Can home care agencies use AI tools without compliance risk?

The question I get from home care administrators is almost always framed as a yes or no. Can we use this. What I have found is that the yes or no framing is what creates the risk, because it pushes agencies toward one of two bad outcomes: a blanket ban that staff quietly ignore, or a blanket permission that nobody scoped.

Both produce the same result. People use the tools anyway, off the record, with whatever data is in front of them. The difference is whether you know about it.

Here is the framework I would use instead.

There is no LHCSA-specific AI rule, and that is the problem

Administrators keep waiting for the Department to publish something that tells them what is allowed. As of this writing there is no LHCSA-specific artificial intelligence regulation to wait for.

That absence gets misread as permission. It is not. It means your exposure comes from rules that already exist and already apply: HIPAA, the New York SHIELD Act, patient confidentiality obligations under 10 NYCRR Part 766, and the plain expectation that a written policy is both maintained and actually followed.

None of those rules mention AI. All of them apply to it.

Question one: does protected health information touch the tool?

Under HIPAA, a business associate is any person or organization that creates, receives, maintains, or transmits protected health information while performing a function or service for a covered entity. That definition is about what the vendor does with the data, not what category of software it is. An AI vendor processing your patient information is a business associate by function.

The Privacy Rule requires that before a covered entity discloses protected health information to a business associate, it first obtain satisfactory assurances, documented through a signed Business Associate Agreement. The sequencing is not negotiable. Without an executed BAA, disclosing protected health information to a vendor is itself an impermissible disclosure, regardless of whether the vendor's infrastructure happens to be secure.

Consumer tiers are the trap here. Reporting on vendor BAA availability as of 2026 indicates that ChatGPT Free, Plus, Pro, Team, and self-serve Business are not eligible for a BAA, while BAAs are available through the API and sales-managed enterprise paths. Vendor terms differ, change, and depend on the specific product tier and configuration. Verify against the vendor's current compliance documentation rather than against an article, including this one.

Question two: the one nobody asks

Here is where home care agencies get caught, and it has nothing to do with patients.

The New York SHIELD Act applies to any person or business that owns or licenses computerized data including the private information of a New York resident. Private information covers Social Security numbers, driver's license numbers, financial account and card numbers, biometric information, and email addresses or usernames paired with passwords or security answers. Every employer with employees in New York is covered, because a name combined with a Social Security number meets the definition. There is no small business exemption, though size is taken into account in judging whether safeguards are reasonable.

Now think about what is in a caregiver personnel file. Social Security number. Driver's license. Direct deposit details. Possibly biometric time clock data.

An agency can create real exposure using AI for recruiting, scheduling, or human resources without a single patient record ever entering the system.

This is the gap I see most often. An administrator correctly locks down patient data, then lets the scheduling coordinator paste a roster with employee identifiers into a general assistant to reformat it, because that feels like an operations task rather than a compliance one.

The SHIELD Act requires reasonable safeguards, and a business is deemed compliant if it maintains a data security program with defined elements. Several of them speak directly to AI adoption:

• Designating one or more employees to coordinate the security program.
• Identifying reasonably foreseeable internal and external risks.
• Training and managing employees in the program's practices and procedures.
• Selecting service providers capable of maintaining appropriate safeguards, and requiring those safeguards by contract.
• Securely disposing of private information within a reasonable time after it is no longer needed.

Read the fourth one again. Selecting capable service providers and binding them contractually is exactly the vendor diligence question an AI tool raises. And breach notification runs on a 30-day clock from discovery.

One useful piece of relief: a business that complies with the HIPAA Security Rule is deemed compliant with the SHIELD Act's data security requirements. If you build one program well, it covers both.

"HIPAA compliant" is not a product badge

Treat any vendor claim of blanket HIPAA compliance as marketing until proven otherwise. Eligibility depends on the exact service, the tier licensed, whether a BAA is executed, and how the deployment is configured. Large vendors typically cover some products and not others under the same corporate BAA, which means an agency needs a current inventory of which specific services are covered rather than a general belief that it has an agreement with the company.

Two further points that matter more than they sound:

A covered entity cannot delegate its HIPAA obligations to a vendor. Using an AI system does not relieve you of the obligation to ensure protected health information is protected, and a signed BAA alone does not establish compliance. If the vendor experiences a breach, you are responsible for breach notification to affected individuals and to HHS.

Second, in 2024 OCR issued guidance clarifying that the Privacy Rule's nondiscrimination provisions apply to AI-driven decision-making, with obligations that overlap Section 1557 of the Affordable Care Act. For a home care agency, the live version of that is any AI-assisted screening, matching, or scheduling logic that could produce disparate outcomes by protected characteristic. If you are piloting AI in hiring, that is the exposure to think through before the pilot, not after.

A workable rule your staff can actually follow

Policies fail when they require judgment at the moment of use. The version that works sorts data into tiers, so a coordinator at 4pm on a Friday does not have to reason from first principles.

The four-tier data rule

Tier	What it is	General AI tool
1. Public	Job descriptions, training outlines, marketing copy, general policy language, meeting agendas with no names.	Permitted
2. Internal	Process documents, de-identified aggregate metrics, draft procedures, anonymized scenarios for training.	Permitted with review
3. Personnel	Anything containing caregiver identifiers: SSN, license number, direct deposit, biometrics, home address.	Covered vendor only
4. Patient	Any protected health information: names, conditions, plans of care, activity sheets, progress notes, addresses.	BAA required, no exceptions

Most of the genuine productivity gain in a home care office sits in tiers one and two, which is the part administrators tend not to believe until they try it. Drafting an in-service module. Rewriting a policy in plain language for a bilingual workforce. Turning a messy set of exit reasons into a themed summary once the names are stripped. Building the supervisor coaching script. None of that requires a single identifier.

What a surveyor will actually ask

This is the part I want administrators to internalize, because it is where the compliance conversation usually stops too early.

New York surveyors already apply a specific two-part test in another domain. In the criminal history record check review, they check whether the agency has a written policy addressing every required element, and then whether the agency has implemented and follows that policy. A policy that exists and a policy that is operated are two different findings.

Assume the same test arrives for AI. That means the defensible position is not a ban and not a memo. It is:

That last item is the one agencies skip, and it is the one that converts a good policy into evidence. This is the same failure I described in the survey readiness piece: the policy exists, the proof that anyone followed it does not.

The competitive read

I will be direct about why I think this matters beyond avoiding a citation.

The agencies that answer the monitoring question first get to use these tools first. Right now most of the sector is frozen, waiting for guidance that is not coming, while their staff use consumer tools unsupervised because the work still has to get done. The gap between official policy and actual practice is the risk. Closing it is not primarily a technology project. It is a policy, training, and documentation project, which is to say it is a workforce project.

An agency that can hand a surveyor a tool inventory, a signed BAA, a data tier rule, and dated training records is not just compliant. It is free to move.