Dental office front desk computer showing approved AI apps with checkmarks and two locked unapproved apps

Your team may already be using AI at work. The practice just hasn’t decided how.

An AI acceptable use policy tells your team which AI tools they can use for practice work, what information can go into them, who checks what they produce, and what happens when someone uses a tool nobody approved. Plenty of templates cover the wording. What they skip is the part an IT team sees every week: the policy holds up best when it’s built into the computers your team uses. Below is how I’d build one for a dental practice or a dental group, or you can jump straight to the template.

Why most dental AI policies don’t hold

Standardization efforts in dental groups tend to stall the same way. The standard gets announced as a policy. It never gets built as a default.

An AI policy is a good example. It says staff may only use approved AI tools. But if the front desk computer still lets anyone install a browser extension or connect an AI notetaker to a staff mailbox, the workstation is saying yes to everything the policy says no to.

Shadow AI is any AI tool your team uses for practice work that the practice hasn’t approved. It’s rarely malicious. Someone uses ChatGPT or Gemini to tidy up a patient letter, it works, and a few weeks later a patient’s name ends up in a prompt without anyone deciding it should.

IBM’s cross-industry breach research shows how common this has become. In its 2026 Cost of a Data Breach Report, IBM found that security incidents involving shadow AI, where workers use unapproved AI, more than doubled to 43% from 20% the year before. And 68% of all the breached organizations studied lacked AI governance to manage AI or detect shadow AI. The study covers 602 breached organizations across 17 industries and 16 countries, not dental practices. It still tells you where to look.

Does HIPAA require an AI policy?

Not by name. As of October 2026 we found no final HHS guidance written specifically about staff using ChatGPT-style tools with patient information. That doesn’t leave you without rules. The HIPAA requirements you already have apply to AI tools the same way they apply to email or a practice management system:

  • Risk analysis. The Security Rule lists it as Required: an accurate and thorough assessment of the risks to electronic patient information (45 CFR 164.308). An AI tool that holds patient information belongs in it today.
  • A sanction policy. Also Required: apply appropriate sanctions to workforce members who don’t follow your security policies. An AI rule with no stated consequence is a suggestion.
  • Activity review. Also Required: regularly review records of system activity, such as audit logs and access reports. Reading that to include which AI tools are in use on your machines is our interpretation.
  • Training. A security awareness and training program for every workforce member, management included. Our post on security awareness training for dental practices covers the training requirements.
  • Written policies. Security policies have to be in writing, kept for 6 years from when they were created or last in effect, whichever is later, and reviewed and updated as your environment changes (45 CFR 164.316).
  • A business associate agreement first. Before a vendor creates, receives, maintains or transmits patient information for you, HIPAA requires satisfactory assurance in a written agreement, in practice a BAA (45 CFR 164.502(e)).
  • Minimum necessary. HIPAA’s minimum necessary standard (45 CFR 164.502(b)) generally covers how you use patient information and what you share with vendors, with defined exceptions. Make it the rule for every AI tool: use only the patient information the task needs.

In January 2025, HHS proposed a Security Rule rewrite that would require a written inventory of the technology assets that can affect patient data. It hasn’t been finalized, and HHS’s regulatory agenda lists it as a long-term action with a projected final action date of July 2027, an agency estimate rather than a deadline. The proposal’s preamble also says HHS expects practices using AI to include those tools in their risk analysis, which the current rule already supports.

What goes on an approved AI tools list?

The list is the working core of the policy. Each entry should say which product and plan it is, what it’s approved for and what it isn’t, what information is allowed in it, whether a BAA is in place and when it took effect, who reviews its output, who can turn it on for a user, and when it’s due for review. If a tool isn’t on the list, it isn’t approved.

For general-purpose assistants, the account type matters more than the brand. Here’s where three common ones stand, from each vendor’s own pages as of October 2026.

Read October 22, 2026 from each vendor’s own documentation. Vendor terms change, so confirm the current terms before you approve a tool.
Tool Business or enterprise version Free or personal version
Microsoft Copilot (formerly Microsoft 365 Copilot) Signed in with a work account, Microsoft says Copilot and Copilot Chat support HIPAA compliance for properly configured implementations, under the BAA in its business terms. Web search queries aren’t covered by the BAA, so before you approve Copilot for patient information, have IT turn off web search for those users and check that the setting took. The consumer terms say you can only use Copilot for your own personal use. They don’t mention a BAA.
Google Gemini Gemini in Google Workspace and the Gemini app are on Google’s list of services covered by its BAA, once your administrator accepts the BAA. Gemini in Google Chrome is excluded. Confirm your Workspace edition and settings against Google’s current list. Google tells users not to enter confidential information they wouldn’t want a human reviewer to see, and keeps reviewed chats for up to three years.
Anthropic Claude Anthropic offers a BAA for its API (not every API feature is covered), and for Claude Enterprise once the organization’s Primary Owner turns on HIPAA and accepts the BAA. Some features, such as connectors, stay outside it. Free, Pro, Max and Team plans can’t enable HIPAA.

ChatGPT has its own set of plans and exceptions, which we went through in our review of dental AI vendors and HIPAA. The short version: OpenAI’s personal plans (Free, Plus and Pro) aren’t on its list of HIPAA-eligible products, and it doesn’t offer a BAA for ChatGPT Business.

Two cautions. A BAA is one requirement for a tool that will handle patient information, and the tool still needs a security and workflow review. Microsoft says using its services “doesn’t on its own achieve HIPAA compliance.” And check the AI features inside software you already pay for. Phone, imaging and practice management vendors may have added AI since you signed, so confirm your existing agreement covers it. Your email platform may already include an assistant that can be covered once the BAA is in place, one more reason the choice between Microsoft 365 and Google Workspace matters.

How unapproved AI turns up on dental workstations

At Medix Dental IT we manage workstations for dental practices and dental groups, and we’ve started paying attention to which AI tools get installed on them, because that’s where a policy meets the machines.

It usually shows up in one of these places:

  • Browser extensions. AI sidebars and writing helpers. In December 2025, OX Security reported two Chrome extensions posing as AI sidebars, with about 900,000 users between them, that sent users’ ChatGPT and DeepSeek conversations and open tab addresses to an outside server every 30 minutes. One carried Google’s Featured badge. A badge isn’t an approval process.
  • Personal accounts. A free chatbot signed in on a shared front desk computer, used by whoever sits there.
  • Desktop apps. AI apps a user installs themselves.
  • AI notetakers. Tools that join video calls through a staff calendar. A proposed class action in federal court in California, In re Otter.AI Privacy Litigation, alleges Otter’s notetaker recorded Zoom and Teams meetings without every participant’s consent. Those are allegations, and the case is still pending. For anything that records patients, see our review of dental AI recording consent.
  • App connections. A “connect your email or calendar” prompt that gives a third-party AI app access to a staff mailbox in Microsoft 365 or Google Workspace.
  • New AI features in existing software. A vendor switches on an AI feature in an update, and nobody reviews it.

These checks help IT find unapproved AI and narrow the routes it can take, by default instead of by memo:

Where it shows up How IT finds it The default that narrows it
Browser extensions Extension inventory from managed Chrome or Edge An allowlist set by browser policy, with everything else blocked
Personal AI accounts Web filtering reports show which AI sites are in use. They usually can’t tell a personal account from a practice account on the same site. Unapproved AI sites blocked or warned, the approved tool pinned, and account restrictions where supported
Desktop apps Software inventory from endpoint management No local admin rights, application control where it fits, and a standard software build. Removing admin rights alone won’t stop apps that install per user.
App connections The list of apps granted access in Microsoft Entra or the Google Workspace admin console Users can’t grant a new app access to practice data without admin approval
Notetakers Meeting attendee lists and connected calendar apps The same app approval control, meeting lobbies for outside participants, and a clear rule for staff on admitting bots
AI features in existing software Vendor release notes and admin settings Review at every renewal and every major update, and confirm the BAA covers the feature before it’s turned on

Start with the controls your existing systems already support. Some environments need extra licensing or tools, and every default needs testing and upkeep as vendors change things. Any capable IT team, in-house or outside, can set these up. For transparency, Medix sells managed IT and security services, including these endpoint, browser and account controls and the vendor reviews described below, so weigh my view with that in mind.

When an unapproved tool turns up

How you respond to the first one decides whether staff report the next one or hide it.

  1. Find out what went in. Which tool, which account, and whether any patient information was entered. Ask before you assume.
  2. Contain it without losing the evidence. Have IT preserve what’s available, such as prompts, logs, account details and the permissions the tool held, without holding up urgent containment. Then remove it, revoke any access it was granted to email or files, and end its active sessions. Reset passwords where there’s reason to think credentials were exposed.
  3. Bring in your privacy officer. If patient information went to an AI vendor without a required BAA, your privacy officer assesses it as a possible impermissible disclosure. HIPAA presumes an impermissible disclosure is a breach unless a documented risk assessment shows a low probability the information was compromised (45 CFR 164.402). Your security incident procedures and your breach response plan take it from there.
  4. Apply the sanction policy, the same way every time. Consistency matters more than severity.
  5. Ask why they needed it. When staff reach for a tool, there’s usually a real job behind it. Sometimes the right outcome is a covered version on the approved list.

Who reviews what AI produces

When a dental group is about to buy an AI tool, I ask one question. Who reviews what it produces? Not who administers it. Who is accountable when it schedules something wrong, codes something wrong or reads an image wrong. In most groups that person does not exist.

The policy should name that person for each kind of use. Patient messages drafted with AI get read by a person before they go out. Clinical notes from an AI scribe get reviewed and signed by the clinician. AI coding and billing suggestions get checked by qualified billing staff, under a process your billing lead owns, before a claim goes out. Anything AI suggests about diagnosis or treatment stays a suggestion, and the decision stays with the licensed clinician.

One AI policy across every location

For a dental group, the short answer is one policy, one approved list and one set of defaults, owned centrally, with exceptions that expire. The work is keeping every office on that same version.

  • One approved list, owned centrally. Offices can request tools. They don’t add them.
  • Defaults pushed from one place. Browser policy, app approval and software builds are set centrally, so a new machine or a new hire starts with the approved setup. Check that the settings actually applied, and that approved work still runs, before you roll them out everywhere.
  • Exceptions that expire. If one office needs a tool the others don’t, approve it centrally, write down why, and set a review date.
  • Acquired offices get inventoried before they join. A newly acquired practice may arrive with its own AI phone service, scribe or chatbot. Add AI tools to your IT due diligence, then move each one onto the list or off the network by a set date.
  • Same training, same acknowledgment, every office. Track it centrally, not in each office manager’s inbox.

AI acceptable use policy template for a dental practice

Copy this, replace the bracketed parts, and adapt it to how your practice or group works. Treat it as a starting point. Short on time? Fill in sections 3, 4 and 6 first and name one policy owner. A practice without an IT provider can start with a one-page approved list. It isn’t legal advice, so have your counsel and your privacy officer review it before you adopt it.

[Practice or Group Name] AI Acceptable Use Policy
Effective date: [date] · Policy owner: [AI tool owner] · Privacy officer: [name]

1. Purpose and scope. This policy covers how workforce members use artificial intelligence tools for practice work. It applies to every workforce member at every location, including owners, clinicians, front office and billing staff, temporary staff and contractors. It covers AI features built into software the practice already uses.

2. Definitions. “AI tool” means software that uses artificial intelligence to generate or analyze text, images, audio or code, including chat assistants, writing tools, browser extensions, notetakers, scribes and AI features inside other software. “Approved AI tool” means a tool on the Approved AI Tools List. “Patient information” means any protected health information, including names, dates, images, chart notes and insurance details. “Confidential practice information” means staff records, passwords and other login credentials, and non-public business, financial and contract information.

3. Approved tools and devices. Workforce members may use only approved AI tools for practice work, signed in with practice-issued accounts, on devices the practice manages. Any other device needs written approval from the policy owner. The Approved AI Tools List records each tool’s plan, approved and prohibited uses, the information it may hold, whether a business associate agreement is in place, its output reviewer and its review date. The policy owner maintains the list and reviews it at least [quarterly].

4. Information you may enter. Patient information may be entered only into a tool the list marks as approved for patient information, with a business associate agreement in place. Even then, use only the minimum information the task needs. Removing a patient’s name does not make information safe to share. Passwords and other login credentials are never entered into an AI tool. Confidential practice information is entered only where the list allows it.

5. Requesting a new tool. Send requests to [request channel]. The policy owner responds within [timeframe] and names an approved alternative to use in the meantime where one exists. Before approval, the practice reviews whether a business associate agreement is required and covers the proposed use, how the vendor stores, keeps and deletes data, whether it uses practice data to train AI, and who can access it. New tools may start with a limited pilot.

6. Installs, accounts and connections. Personal AI accounts may not be used for practice work. Installing AI apps or browser extensions, or connecting AI tools to practice email, calendars or files, requires approval. AI notetakers and scribes may be used only where approved and only with the practice’s notice and consent process.

7. Reviewing AI output. AI-generated communications are reviewed before they’re sent, except inside a separately approved automated workflow described below. AI-generated clinical documentation is always reviewed by the clinician. Each approved use has a named reviewer: [role] for patient communications, the treating clinician for clinical notes, qualified billing staff under [role]’s process for coding and billing. Diagnosis and treatment decisions remain with the licensed clinician. Any AI that talks to patients or books appointments on its own needs separate approval, set limits, a named owner, monitoring and a way for the patient to reach a staff member.

8. Reporting. Report any use of an unapproved AI tool for practice work, any misuse of an approved tool, and any suspected unauthorized use or disclosure of patient information involving AI to [privacy officer] within [timeframe]. Report AI behavior that could affect patient care or information security through the same channel. [Optional: Prompt self-reporting will be considered when applying sanctions.]

9. Monitoring. The practice monitors software, browser extensions, connected apps and web use on practice systems to enforce this policy.

10. Sanctions. Violations are handled under the practice’s sanction policy.

11. Training and acknowledgment. Workforce members complete training on this policy at hire and [annually], and sign an acknowledgment each time.

12. Review. The policy owner reviews this policy at least [annually] and whenever the practice adds an AI tool, a vendor changes its terms or the law changes. Each version is kept for six years from the date it was created or last in effect, whichever is later.

One last thing about that template. Signed and filed, it sits in a drawer. Make the approved tools the easiest ones to reach on every workstation and block what isn’t approved, and your team faces the same tool decision far less often. Training still covers patient information, consent and reporting. If your standard depends on anybody remembering it, you do not have one yet.

Posted in AI in Dentistry, Dental Cybersecurity

Filter By: