Lesson objective

Attach AI governance to the policies, training, sanctions and evaluation a practice already maintains, instead of running a second programme beside them.

Plain-language explanation

Why this matters now

The commonest failure of a practice AI programme is not that it is wrong. It is that it is separate. A second binder, a second training session, a second review cycle, a second owner, and within a year it is the binder nobody opens while the real compliance work continues without it.

The fix is to attach every part of the AI programme to a duty the practice already performs. Those duties are specific and already have owners.

  • Law / rule (45 CFR 164.308(a)(1)(ii)(A) and (B); binds covered entities and business associates): risk analysis and risk management are Required implementation specifications. (C) Sanction policy is Required, and (D) Information system activity review is Required, such as audit logs and access reports.
  • Law / rule (45 CFR 164.308(a)(5)(i)): implement a security awareness and training programme for all workforce members, including management.
  • Law / rule (45 CFR 164.308(a)(8)): perform periodic technical and nontechnical evaluation, including in response to environmental or operational changes affecting the security of electronic protected health information.
  • Law / rule (45 CFR 164.306(b)(2)): in choosing security measures a practice must consider its size, complexity and capabilities; its technical infrastructure; the costs; and the probability and criticality of potential risks. 45 CFR 164.306(d) marks implementation specifications Required or Addressable, and 164.312 sets unique user identification as Required, with encryption at rest and transmission encryption Addressable, alongside audit controls and person or entity authentication.

Read that list as a set of hooks. The practice already has a risk analysis, a sanction policy, an activity review, a training programme and an evaluation trigger. Every one of them has room for AI, and none of them needs a parallel version.

Note what the Security Rule does not currently say. Encryption is Addressable, not Required, and a January 2025 proposal to change that has not been finalised. Teach the current rule. "Addressable" also does not mean optional: the practice implements the specification where reasonable and appropriate, or adopts an equivalent alternative and documents why.

A question to sit with

If your AI policy disappeared tomorrow and only your existing compliance documents survived, how much of your AI governance would still be running?

What goes wrong

The parallel programme. AI gets its own policy, its own training and its own annual review, all on a different cycle from everything else. Two cycles means the weaker one lapses.

Training that is a slide. A single slide reading "do not put patient information into AI tools" changes nothing, because it does not tell anyone what to do instead, and it does not match the inventory the practice actually approved.

The sanction policy nobody extended. The practice has a sanction policy, as the Security Rule requires. Nobody has said whether pasting a patient record into an unapproved tool falls under it. When it happens, the response is improvised, and improvised responses are inconsistent.

Audit logs that show nothing about AI. Information system activity review is Required, but the AI features sit inside applications whose logs nobody reads, or behind shared logins that make the logs meaningless.

Adopting a tool without triggering evaluation. 164.308(a)(8) requires periodic evaluation including in response to operational changes affecting the security of ePHI. Switching on a transcription feature in the phone system is such a change. Practices treat it as a settings tweak.

Buying a framework instead of running a programme. Voluntary material is genuinely useful for structure. It is not a substitute for the duties above, and it confers nothing. Lesson 4 handles this in full.

The practical answer

Attach each AI element to the compliance element it belongs inside. The mapping is an AI Coalition Network recommendation; the left column is the duty, the right column is what to add.

Existing duty What AI adds to it
Risk analysis, 164.308(a)(1)(ii)(A) Each inventory row that touches ePHI becomes an entry: what information, where it goes, who can reach it
Risk management, (B) The controls chosen for each row, including review depth and access restriction
Sanction policy, (C) Explicit examples: patient information into an unapproved tool; sending a clinical output without the required review; marking an item reviewed without reading it
Activity review, (D) Who accessed which AI feature, and who marked outputs reviewed; only meaningful with unique logins
Training, 164.308(a)(5)(i) Fifteen minutes inside the existing session: the three lists, the four review depths, what to do when a tool would help but is not approved
Evaluation, 164.308(a)(8) Any new AI feature or vendor default change is an operational change that triggers the evaluation
Policies and procedures The AI programme is a section, not a separate binder
Unique user identification, 164.312 Required, and the precondition for every record of who reviewed what

Write the training so it matches the inventory. An AI Coalition Network recommendation for the fifteen minutes: name the three lists and where they live; show two approved uses and two declined ones, with reasons; state the four review depths in the learner's own role; give the exact route for requesting a new tool and the turnaround; name the stop authority holder; and say plainly what happens if someone has already used something unapproved, including that reporting it promptly is the expected behaviour.

Extend the sanction policy with worked examples, not new categories. A practice rarely needs a new sanction. It needs to say which existing category covers which AI behaviour, and to distinguish three cases: a genuine mistake reported promptly; a repeated breach after training; and deliberate concealment. Writing those distinctions before an incident is what makes the response consistent.

Use the voluntary security material for structure, honestly.

  • Professional guidance (HHS 405(d) Program, Health Industry Cybersecurity Practices, 2023 Edition; voluntary, industry-led and consensus-based, not a HIPAA requirement): it names five current threats — social engineering; ransomware; loss or theft of equipment or data; insider, accidental or intentional data loss; and attacks against network connected medical devices — and sets out ten Cybersecurity Practices, including email protection, endpoint protection, access management, data protection and loss prevention, asset management, network management, vulnerability management, incident response, network connected medical devices, and cybersecurity oversight and governance. The practices "are not prioritized in any specific order", the document recommends beginning with a risk assessment, and Technical Volume 1 covers small health care organizations. Public Law 116-321 includes approaches promulgated under section 405(d) among "recognized security practices". The 2023 edition does not mention generative AI.
  • Professional guidance (NSA AI Security Center with CISA, FBI and international partners, Deploying AI Systems Securely, released April 15, 2024; voluntary joint guidance): it is intended for organisations deploying and operating externally developed AI systems, which is exactly what a dental practice does. Its stated goals are to improve the confidentiality, integrity and availability of AI systems, ensure appropriate mitigations for known vulnerabilities, and provide methodologies and controls to protect, detect and respond to malicious activity against AI systems and related data and services.

Both are voluntary and neither is dental. Their value is that a practice adopting an externally developed tool is doing the thing the second document was written about, and the first tells it which threat categories to think about. Neither creates a duty, and neither can be certified to.

Remember this

  • Attach the AI programme to the risk analysis, risk management, sanction policy, activity review, training and evaluation the practice already runs.
  • Adopting or enabling an AI feature is an operational change under 164.308(a)(8) and triggers evaluation.
  • Unique user identification is Required under 164.312 and is what makes any record of review meaningful.
  • Extend the sanction policy with worked examples and distinguish a reported mistake from concealment.
  • HICP 2023 and the 2024 joint deployment guidance are voluntary, non-dental and confer nothing; use them for structure and say so.

Dental practice example

Larkspur Dental Group PLLC, fictional, has a compliance binder that predates every tool in this module. Its security official, Priya Ramanathan, is handed a nine-page draft AI policy by the group office and asked to roll it out. She does the one-binder test instead. Page by page, she tries to place each paragraph inside an existing section. The paragraph on what patient information may enter a tool belongs in the risk analysis, where the group already lists its systems; she moves it and deletes the standalone version. The paragraph on staff misuse belongs in the sanction policy; rather than writing a new category she adds three worked examples and the distinction between a mistake reported the same day, a repeat after training, and concealment. The paragraph on training becomes fifteen minutes inside the annual session that already exists, rewritten so it matches the group's actual approved list rather than a generic one. The paragraph headed "AI annual review" she deletes outright, because the group already performs periodic evaluation under 164.308(a)(8), and she instead writes one sentence into that procedure: enabling a new AI feature, or accepting a vendor default change, is an operational change that triggers the evaluation. Two paragraphs survive as genuinely new, and she keeps them with a note saying why: the four review depths, and stop authority. Then she finds the gap the test was worth doing for. Activity review is Required, and the group does read logs, but three of the eleven hygienists share a clinical workstation login at one site, so the log cannot show who marked a drafted note reviewed. Unique user identification is Required under 164.312, and the sharing predates the AI question entirely. She raises it as a Security Rule finding rather than an AI finding, which is both accurate and the reason it gets fixed within the month. The nine-page policy becomes two pages inside the existing binder. The group office objects that the draft has been cut. Priya's answer is that a programme kept in one binder, on one cycle, with one set of owners, is the only kind that is still running next year.

Responsible-use warning

Do not run a second programme. An AI policy with its own binder, its own training and its own annual cycle competes with the compliance work that already has owners and deadlines, and it is the one that lapses. Attach each element to the duty it belongs inside, and delete whatever is duplicated. Do not describe HICP, the joint deployment guidance or any framework as a requirement, and do not write the proposed Security Rule encryption changes into a policy as current duties; encryption remains Addressable as the rule stands.

Human-review requirement

The security official owns the risk analysis, risk management, activity review and evaluation entries, and the privacy official owns the training and sanction content; both act under the practice owner, who owns the programme. They check that every AI element sits inside an existing compliance section with the same owner and cycle, that enabling an AI feature triggers the evaluation required after an operational change, that unique user identification is in force so review records mean something, and that no voluntary framework is described as a requirement or as something the practice can be certified against.

Practical exercise

Work with your own compliance documents, or fictional equivalents, and no patient information.

  1. List the sections your practice already maintains: risk analysis, risk management, sanction policy, activity review, training, evaluation, policies and procedures.
  2. Take your draft AI policy and place each paragraph into one of those sections. Mark anything that will not fit as either genuinely new, with a reason, or duplicated, and delete the duplicates.
  3. Add each inventory row that touches electronic protected health information to the risk analysis, recording what information enters the tool, where it goes and who can reach it.
  4. Write three worked examples into the sanction policy, and write the distinction between a mistake reported promptly, a repeat after training, and deliberate concealment.
  5. Draft the fifteen minutes of training as a script, using your own approved and declined lists rather than generic examples. Include the request route and its turnaround.
  6. Add one sentence to your evaluation procedure making an enabled AI feature or an accepted vendor default change an operational change that triggers evaluation.
  7. Check unique user identification across every system with an AI feature. List every shared login, name its owner and write the date it will be removed.
  8. Pull one month of access logs for the system with the most AI activity and try to answer: who marked outputs reviewed. Record what the logs can and cannot show, and treat any gap as a Security Rule finding rather than an AI finding.

Knowledge check

  1. Which Required implementation specifications under 45 CFR 164.308 give an AI programme its hooks into existing compliance work?

    Answer guidance: Risk analysis at (a)(1)(ii)(A), risk management at (B), the sanction policy at (C) and information system activity review at (D) are all Required, along with the security awareness and training programme at (a)(5)(i) and periodic evaluation at (a)(8). Each already has an owner and a cycle, so AI content belongs inside them rather than in a parallel programme.

  2. Why does enabling a new AI feature trigger the evaluation required by 45 CFR 164.308(a)(8)?

    Answer guidance: That provision requires periodic technical and nontechnical evaluation, including in response to environmental or operational changes affecting the security of electronic protected health information. Switching on a transcription, drafting or summarising feature changes how ePHI is created, handled or transmitted, so it is an operational change rather than a settings tweak.

  3. What is the status of encryption under the current Security Rule, and what must a practice not do with the January 2025 proposal?

    Answer guidance: Under 45 CFR 164.312 as the rule currently stands, encryption and decryption at rest and transmission encryption are Addressable, not Required, while unique user identification is Required. Addressable means the practice implements the specification where reasonable and appropriate or adopts an equivalent alternative and documents why. The January 2025 proposal is not final and must not be taught or written as a current duty.

  4. What do HICP 2023 and the 2024 joint guidance on deploying AI systems offer a dental practice, and what do they not?

    Answer guidance: HICP 2023 is voluntary, industry-led guidance naming five current threats and ten cybersecurity practices that are not prioritised in any order, with Technical Volume 1 for small organisations; it does not mention generative AI. The April 15, 2024 joint guidance is voluntary and is written for organisations deploying externally developed AI systems. Neither creates a duty for a dental practice, neither is dental, and neither can be certified against.

Sources and further reading