
Kurt FischmanFounder, Machine-like
Kurt is the CEO of Machine-like, the Managed Agent Operations platform.

Before sharing protected health information with an AI vendor, practice owners must verify five contractual requirements: an executed Business Associate Agreement covering the active product tier, an explicit ban on model training, a documented list of downstream subcontractor agreements, zero data retention for prompts and transcripts, and rapid incident notification. A vendor's marketing claim of HIPAA compliance creates no legal shield without these binding terms.
Practice owners often hear that modern artificial intelligence will eliminate administrative overhead, transcribe patient calls, and organize intake records in seconds. The software demo looks seamless: an intake form arrives, an automated assistant extracts the relevant history, drafts an appointment summary, and stages the record for staff review. Yet beneath the user interface lies a serious legal and operational reality. When your clinic, dental group, or therapy practice routes patient data into an external AI tool, you aren't just buying software. You're handing protected health information to a chain of cloud services, foundation model providers, and data pipelines.
A badge on a vendor's website that says "HIPAA Compliant" doesn't protect your practice from federal penalties.1 Under the Health Insurance Portability and Accountability Act, protected health information includes any individually identifiable health data transmitted or maintained in any medium. If an AI vendor processes patient names, clinical notes, intake forms, or phone recordings without an executed Business Associate Agreement, that disclosure is unlawful the moment the data leaves your system.2
Protecting your business requires looking past sales demonstrations and asking five precise questions about legal agreements, data retention, access boundaries, subcontractor architecture, and incident handling.
Software vendors frequently confuse security certifications with statutory compliance. A vendor might hold a SOC 2 Type II report, encrypt databases at rest, and host application servers on certified cloud infrastructure. Those safeguards are essential technical hygiene, but they don't fulfill federal privacy mandates.3
Under federal rules, any external entity that creates, receives, maintains, or transmits protected health information on behalf of a covered healthcare practice qualifies as a business associate.1 The law requires the practice and the vendor to sign a Business Associate Agreement before a single patient identifier enters the platform. The agreement specifies permitted uses of the data, mandates administrative and technical safeguards, and establishes legal accountability.
The crucial trap for practice founders is product-tier scope. Major software providers often offer business associate agreements only on dedicated enterprise plans or through custom enterprise contracts.4 If your staff signs up for a standard business subscription or uses a self-serve web interface with a credit card, that account operates under standard consumer or commercial terms of service. Those commercial agreements almost always state that the user agrees not to input protected health information. In that scenario, your practice bears full liability for any unauthorized disclosure, while the vendor's terms explicitly disclaim responsibility.
You can't rely on verbal promises from account managers or generic compliance landing pages. Your practice needs an executed agreement signed by an authorized corporate officer that identifies the exact software tier, feature set, and application programming interfaces your team uses.5
The table below contrasts standard sales reassurances with the contractual evidence your practice must inspect before sharing protected health information.
| Diligence Area | What the Vendor Claims | The Hidden Exposure | What to Verify in Writing |
|---|---|---|---|
| Agreements and Scope | "Our platform is fully HIPAA compliant." | Marketing claims don't create legal protection; without a signed agreement, sharing data is an immediate violation.1 | An executed BAA specifically covering the exact product tier, features, and API endpoints you deploy. |
| Model Training | "We take data privacy seriously and protect user inputs." | Standard terms often claim licenses to use prompts and text for service enhancement or product tuning.3 | An explicit contractual clause stating patient data, prompts, and audio will never train or tune any model. |
| Downstream Subcontractors | "We run on enterprise-grade cloud servers." | Application vendors pass prompts to third-party foundation models that may lack downstream BAAs.4 | A complete subprocessor list confirming valid subcontractor BAAs with every underlying model and host. |
| Data Retention | "Your session data is encrypted and secure." | Commercial model endpoints often store prompts and outputs for 30 days for abuse monitoring.6 | Contractual Zero Data Retention or an automated, verified schedule for purging prompts, logs, and embeddings. |
| Incident Notification | "We notify customers in accordance with HIPAA standards." | Federal law allows up to 60 days to report a breach, leaving you too little time to investigate.7 | A binding requirement to report suspected security incidents and breaches within 24 to 72 hours. |
Never treat verbal security assurances or marketing logos as legal protection; every safeguard requires explicit language in your Business Associate Agreement.
Evaluating an AI vendor requires inspecting how the software processes text, where information travels, and who holds legal custody along the way.
Every founder should require clear written answers to the areas outlined above before granting software access to practice management tools or electronic health records. When a vendor hesitates or provides vague generalities about encryption, that's your signal that their operational architecture hasn't caught up with healthcare regulations.
Most healthcare AI applications don't build or run their own proprietary foundation models from scratch. Instead, the application serves as an orchestrator: it captures patient input, constructs a prompt, and transmits that prompt over an application programming interface to an external foundation model provider or specialized speech recognition service.4
Under the HIPAA Omnibus Rule, any downstream provider that processes or stores protected health information on behalf of your primary vendor is legally classified as a subcontractor business associate.1 The primary vendor is legally obligated to execute a matching subcontractor agreement with every external service in its pipeline.2
That legal chain is only as strong as its weakest link. If your practice executes a solid agreement with an AI intake software vendor, but that vendor routes patient audio through a third-party transcription service without a subcontractor agreement, your patient data has been unlawfully disclosed. Similarly, if the vendor's primary server fails over to an unapproved model host during a traffic spike, that fallback path violates your agreement.
Ask the vendor for a current, written list of all sub-processors that touch data. The list must identify the specific model providers, hosting facilities, database managers, and monitoring tools involved in the workflow.4 Verify that the vendor's agreement commits them to notifying you in advance whenever a sub-processor is added or replaced, with a contractual right to terminate if the new provider doesn't meet your compliance standards.
When an automated agent drafts a patient summary or processes an intake packet, data moves through several processing states: temporary memory buffers, application logs, vector databases, and model caches. Standard commercial terms often allow vendors to retain input prompts and completions for thirty days or longer to monitor for abuse or debug software errors.6
In a healthcare environment, thirty days of unencrypted or vendor-accessible prompt logs creates unacceptable exposure. If an unauthorized party compromises the vendor's internal logging dashboard, thirty days of patient histories, medication lists, and contact records sit exposed.
Healthcare practices should demand contractual Zero Data Retention on all model processing endpoints.6 Zero data retention ensures that once the model generates its completion, the input prompt and output text are instantly purged from the model provider's memory and disk caches, rather than saved in an abuse-monitoring queue.
Beyond technical storage, you must examine human access boundaries. Inquire whether the vendor's engineering staff, customer support technicians, or external data annotators have access to raw patient records.7 Some AI startups employ human reviewers to inspect transcripts when a model encounters low confidence scores. If humans review patient records, those reviewers must undergo background screening, receive documented privacy training, and operate under strict role-based access controls with immutable audit logging.
Furthermore, your agreement must explicitly bar secondary model training.3 In standard enterprise software, vendors often reserve the right to aggregate customer data to train future models or improve system accuracy. Under HIPAA, secondary model training on protected health information isn't a recognized administrative right unless explicitly permitted by the agreement.3 Your contract must state clearly that customer data will never be used to train, tune, or evaluate public or proprietary models.
Managing regulatory risk doesn't mean your practice must avoid modern automation entirely. The key to successful adoption is setting disciplined operational boundaries: scope AI systems exclusively for bounded, nonclinical administrative tasks, while keeping clinical and advisory decisions firmly in the hands of qualified human staff.5
AI agents excel at administrative coordination. They can check whether an incoming intake packet contains all required insurance cards, flag missing signature pages, coordinate calendar availability, and confirm that external lab reports have arrived before a consultation. These tasks involve structured data verification and schedule management rather than medical judgment.
By contrast, an automated system should never triage patient symptoms, recommend diagnostic steps, interpret test results, or offer medical advice. If an intake agent notices a patient mentioning chest discomfort or severe shortness of breath, the software shouldn't attempt to assess urgency. It must immediately escalate the interaction to front-desk staff or clinical triage according to practice protocols.
Establishing clear boundaries also simplifies vendor evaluation. When you restrict your automated workflows to administrative tasks, such as verifying missing documents or coordinating appointment reminders, you minimize the clinical exposure of the workload. You maintain a documented audit trail showing that licensed clinicians make all clinical decisions, while automated agents simply handle the repetitive clerical follow-through that keeps the practice running on schedule.
No. An AI vendor cannot use protected health information to train, fine-tune, or evaluate machine learning models unless your Business Associate Agreement explicitly authorizes that secondary use.3 Under federal privacy regulations, business associates can only use patient information for purposes directly necessary to deliver the contracted service, or for narrowly construed administrative functions.3 Standard terms of service that grant broad product-improvement licenses conflict with HIPAA requirements and shouldn't be accepted by healthcare practices.
The vendor must notify your practice promptly in writing, providing details on what records were exposed, which individuals were affected, and what immediate containment measures were taken.7 Although HIPAA regulations allow business associates up to sixty calendar days to report a formal breach, healthcare practices should contractually require notification within twenty-four to seventy-two hours of discovering any suspected incident.7 This rapid notification window ensures your practice has adequate time to investigate the scope of the exposure and meet mandatory reporting obligations.
No, your practice doesn't need to sign a direct agreement with the vendor's cloud host, provided your primary vendor has executed a valid subcontractor Business Associate Agreement with that host.1 Under the HIPAA Omnibus Rule, the chain of custody flows from your practice to the primary vendor, and from that vendor down to its infrastructure subcontractors.2 However, your practice must verify that this downstream agreement is actively in place and that the primary contract binds the vendor to maintaining compliant agreements across its entire infrastructure.4
Your work should be about creativity, solving problems, and building relationships.
Leave the machine-like tasks to us.