Is Customer Data Safe with AI Tools? A Guide for Veterinary Practices

The Adoption Paradox: High AI Use, Low Data Awareness

Somewhere in your clinic right now, an AI tool is probably listening. It may be transcribing consultations, drafting medical records, or summarizing client communications — and the question of what happens to that data afterward is one most practices have never asked out loud.

The scale of adoption is no longer in doubt. A 2024 survey of 3,968 veterinary professionals found that 39.2% used AI in medical practice and 69.5% for administrative tasks. In a separate 2025 survey, 71.0% of veterinary professionals reported using AI tools professionally, with nearly 20% of users relying on them daily.

Yet the same research reveals a striking gap in understanding. Among those active Chinese users, 44.6% reported low familiarity with the technology — what the study authors call an adoption paradox. Clinics are adopting tools faster than they're learning how those tools work, including how they handle the sensitive information flowing through them.

That information is anything but trivial. AI tools in a typical practice may process consultation recordings, client details, medical histories, diagnostic results, and draft records. Most clinics have never verified what these vendors do with any of it.

• Consultation recordings and transcripts that may be retained for model training
• Client contact details and communication histories
• Medical records and diagnostic results
• Prompts, corrections, and generated documents fed back into the system

The hidden risk is that a tool can look secure without actually protecting your data. As one veterinary AI privacy guide puts it, "information can be encrypted while still being permitted for training under a provider's terms." In other words, encryption alone is not data protection — a vendor can encrypt everything in transit and still legally use your clinic's records to train its models.

Compounding the problem, veterinary AI operates in a regulatory gap. Unlike human healthcare, there's limited guidance for AI-based clinical tools in veterinary medicine, and many commercial systems use proprietary algorithms whose inner workings practitioners can't inspect. Data privacy and security stewardship has been identified as one of four key ethical areas for veterinary AI implementation — but ethics frameworks don't verify vendor practices on your behalf.

The result is a profession running ahead of its own due diligence: high adoption, low awareness, and client data flowing into tools whose data practices remain unverified. Closing that gap starts with treating every AI vendor like an unknown quantity until proven otherwise.

Encryption Isn't Protection: The Hidden Training-Data Risk

Here's the uncomfortable truth most veterinary practices miss when evaluating AI tools: a vendor can truthfully advertise bank-level encryption while your clients' data is still being fed into their AI models. As one veterinary AI privacy guide puts it, "information can be encrypted while still being permitted for training under a provider's terms." Encryption protects data in transit and at rest — it says nothing about what a provider is contractually allowed to do with that data afterward.

This distinction matters because AI tools routinely process sensitive material: consultation recordings, client names and contact details, medical histories, diagnostic results, and draft medical records. If any of that is permitted for model training — including by third-party AI providers the vendor relies on — your clients' information could effectively leave your control forever, even though every transmission was fully encrypted.

The scale of adoption makes this urgent. A 2024 survey of 3,968 veterinary professionals found 39.2% already using AI in medical practice and 69.5% using it for administrative tasks. Yet many users don't understand what they've signed up for: in a separate survey, 44.6% of active AI users reported low familiarity with the technology. That combination — widespread use, limited understanding — is exactly where data risks go unnoticed.

Before adopting any tool, get written answers to these questions:

• Are recordings, transcripts, files, corrections, prompts, or generated documents used to train or evaluate AI models — including third-party providers?
• Where is data stored, and which subprocessors or support teams can access it, including across borders?
• What does "delete" actually mean — UI removal only, or removal from backups and logs too?
• Who owns the submitted data and the generated records, and what happens at end of contract?

Verbal assurances and vague "kept private" claims are insufficient. Insist on a data-processing agreement covering permitted uses, safeguards, subprocessors, incident reporting, and international transfers — and verify credentials rather than trusting badges. There is no official HIPAA certification, and HIPAA doesn't automatically apply to veterinary records anyway; ask providers to explain the actual safeguards behind any compliance claim. A current ISO/IEC 27001 certification from an accredited body and access to a SOC 2 Type II review are far stronger evidence than a sales rep's promise.

Practices extending AI into broader workflows face a parallel challenge: published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory where skills are security-scanned and version-pinned before you connect them to your agents. The principle is the same everywhere: security should be clear before adoption, not pieced together afterward.

Build Your Data-Safety Checklist Before You Adopt

Before signing up for any AI tool that touches client data, a practice should be able to answer one question in writing: what happens to the information you send it? The urgency is real — with 69.5% of veterinary professionals already using AI for administrative tasks, data safety questions are no longer hypothetical. And the most common assumption clinics make is also the most dangerous one.

That assumption is that encryption equals protection. As one veterinary AI privacy guide puts it, "information can be encrypted while still being permitted for training under a provider's terms." In other words, your data can be fully secured in transit and still legally used to train models. Verbal assurances or vague "kept private" claims are not enough — get written confirmation covering recordings, transcripts, prompts, corrections, and generated documents, including any third-party AI providers involved.

A practical pre-adoption checklist should cover four areas:

• Training use in writing: confirm whether any submitted data trains or evaluates AI models, now or in the future.
• A data-processing agreement (DPA): covering permitted uses, subprocessors, incident reporting, international transfers, and end-of-contract deletion — with the clinic retaining ownership of its data.
• A full security program: individual accounts, strong authentication, role-based access for vets, technicians, students, and locums, activity logs, and prompt removal of former employees.
• Deliberate consent and retention policies: recording consent should explain why data is collected, how it's used, how long it's kept, and what happens if a client declines.

It also helps to separate three obligations that clinics often blur together. Privacy governs what data is collected and how it's used. Security covers the safeguards protecting it. Record-keeping rules determine what must be preserved — which matters because "delete" in an AI tool may mean UI removal only, with backups and logs retaining copies temporarily, while professional obligations may require medical records to be kept longer.

One misconception worth correcting: HIPAA does not automatically apply to veterinary records. That doesn't mean veterinary data is unregulated — state privacy, recording, and breach-notification rules may still apply, and there is no official HIPAA certification, so any provider claiming compliance should be able to explain the safeguards behind the claim.

The same scrutiny applies as practices extend AI into broader workflows. When connecting agents to third-party skills, note that published research found roughly 26% of agent skills carry at least one vulnerability — which is why SkillzBench is building a free directory where skills are security-scanned, version-pinned, and benchmarked before you connect them. Security should be clear before adoption, not something the clinic has to piece together afterward.

Keep the Veterinarian in the Loop — and Vet Every Tool

An AI tool can encrypt your data perfectly and still use it to train someone else's model. That's why the final safeguard isn't technical — it's you. The vetting mindset you apply to vendors has to extend to every output the tool produces, and to every extension you bolt onto it.

The clinical reason is straightforward. In the Vetology AI study, the system achieved 93.75% specificity but only 55.4% sensitivity — meaning it was far better at ruling out disease than confirming it. A SignalRAY study found the same pattern: AI performed on par with the highest-performing radiologist yet was more specific and less sensitive than human reviewers.

For a practice, that asymmetry has a practical meaning: these tools are good at recognizing normal images and weak at catching subtle abnormalities. A clean AI read is not a clean bill of health. That's why peer-reviewed guidance concludes that AI should aid, not replace, veterinary decision-making, and why every AI-generated record should remain subject to clinical review and approval before it enters the medical file.

Your review checklist should be as deliberate as your vendor checklist:

• Assign a named veterinarian to review and sign off on every AI-drafted record or diagnostic suggestion.
• Treat AI "normal" findings with extra caution — low sensitivity means abnormalities can slip through.
• Remember that many veterinary AI systems use proprietary algorithms, so you may not know what factors drove a diagnosis.
• Watch for bias: AI trained mostly on canine and feline patients (dogs appeared in 59.1% of reviewed studies) may be unreliable for exotics or livestock.

The same vetting discipline applies when you extend AI beyond records — into agent skills and automated workflows. This is not a hypothetical risk: published research found that roughly 26% of agent skills carry at least one vulnerability, and installing an unvetted skill into a workflow that touches client data undoes every safeguard you negotiated.

This is where a dedicated resource helps. SkillzBench is building a free, independent directory of AI agent skills in which every skill is security-scanned, version-pinned to a specific commit, license-checked, and benchmarked in a sandbox — with a 0–100 trust score that separates security signals from marketing claims. Pairing that kind of verification with your own clinical review closes the loop: the tool is checked before it touches data, and the output is checked before it touches the patient record.

Vet the vendor. Vet the skill. Vet the output. All three, every time.