in a smaller group of respondents’ own words. Time is the outcome practices notice. Well-being is how they spend it.

Ownership doesn’t explain maturity or adoption. Corporate and private practices are statistically indistinguishable on both, and private practices use AI only slightly more often. What ownership does explain is experience. Private practice respondents rate readiness higher on most measures, express more satisfaction with the tools, and report more business outcomes, with the largest gaps on client retention and operational decision making. The survey can’t say why. Plausible mechanisms include shorter decision chains, a closer link between the person choosing the tool and the person using it, and less standardization of workflow across sites. Those are hypotheses, not findings. But corporate groups should treat the experience gap as a design problem rather than a talent problem. The tools are in the building. The results aren’t landing the same way.

Practice size works differently. Larger hospitals score higher on the organizational dimensions and no higher at all on adoption depth or workflow integration. Small practices spread AI across their teams as widely as large ones. What they lack is the scaffolding. Governance barely exists below seven veterinarians. That’s a structural observation, not a moral one. A three-doctor practice doesn’t need an AI officer. It does need one person responsible for it, a one-page rule for what may and may not go into a prompt, and a habit of asking vendors where the data lives. Those are cheap. The cost of not having them will rise as tools move from drafting emails to touching the medical record.

The worry changed its object

Reliability and accuracy remain the leading concern, as in 2024. Almost everything underneath that headline has moved. Fear of job displacement fell by almost half. Concern about data security rose, and concern about training gaps rose with it. The profession is less afraid of being replaced and more afraid of operating the software badly. That’s a more useful fear.

Displacement anxiety isn’t evenly distributed. It falls steadily with age, and by role it concentrates at the front desk, where it runs at more than double the rate among veterinarians. The same roles that adopted AI fastest are the ones most worried it will take the job. That’s not a contradiction. It’s what it looks like when a tool arrives in the part of the hospital whose work is most legible to software. The constructive response is to redefine those roles around the work AI can’t do. Judgment with a distressed client, triage when the waiting room is full, the conversation that happens after the draft is generated. And to train for that work out loud.

A new note appears in the free text. Some respondents object to AI on environmental grounds and still expect it to become standard. That combination is going to get more common. Practices and vendors that can talk plainly about compute, data retention, and model choice will have an easier time with a workforce that has already moved past “should we use this” and on to “on what terms.”

The do-it-yourself current

Two in five AI users have tried to build their own agent in tools like ChatGPT or Claude, and about one in five says it worked. Nearly a third more are considering it. For a profession that historically buys software rather than writes it, that’s a large number, and it cuts two ways.

On one side it’s evidence of appetite, and of a workforce that’s no longer waiting for the vendor roadmap. A technician who can stand up a discharge-instruction assistant on a Saturday isn’t a problem to be managed. On the other side, unmanaged agents are how client data leaves the practice by accident. The same survey that documents this appetite documents the absence of guidelines in nearly a third of hospitals. Those two facts shouldn’t be allowed to coexist for another two years. The answer isn’t to ban shadow AI. It’s to give people a sanctioned path, with approved tools, approved data classes, and a named reviewer, so the energy goes into useful agents instead of workarounds. It’s also a signal to every PIMS vendor. Practices want software that acts on their data. The systems of record that let it in, on the practice’s terms, will be the ones the agents get built around.

What this means for the industry

Vendors should stop selling adoption. Adoption has occurred. The buying criteria that will matter in the next cycle are integration with the PIMS, consistency of output, performance review after deployment, and training a three-doctor practice can finish. A feature being inside the PIMS doesn’t settle any of those. The bundled option is a candidate, not a shortlist. Diagnostic products have a specific credibility problem that administrative products don’t, and closing it takes published validation in veterinary populations, not testimonials.

Practice groups, corporate and independent, should treat maturity as an operating discipline rather than a technology project. The index isn’t a trophy. It’s a punch list. Put one person in charge of it. Write down what the tools are for. Decide how a new product gets evaluated, including where the data goes. Connect the tools or stop adding them. Redesign one workflow end to end, and medical records is the obvious candidate, instead of sprinkling copilots onto an unchanged day. Measure something other than whether people logged in. The practices already doing this are the ones reporting results. The ones that aren’t are the ones reporting that AI did nothing.

Educators and associations have a narrower, more urgent job than they did in 2024. The last paper called for AI in the veterinary curriculum. That call still stands, but the content must change. Students who arrive fluent in chat interfaces don’t need another lecture on what a large language model is. They need supervised practice in checking a draft, declining a bad suggestion, documenting the human decision, and explaining an AI-assisted recommendation to a client. The training gap in this survey isn’t conceptual. It’s operational. Continuing education should follow the same rule.

Professional bodies should notice that the profession is writing its own rules in the absence of theirs. Informal guidelines, peer recommendations, and bundled tools are how most hospitals currently decide. That’s fine for a drafting assistant. It isn’t fine for anything that touches diagnosis, prescribing, or the official medical record. Clearer guidance on data handling, disclosure to clients, and the veterinarian’s retained responsibility wouldn’t slow adoption. Adoption is done. It would make the next phase safer.

Recommendations

Put one person in charge of AI this quarter. A practice doesn’t need a chief AI officer. It needs one person who can say which tools are in use, what they’re for, and who reviews them. Nearly half of AI-using practices currently have neither a plan nor that person.
Write a one-page usage guideline before adding another product. Say what data may be pasted into an external model, what must stay inside the PIMS, and when a human must sign the output. This is the cheapest governance step available, and the one most hospitals haven’t taken.
Evaluate vendors for data handling and security, then review performance after 90 days. One in five practices does this today. A short, repeatable scorecard is enough. Accuracy in your caseload, integration cost, support, and whether the tool still earns its fee. Four questions, asked on a schedule, would move a large share of the market from ‘bundled and hoped’ to ‘chosen and checked’.
Fix the medical record workflow first. Documentation is the dominant use case, the leading source of recovered time, and the leading source of after-hours burnout. Redesign that path so the note is captured, checked, and filed without a copy-paste step. Practices that bolt a scribe onto an unchanged records process shouldn’t be surprised when the evening doesn’t come back.
Train the front desk and the technical staff as carefully as the doctors. Those roles adopted fastest and worry most about displacement. Training that only addresses veterinarians will miss the people who now run the tools all day.
Don’t confuse automated reminders with proactive care. Most hospitals have the first. Almost none have the second. The next clinical dividend is individual, history-aware outreach, not another six-month vaccine text.
Publish veterinary validation for diagnostic tools. Communication products have earned their satisfaction scores. Imaging, cytology, and lab interpretation haven’t. Until independent performance data exist for the species and settings these tools are sold into, skepticism about diagnostic AI is a professional standard, not a lag.
Measure well-being alongside throughput. The write-ins are unambiguous. Teams are spending the saved time on leaving work, not on adding appointments. Hospitals that capture the hours and then refill the schedule will have used AI to speed up burnout rather than relieve it.

Limits of this reading

I want to include two caveats with the conclusions. First, maturity and outcomes are both self-reported at a single point in time, so the survey can’t establish which way the relationship runs. Deliberate adoption may produce results. Early good results may persuade a practice to invest in structure. Or a bad early experience may cause a practice to pull back and stay at Level 1, which would mean some of the least mature practices are there because AI disappointed them, not the other way round. All three are plausible, and the data can’t separate them. What it can say is that practices that adopt deliberately report better results, and that the gap is large. Second, the 2026 sample is smaller than 2024 and isn’t a matched panel. Comparisons across waves describe the profession as sampled, not the same people two years later. None of that weakens the central pattern. It does mean the maturity index should be used as a diagnostic, not a ranking.

Closing

The 2024 paper ended in a posture of preparation. Train people. Publish case studies. Write ethical guidelines. Shift the mindset. Those remain good ideas. They’re no longer the bottleneck. The bottleneck is operational. Someone in charge, written rules, integration, redesigned work, and training that matches the jobs people are already doing.

AI is now ordinary veterinary infrastructure, about as remarkable to most of the workforce as the practice management system it still, too often, fails to talk to. Ordinary infrastructure still must be governed. The hospitals that treat it that way are already seeing more time, better retention, and clearer decisions. The hospitals that treat it as a collection of clever tabs will keep reporting that nothing measurable happened. The difference isn’t the software. It’s whether anyone oversees it.

We’d like to thank Adam Wysocki for his generous contribution to this study. His experience and perspective brought expert insight into what the findings mean for the industry. We hope veterinary clinics find genuine value in his recommendations as they build toward smarter, more deliberate AI adoption.

After the Breakthrough Comes the Work: Making AI Matter in Veterinary Practice