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

Why Dental Practices Are Adopting AI — and What's at Stake

Walk into a dental practice today and there's a good chance AI is already quietly at work — scheduling patients, verifying insurance, drafting chart notes. According to an ADA Health Policy Institute survey, 43.3% of U.S. dentists now use AI for at least one task, with another 26.4% planning to adopt it soon.

The adoption pattern is telling. Current use skews heavily toward administrative work: 13.6% of dentists use AI for insurance verification, 10.1% for reception and front-desk check-in, and 10.1% for business analytics. The most anticipated future uses are charting and note taking (34.8%) and insurance verification (32.6%). The driver is practical — 32.4% of dentists report being overworked, and staffing shortages have pushed non-clinical workload to a breaking point.

Clinical AI is a different story. Fewer than 5% of dentists currently use AI for treatment recommendations, and 82.6% say they won't. That caution is healthy, but it can create a false sense of security: practices that avoid AI in the operatory may still be feeding patient data into administrative AI tools without realizing the exposure.

The stakes are not hypothetical. Healthcare reported a record 772 large breaches in 2025, and hacking now accounts for over 80% of large breaches, up from 49% in 2019. Dental entities are firmly on the target list — HIPAA Journal's breach data shows DentaQuest alone affected 15 million individuals, alongside Delta Dental of California (7 million) and Managed Care of North America (8.6 million).

Small practices are vulnerable too, often through vendors rather than direct attacks. In a recent wave of dental breaches across six states, phishing and email compromise were the dominant vectors, and a single third-party vendor breach cascaded to at least four dental practices — including Stafford Oral Surgery (7,019 patients) and Garrisonville Dental (5,204).

Here's the critical point many practices miss: adopting AI doesn't change your legal obligations. There is no HIPAA rule written specifically for AI, so every tool that receives, stores, or transmits protected health information falls under existing law — and the vendor becomes a HIPAA business associate the moment it touches PHI. That means a signed Business Associate Agreement is legally required before use, and "HIPAA compliant" marketing claims are no substitute. The same vetting logic applies to any AI agent or skill a practice plugs into its workflows; understanding what a tool does with data before connecting it is exactly the discipline platforms like SkillzBench are built to support, with security-scanned, version-pinned agent skills.

AI can lighten the administrative load. It cannot lighten the liability — and practices that understand that distinction from day one are the ones that adopt safely.

The BAA Rule: Your Legal Linchpin for AI Tools That Touch PHI

If an AI vendor touches patient data, HIPAA treats that vendor exactly like a member of your staff — and the paperwork to prove it is non-negotiable. The Business Associate Agreement (BAA) is the single most important safeguard a dental practice can put in place before any AI tool ever sees protected health information (PHI).

The rule is straightforward: if an AI system will receive, store, or transmit PHI on behalf of your practice, the vendor legally qualifies as a business associate, and a written BAA is required before use. There is no HIPAA rule written specifically for AI yet, but that doesn't create a gap — AI companies with PHI access are simply treated as business associates under existing law.

Here's where many practices get burned: "HIPAA compliant" is a marketing claim, not a legal protection. A vendor can plaster the phrase across its website, but unless a BAA is actually signed, your practice carries the liability. As compliance guidance from OSHA Review puts it, if a vendor refuses to sign a BAA, the practice cannot lawfully use that tool with patient data — regardless of how the tool markets itself.

Free consumer tools are the sharpest edge of this problem. The California Dental Association warns that dentists should never use the free, public version of ChatGPT with patient data, because anything entered could be used to train the model. The same caution applies to Google Gemini and similar platforms unless they've been specifically approved for compliant use with a signed BAA. Even ChatGPT's enterprise offerings only recently began offering BAAs on a case-by-case basis for API customers.

The stakes aren't theoretical. Healthcare reported a record 772 large breaches in 2025, with hacking behind more than 80% of them, according to HIPAA Journal. Dental-sector entities have been hit hard: DentaQuest (15 million individuals), Delta Dental of California (7 million), and Managed Care of North America (8.6 million) all rank among major breaches. Third-party vendors — the same category as AI vendors — are a leading breach point, and one vendor compromise cascaded to at least four dental practices.

Before adopting any AI tool, run through a short vendor screen:

• Will staff enter PHI into the tool, and how much?
• Does the vendor retain or use your data for model training?
• Who at the vendor can access the data, and for how long?
• Will the vendor sign a BAA — and can they produce one today?

A vendor with an easily accessible HIPAA policy on its website is a good sign; a vendor that dodges the BAA conversation is a dealbreaker. For practices evaluating AI agent skills and automation tools, verifying what a tool actually does with data before connecting it matters — resources like SkillzBench's free skills directory (currently in development, with a live waitlist) aim to make security and quality signals visible before adoption. And if your existing practice software adds new AI features, confirm your current BAA covers them.

The bottom line: no signed BAA, no patient data. It's that simple.

A Practical Compliance Framework: De-identify, Vet, and Document

The good news is that protecting patient data with AI doesn't require new law — it requires the same discipline dental practices already apply to any other vendor. HIPAA's obligations haven't changed; the workflow for applying them to AI tools can be broken into three steps: de-identify, vet, and document.

Step one: de-identify before anything leaves the practice. Under HIPAA's Safe Harbor method, de-identification requires removing 18 specific identifiers, and once data is properly de-identified, there is no restriction on its use or disclosure under HIPAA, according to a peer-reviewed legal analysis. But de-identification isn't a magic shield. Researchers warn that dominant tech companies with massive access to personal information — a risk highlighted in Dinerstein v. Google — create real potential for re-identifying "anonymous" datasets, and that combining anonymized data with other sources increases that risk. The practical takeaway: de-identify aggressively, share only the minimum necessary information, and audit your anonymization process continually rather than treating it as a one-time task.

Step two: vet every AI vendor before adoption. HIPAA's Security Rule requires that every AI tool be included in the practice's security risk analysis, with vulnerabilities identified and risk management steps documented. The free Security Risk Assessment (SRA) Tool from HHS is designed specifically for small and medium providers and walks you through this process. Before adopting any tool, ask four questions:

• Will staff enter PHI into this tool?
• Will the vendor retain the data or use it to train its models?
• Who at the vendor has access to the data, and for how long?
• Does the vendor offer a Business Associate Agreement?

That last question is the dealbreaker. If a vendor refuses to sign a BAA, the practice cannot lawfully use that tool with patient data — regardless of how the vendor markets itself. A vendor's easily accessible HIPAA policy page is a good early signal of how seriously it treats PHI security.

Step three: document everything. Record which AI tools are in use, what data they touch, what the vendor agreed to, and what risk management steps you've taken. This documentation matters because third-party vendors are a major breach point in healthcare — one recent vendor breach cascaded to at least four dental practices, per HIPAA Journal reporting — and practices with documented habits will be better positioned when HHS's proposed Security Rule updates take effect.

Vetting applies to AI tools themselves, not just vendors. As agent skills and AI workflows proliferate, practices and businesses alike should verify what a tool actually does before connecting it — the same principle behind SkillzBench's free directory, which security-scans and benchmarks agent skills before you equip them. De-identify, vet, document: three habits that keep patient data safe no matter which AI tools your practice adopts next.

Strengthening the Human and Technical Defenses Around Your AI Stack

The most damaging attacks on dental practices rarely involve sophisticated AI exploits — they arrive by email and through vendors. A wave of dental breaches across six states traced back to phishing and email compromise, and one third-party vendor breach cascaded to at least four separate practices. Your AI stack inherits these same vectors.

The good news is that baseline defenses are inexpensive and well-documented. HIPAA Journal's analysis of recent dental incidents recommends the fundamentals:

• Strong, unique passwords for every employee email account
• Multifactor authentication (MFA) on all accounts that touch patient data
• A dedicated email security solution to filter phishing attempts
• Ongoing security awareness training for the whole team

Vendor risk deserves equal attention, because hacking now accounts for over 80% of large healthcare breaches, and many of the largest incidents involved business associates rather than providers directly. Every AI tool you adopt is a new third party with access to your systems, so vendor vetting belongs in the same conversation as firewalls and passwords. HHS offers a free Security Risk Assessment tool designed for small and medium providers, and every AI tool should appear in it.

Human oversight remains your strongest safeguard. The ADA survey captures the profession's stance plainly: "It has its place but still must be monitored and verified by humans." Compliance guidance is equally direct — a clinician should review every AI output against the standard of care. That review habit also guards against over-reliance, which legal reviewers warn can erode practitioner skills over time.

Two administrative steps complete the picture. Update your consent forms to explain how patient data may be used in AI training, revisiting the language as tools evolve. Then review your malpractice and cyber insurance policies for AI-related gaps — coverage should explicitly address AI-driven errors and data breaches.

If your team uses AI coding or workflow agents, the same vetting discipline applies to the skills and plugins those agents run. SkillzBench is building a free directory that security-scans and benchmarks agent skills before you deploy them, so you can spot risky ones before they touch practice data.

Building a Vetted AI Toolkit You Can Trust

A dental practice's AI safety posture is only as strong as its weakest component. As practices move beyond single tools and begin assembling AI into workflows — charting assistants connected to scheduling agents, insurance verifiers feeding analytics dashboards — every piece of that chain becomes a potential entry point for patient data to leak.

The stakes are not hypothetical. Third-party vendors are a major breach point in healthcare, and a wave of recent dental practice breaches showed how a single vendor compromise cascaded to at least four practices at once. Meanwhile, hacking and IT incidents accounted for more than 80% of large healthcare breaches in 2025, a record year with 772 large breaches reported. When you stack multiple AI components together, you multiply the number of vendors, data flows, and failure points you must vet.

Vetting each component systematically means checking four things before anything touches your systems:

• Security: Does the tool or skill have known vulnerabilities, and does the vendor sign a BAA if PHI is involved?
• Licensing: Is the component legally cleared for commercial use in a practice setting?
• Quality: Does it actually perform better than working without it, or does it just add risk?
• Version control: Are you running the exact version you evaluated, not whatever silently changed upstream last week?

That last point deserves emphasis. Agent skills — instruction packages that extend tools like Claude Code, Cursor, and Gemini CLI — are proliferating quickly, and published research found that roughly 26% of them carry at least one vulnerability. Many are community-published on GitHub with no review at all, and an upstream update can change behavior after you have already adopted it.

This is the problem SkillzBench is being built to solve. As an independent, free directory of AI agent skills, it is designed to make component vetting systematic rather than ad hoc: every skill is pinned to a specific commit so you know exactly what you are running, security scans and license checks are applied before listing, and each skill is benchmarked in a sandbox — tested with and without the skill, using the same model — so you can see the measured performance lift rather than a vendor's marketing claim. A 0-100 trust score combines security, quality, license, maintenance, and community signals, and upstream changes trigger re-scans and alerts so your vetting does not go stale. A free MCP connector will let you connect only the skills you have chosen to your agents, keeping the blast radius small.

The principle here mirrors the compliance guidance that runs through this entire guide: verify before you trust, and keep humans in the loop. Whether the component is a diagnostic imaging tool that requires a signed BAA before lawful use or an open-source scheduling skill from GitHub, the discipline is the same — evaluate it deliberately, document what you found, and re-check when things change. Practices that build that habit now will be far better positioned as AI workflows, and the rules governing them, continue to evolve.