Until recently, CMS guidance on the use of AI in home health has been patchy at best, leaving agencies unsure of how to incorporate AI in a way that’s fully compliant.
At last, CMS has issued clear guidance in the OASIS E2 update: AI is permitted, with firm guardrails.
The clearest CMS guidance on AI and OASIS documentation comes from the OASIS-E2 Manual. It essentially states that: while AI may be used to assist clinicians in assessing the patient, the human is ultimately responsible for the accuracy of the final answer.
Here’s a direct quote from CMS (emphasis ours):
An agency’s software may not “answer” or “generate” a final code for the OASIS items. Following agency policies, the assessing clinician is responsible for considering available information and ensuring the appropriate OASIS item response(s) were selected, within the appropriate timeframe and consistent with data collection guidance
This isn’t merely Andy’s interpretation. CMS restated the same position in their April 2026 Q&A, in response to this question from an agency:
Question 1: A clinician is completing a Start of Care assessment. Their iPad uses an ambient listening AI platform, with the patient’s consent, that populates some of the OASIS responses while the assessment is being conducted. After completing the assessment, the clinician reviews the OASIS items one by one, making corrections on any items they feel were incorrectly coded by the AI platform.
Answer 1: As stated in the final version of the OASIS-E2 Guidance Manual, “an agency’s software may not “answer” or “generate” a final code for the OASIS items. Following agency policies, the assessing clinician is responsible for considering available information and ensuring the appropriate OASIS item response(s) were selected, within the appropriate timeframe and consistent with data collection guidance.”
The evolution of CMS’ guidance, from the draft, to the final version, and through the Q&A all reflect the same underlying principle that while AI tools may assist, clinicians are ultimately responsible for the final accuracy.
The draft manual left things vague and invited questions:
An agency’s software may not “answer” or “generate” the OASIS response for the assessing clinician
Between the publication of the draft and final version, CMS clarified 2 things. First, that the software may not generate the final OASIS response:
An agency’s software may not “answer” or “generate” a final code for the OASIS items.
Second, that clinicians are responsible for verifying the correct response was entered, however the initial responses were entered in the first place.
Following agency policies, the assessing clinician is responsible for considering available information and ensuring the appropriate OASIS item response(s) were selected, within the appropriate timeframe and consistent with data collection guidance
The guidance around AI builds on existing principles, namely collaboration with other professionals to gather information for OASIS items.
The manual explicitly allows collaboration with the patient’s family members, caregivers, other health personnel, whether they’ve had direct contact with the patient, or have other ways to understand the patient’s situation.
See the quote from CMS below:
While only the assessing clinician is responsible for accurately completing and signing a comprehensive assessment, they may collaborate to collect data for all OASIS items, if agency policy allows. Collaboration may consider information from others such as the patient, caregivers, and other health care personnel, including the physician, pharmacist, and/or other agency staff who have had direct contact with the patient or had some other means of gathering information to contribute to the OASIS data collection.
Saying that humans must review every answer is not the same as designing a product that helps the review actually happen.
At a minimum, a tool shouldn’t block or discourage clinicians from viewing OASIS responses. If an AI tool makes it difficult for the clinician to access the final scores, it’s unlikely to be compliant.
At the other extreme, a tool which raises too many errors or questions will quickly create notification fatigue. Clinicians who are buried with flags and errors will quickly learn to tune them out.
In other words, the most compliant tool should not hinder human review. In an ideal world, it can intelligently direct a clinician’s attention to the most important questions.
Andy