Lesson objective
Map each part of an AI governance programme onto the compliance elements a practice already operates, so nothing runs as a separate structure.
Plain-language explanation
Why this matters now
A practice that builds AI governance as a separate structure builds something with its own meetings, its own documents, its own reporting line and its own quiet death. The programme survives when it is carried by activity the practice already performs for other reasons: the policies it already writes, the training it already delivers, the auditing it already runs, the disclosure route it already publishes.
The frame to hang it on is already public and already voluntary.
- Professional guidance (HHS-OIG, General Compliance Program Guidance, published 6 November 2023; OIG states it "is voluntary guidance" and "is not binding on any individual or entity", and uses "should" throughout to present voluntary, nonbinding guidance): its seven elements are written policies and procedures; compliance leadership and oversight; training and education; effective lines of communication with the compliance officer and disclosure programs; enforcing standards through consequences and incentives; risk assessment, auditing and monitoring; and responding to detected offenses and developing corrective action initiatives. Section IV.A names individual and small-group physician practices as small entities and suggests a small entity unable to support a compliance officer consider designating a compliance contact who, whenever possible, is not involved in billing, coding or claim submission and who reports at least quarterly to the owner or CEO. Element 6 recommends data analytics: all entities should be able to compare standard metrics internally to find outliers, and may use billing software and electronic health records with analysis components. The GCPG states that even an entity making "an innocent billing mistake" still has an obligation to repay the money, and cites the requirement to report and return overpayments within 60 days of identification at 42 U.S.C. 1320a-7k(d).
Two limits on that paragraph matter enormously. The guidance is voluntary and not binding on anyone, and it does not mention artificial intelligence anywhere. Using its seven elements as the skeleton for an AI programme is this programme's recommendation and the practice's own choice. Write that sentence into the document, because a policy that implies OIG asked for an AI programme has invented an authority in its first paragraph.
There is a second, voluntary structure some practices meet through a supplier or a health-system partner.
- Standard (ISO/IEC 42001:2023, published 18 December 2023, Edition 1, developed by ISO/IEC JTC 1/SC 42; voluntary, binding only where adopted by contract or by an organisation seeking conformity): ISO describes it as specifying "requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS)", as a management system standard using the Plan-Do-Check-Act methodology that, "rather than looking at the details of specific AI applications", manages AI risks and opportunities across an organisation, and as intended for "organizations of any size involved in developing, providing, or using AI-based products or services". ISO calls it "the world's first AI management system standard". Its text is paid, so nothing beyond ISO's public description can be quoted, and no clause numbers or Annex controls may be cited from that description.
Plan-Do-Check-Act and the seven elements are not rivals. A practice that already thinks in compliance elements should stay there. A practice whose group or supplier speaks the management-system language can map onto that instead. What no practice should do is run both and neither.
A question to sit with
Name the meeting at which your AI programme will be discussed in eighteen months' time, when the novelty has gone and the person who built it has left. If no such meeting already exists in the diary for another purpose, the programme has no carrier.
What goes wrong
The parallel structure. An AI committee, an AI policy binder and an AI report, all separate from compliance. It runs for two quarters on enthusiasm and then competes with clinic for time, and clinic wins.
Training that is a one-off. The practice runs an AI session in March. New staff in June learn from whoever sits nearest. Training belongs in onboarding and in the annual cycle the practice already runs, not in a memorable afternoon.
No route to raise a concern. A medical assistant notices a drafted reply that invents a medication instruction and has nowhere to put it. OIG's fourth element is effective lines of communication with the compliance contact and a disclosure route; the AI programme uses the same route rather than inventing a form nobody remembers.
Standards enforced only downwards. Element 5 is enforcing standards through consequences and incentives. A rule that binds medical assistants and not partners is not a rule; it is a class system, and the staff can all describe it precisely.
Monitoring that never produces a number. Element 6 asks for risk assessment, auditing and monitoring, and suggests comparing standard metrics internally to find outliers. An AI programme with no measurement is a set of intentions.
Finding an error and stopping there. Element 7 is responding to detected offenses and developing corrective action. If AI-assisted work produced a claim that should not have been paid, the response is not confined to fixing the tool: the GCPG notes that even an innocent billing mistake carries an obligation to repay, and cites the 60-day report-and-return requirement. That is a legal exposure with its own clock, handled with the practice's advisers.
The practical answer
Map the AI programme onto the seven elements, one line each. This mapping is this programme's recommendation; the elements are voluntary guidance that does not mention AI.
| Element | What the AI programme adds | Where it lives |
|---|---|---|
| 1. Written policies and procedures | The permitted-use inventory, the approved tool list, the non-delegable list | With the existing policies and procedures |
| 2. Compliance leadership and oversight | The owner register from Module 01, including the programme owner and stop authority | The same designations file |
| 3. Training and education | An AI module in onboarding and in the annual cycle, role-specific | The existing training record |
| 4. Lines of communication and disclosure | The existing route, explicitly open for AI concerns, with no sanction for raising one | The published route |
| 5. Enforcing standards | The same consequences and incentives, applied to partners as to staff | The existing sanction policy |
| 6. Risk assessment, auditing and monitoring | The sampling from the inventory, with results compared over time to find outliers | The existing audit calendar |
| 7. Responding and corrective action | What happens when AI-assisted work produced an error, including the repayment question | The existing incident and corrective action process |
Add three AI-specific practices the elements do not supply. Each is voluntary and each should be labelled.
- A change trigger. A supplier release, a payer policy change or a new state statute re-opens the relevant inventory row. Module 03 builds the record this leaves.
- A blinded internal reporting route for AI-related safety concerns. The Joint Commission and CHAI guidance includes voluntary, blinded reporting of AI safety-related events among its seven elements, and describes itself as an initial, high-level document written mainly for hospitals. A small practice cannot run a national reporting system, but it can let staff raise an AI concern without attribution, which is often the only way a concern about a partner's use ever surfaces.
- Role-specific documentation and training. The same guidance says organisations should, at a minimum, define and document how users will receive AI tool documentation and role-specific training. A front-desk registrar and a coding professional need different sessions, and a single all-staff session satisfies neither.
Size the programme honestly and write the sizing down. OIG's guidance for individual and small group physician practices states that, unlike other OIG guidances, it "does not suggest that physician practices implement all seven components of a full scale compliance program." A two-clinician practice may run a one-page inventory, an annual review, quarterly sampling of one row and a single named owner. What it may not do is have nothing and call it proportionality.
Remember this
- Build the AI programme inside the compliance activity the practice already runs; a parallel structure dies in two quarters.
- OIG's seven elements are voluntary, not binding, and say nothing about AI; using them as the skeleton is the practice's own choice and should be stated.
- Map each element to a concrete AI artefact and to the existing file, meeting or calendar entry that carries it.
- Add a change trigger, an unattributed route for AI safety concerns, and role-specific training, each labelled voluntary.
- An error found in AI-assisted claim work raises a repayment question with a 60-day clock; handle it with the practice's advisers, not with a software fix alone.
Medical practice example
Larkspur Internal Medicine PLLC already runs a modest compliance programme: a policies binder, an annual training afternoon, a quarterly claims sample and a posted route for raising concerns. Roland Pike, the administrator, is asked by the partners to write the AI programme and arrives with a proposal for an AI governance committee meeting monthly. Dr Imani Sowande asks the eighteen-month question and the proposal does not survive it, because the practice has never sustained a monthly meeting about anything. They rebuild it as seven lines instead. The permitted-use inventory goes into the policies binder as section 12, behind the existing policies rather than in a new folder. The owner register joins the HIPAA designations file. The AI training becomes twenty minutes inside the annual afternoon plus a page in onboarding, with a separate fifteen minutes for the billing office because their row is different. The existing concerns route gets one added sentence saying it covers AI, and a second saying nobody is sanctioned for raising something. The sanction policy is not changed at all, which is the point: the partners are covered by it already, and Dr Sowande insists on saying so aloud at the training. The quarterly claims sample gains a column recording whether AI assisted the work, so that in a year the practice can compare AI-assisted and other work rather than guessing. Corrective action takes the longest conversation. Roland asks what happens if a sample finds a claim that should not have been paid, and the answer he had assumed, namely that they would fix the tool, is only the first half. The partners agree with their adviser that the second half is a repayment question with its own 60-day clock from identification, and they write that into the corrective action page in one sentence with a name against it. They add two voluntary practices: an unattributed route for AI safety concerns, which Roland argues for on the grounds that a medical assistant will never put her name to a concern about a partner draft; and a standing item where whoever reads the supplier release notes reports anything that changes an inventory row. The whole programme is eleven pages. Nothing in it required a new meeting, and eighteen months later it is still running, which is the only test that matters.
Responsible-use warning
Never let the AI programme become a separate structure with its own committee, binder and report. It will compete with clinic for time and it will lose, and the practice will be left with an impressive document describing something that stopped happening. Equally, do not write that OIG or any other body requires an AI governance programme. The General Compliance Program Guidance is voluntary, states that it is not binding on any individual or entity, and does not mention artificial intelligence at all; ISO/IEC 42001 is voluntary and binds only through a contract or a conformity decision. Adopt the structure by choice and say in the document that you chose it.
Human-review requirement
The practice owner or entity CEO owns the decision to adopt a compliance frame for the AI programme and owns the sizing judgement behind it. The compliance contact carries the mapping and reports on it at least quarterly to the owner or CEO, following the cadence OIG suggests for small entities. The treating clinician retains every clinical judgement inside the programme. The owner checks that each of the seven rows names an existing file, meeting or calendar entry rather than a proposed one, that voluntary material is labelled as voluntary, and that any finding touching a paid claim is escalated as a repayment question with the practice advisers rather than treated as a software defect.
Practical exercise
Work with a fictional practice or your own.
- List the compliance activity your practice already performs: policies, designations, training, the route for raising concerns, the sanction policy, any auditing, and the incident process. Note where each physically lives.
- Draw the seven-row mapping table and fill the third column with an existing file, meeting or calendar entry for each row. Mark every row where the honest answer is a new structure.
- Rework each marked row so it is carried by something that already happens.
- Write the sentence your policy will use to describe why it adopts the seven elements, making clear the guidance is voluntary, not binding and silent on AI.
- Draft the AI content for onboarding and for the annual training cycle, with at least two role-specific variants.
- Add one sentence to the existing concerns route naming AI, and one stating that raising a concern carries no sanction. Decide who publishes it and when.
- Check your sanction policy applies to owners and partners in the same terms as to staff. If it does not, write the change and who must approve it.
- Add a column to your existing audit or sampling work recording whether AI assisted the item, so comparison becomes possible within a year.
- Write the corrective action paragraph for the case where AI-assisted work contributed to a claim that should not have been paid, naming who is contacted, the repayment question and the 60-day clock from identification.
- Write your sizing paragraph: what the practice is running, what it is not running, and why that is proportionate for its size and risk.
Knowledge check
-
What are OIG's seven elements, and what two limits must a practice state when using them for AI governance?
Answer guidance: Written policies and procedures; compliance leadership and oversight; training and education; effective lines of communication with the compliance officer and disclosure programs; enforcing standards through consequences and incentives; risk assessment, auditing and monitoring; and responding to detected offenses with corrective action. The two limits are that OIG states the guidance is voluntary and not binding on any individual or entity, and that it does not mention artificial intelligence, so the AI content is the practice own choice.
-
Why does a separate AI committee usually fail in a physician practice?
Answer guidance: Because it has no carrier. A structure with its own meeting, binder and report competes with clinic for time and loses once the novelty fades or the person who built it leaves. A programme survives when each part is carried by activity the practice already performs for another reason: the existing policies file, the existing training cycle, the published concerns route, the existing sanction policy, the existing audit calendar and the existing incident process.
-
What does ISO/IEC 42001:2023 offer a practice, and what may not be quoted from it?
Answer guidance: ISO describes it as specifying requirements for establishing, implementing, maintaining and continually improving an artificial intelligence management system, as a management system standard using Plan-Do-Check-Act that manages AI risks across an organisation rather than examining specific applications, and as suitable for organisations of any size. It is voluntary and binds only through a contract or a conformity decision. Its text is paid, so no clause numbers or Annex controls may be cited, only ISO public description.
-
What does OIG guidance say about an innocent billing mistake, and why does that matter to an AI programme?
Answer guidance: The General Compliance Program Guidance states that even an entity making an innocent billing mistake still has an obligation to repay the money, and cites the requirement to report and return overpayments within 60 days of identification at 42 U.S.C. 1320a-7k(d). An AI programme that finds an error in AI-assisted claim work therefore cannot stop at fixing the tool; the finding raises a repayment question with its own clock, to be handled with the practice advisers.
Sources and further reading
- U.S. Department of Health and Human Services, Office of Inspector General (HHS-OIG) — General Compliance Program Guidance (GCPG) (2023-11-06). Accessed 2026-09-15.
- International Organization for Standardization (ISO) and International Electrotechnical Commission (IEC), ISO/IEC JTC 1/SC 42 — ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system (2023-12-18). Accessed 2026-09-15.
- Joint Commission and Coalition for Health AI (CHAI) — Joint Commission and Coalition for Health AI (CHAI) Guidance on The Responsible Use of AI in Healthcare (RUAIH) (2025-09-17). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.