Healthcare & legal · 2027 planning guide

AI in healthcare and legal work: contracts, review and confidentiality

A planning guide for hospitals, clinics and law firms. Evidence checked 20 November 2026. This page is not legal or medical advice.

Short answer. A BAA is a contract, not a certification, and it does not make a model clinically or legally safe. Hospitals need a signed BAA, PHI minimization, access controls and clinician review. Law firms need confidentiality controls, matter isolation and lawyer supervision. In both settings, the vendor supplies infrastructure; the covered entity or firm remains accountable for the decision. Start with a narrow workflow and document the review.

Quick answers

These answers separate what a contract can provide from what your clinicians, lawyers and compliance team must still own.

Quick answers for healthcare and legal AI procurement, checked 20 November 2026.
Buyer questionOne-sentence answer
Does Plugsky sign a BAA?A BAA is available by contract, subject to review, and the signed document defines the covered scope.
Is Plugsky HIPAA certified?No; HHS does not issue HIPAA certifications, and Plugsky makes no such claim.
Can we send PHI to the API?Only under a signed BAA and a documented minimum-necessary workflow; do not assume a standard plan covers PHI.
Does the model give medical advice?No; output is draft text for a qualified clinician to review, and the vendor does not practice medicine.
Can law firms use it?Yes, with confidentiality controls, matter separation and lawyer supervision; privilege and consent need counsel review.
Is client data used for training?Ask for a contractual commitment; do not infer it from a marketing page.
Does it support Arabic patient records?Arabic OCR and generation are limited; see the Arabic guide and test before use.
What should we pilot?A read-only summarization or intake workflow with a named reviewer and a deletion date.

Healthcare: the BAA is the starting line

A business associate agreement allocates obligations; it does not validate the model or transfer clinical responsibility. Read it with the data flow and the Security Rule in view.

Does Plugsky sign a BAA?

Plugsky offers a BAA by contract, subject to review. That means the document is available, not automatic: the scope, the covered services, the permitted uses, the safeguards and the breach-notification terms have to be agreed. A BAA with the API provider does not cover your other vendors, your internal applications or a downstream model provider unless the contract says so. Put the BAA in place before the first PHI prompt, and confirm that the deployment you are using is the one covered. The signed BAA, not a features page, is the evidence your compliance team needs.

Is Plugsky HIPAA certified?

No. HHS does not issue HIPAA certifications, so no vendor holds an official HIPAA certification. HHS OCR enforces the HIPAA Privacy, Security and Breach Notification Rules, and covered entities and business associates demonstrate compliance through policies, risk analysis, safeguards and documentation. A vendor may hold SOC 2 or ISO 27001 reports, but those are not HIPAA certifications either. If a vendor says it is “HIPAA certified”, ask who issued the certificate. The correct question is whether it will sign a BAA and whether its controls meet the Security Rule.

What counts as PHI in an AI prompt?

PHI is individually identifiable health information held or transmitted by a covered entity or business associate. In a prompt, that can include names, dates, record numbers, images, free-text clinical notes, and combinations that re-identify a patient even without an explicit identifier. De-identified data is outside HIPAA only if it meets the Safe Harbor or Expert Determination standard. Do not assume that removing a name is enough. Classify the prompt before it leaves your environment, and apply minimum necessary at the application layer, not in a policy document.

Can we de-identify data before sending it?

Yes, and it is often the right first step. Safe Harbor requires removal of 18 categories of identifiers, including dates, geographic subdivisions smaller than a state, device identifiers and biometric identifiers. Expert Determination requires a qualified person to determine that the risk of re-identification is very small and to document the methods and result. Neither is a one-time check: re-identification risk changes with the dataset and the model's ability to combine fields. If you use de-identified data, keep the mapping and the determination, and verify that the vendor's retention does not recreate identifiability.

Who reviews model output before it reaches a patient?

A qualified clinician, under your clinical governance. The model can draft a summary, extract a finding or suggest a code, but it should not diagnose, prescribe or discharge without human review. Build the review into the workflow: show the source evidence, mark uncertainty, require an explicit accept or edit, and record the reviewer. For patient-facing content, add a second check for tone, readability and health literacy. Do not describe the product as a medical device or claim clinical accuracy unless you have the regulatory evidence and the clearance that claim requires.

What should a hospital's security review cover?

A hospital security review should cover the data path for PHI, the BAA, access controls, audit logging, encryption and key custody, retention and deletion, incident notification, subprocessors, and the model change policy. Add clinical safety: intended use, contraindications, evaluation on your patient population, and the escalation path when the model is wrong. Check whether the deployment is shared or private, and whether support staff can see prompts. Review the vendor's SOC 2 report and bridge letter, but do not treat them as a HIPAA certification. If the workflow touches substance-use-disorder records, separate consent rules may apply; verify with counsel.

Legal: confidentiality, privilege and supervision

The ethics analysis starts with the client and the matter, not with the model. A firm-wide policy is necessary but not sufficient.

Can a law firm send client documents to an AI API?

It can, but not by default. The firm must assess confidentiality under Model Rule 1.6, competence under Rule 1.1, supervision, communication and fees. ABA Formal Opinion 512 says lawyers must understand the tool's terms, who can access the data, and the risk of disclosure before inputting client information. If the tool learns from inputs, informed consent may be required. A private deployment with no training on your data and strict access controls reduces the risk, but it does not eliminate the need for a firm-level policy and a client-specific assessment.

Does using an AI tool waive privilege?

It does not automatically waive privilege, but disclosure to a third party can create risk, and the analysis is fact-specific. Privilege protects confidential communications between a lawyer and client for legal advice; it is not a shield around every document that passes through a vendor. Using a tool under a confidentiality agreement, with access limited to the firm and no independent use by the vendor, is different from posting a document into a public consumer chatbot. Counsel should assess the specific workflow, the vendor terms and the jurisdiction. Do not rely on a vendor's “privileged” label. This is not legal advice.

What does ABA Formal Opinion 512 require?

Formal Opinion 512, issued in July 2024, applies the Model Rules to generative AI. It covers competence and verification of output, confidentiality and informed consent, communication with clients, supervision of lawyers and non-lawyers, candor to tribunals, and reasonable fees. It does not create a new rule or a safe harbor. The practical takeaways are: read the tool's terms, know who can access the data, verify citations and analysis, disclose when required, and do not bill for time you did not spend. State and local bar guidance may add requirements, so check your jurisdiction.

How do we keep matter data separate?

Separate at the application layer: one workspace or key per matter or client, role-based access tied to the matter team, and no shared prompt history across matters. If the vendor supports organizations, projects or keys, map them to your ethical walls. Log access and review it. Do not rely on the model to remember an instruction not to disclose another client's data; the model has no concept of your ethical obligations. Test with a user who should not see a matter, and confirm the retrieval layer cannot return it. Confidentiality is an architecture property, not a prompt.

Can we disable training on our data?

Ask for the commitment in the contract; do not infer it from the word “private”. The DPA or security addendum should state whether prompts, completions, embeddings and feedback are used to train or improve models, whether human reviewers can see them, and how long they are retained. If the answer is “no training”, confirm that it survives changes to the service and applies to subprocessors. For law firms, also ask about telemetry, support access and evaluation datasets. A vendor's public FAQ is not a substitute for the signed term.

What should be in the engagement letter?

If the firm uses AI in the matter, the engagement letter can describe the tools, the data handled, the supervision model and the billing approach. It should not be a generic boilerplate consent. For higher-risk matters, seek informed consent for the specific disclosure and explain the benefit and the risk. Address who reviews output, how client instructions are recorded, and whether the client prohibits AI use. Check the rules in your jurisdiction and your professional-indemnity insurer's conditions. This is a counsel drafting task, not a marketing template. This page is not legal advice.

Shared workflows and evaluation

Healthcare and legal teams use different vocabulary for the same controls: evidence, review, retention and a named accountable person.

How should we log and retain prompts and outputs?

Decide by data class. For PHI or privileged material, prefer metadata-only logging by default, with content retained only when necessary for review and then deleted on a schedule. If you retain content, encrypt it, restrict access, log every read, and support legal hold. Keep the reviewer's decision and the source evidence, not just the model output. Confirm the vendor's retention separately: deleting your copy does not delete the vendor's logs or backups. Document the retention schedule and test a deletion request end to end.

What is the role of a human reviewer?

The reviewer is the accountable decision-maker, not a click-through. That means enough context to judge the output: the source documents, the prompt, the model and version, the retrieval results and any flagged uncertainty. The reviewer should be able to accept, edit or reject, and the system should record why. For clinical content, the reviewer must be qualified for the task; for legal content, a lawyer must supervise. “Human in the loop” is meaningless if the reviewer sees only a polished paragraph with no evidence and no time to check it.

How do we evaluate accuracy in a clinical or legal workflow?

Build a task-specific set with expected outputs or rubrics, reviewed by clinicians or lawyers. Include easy, ambiguous and adversarial cases, plus cases from your own population or practice. Measure correctness, omission, hallucination, citation accuracy, refusal and harm severity, not just fluency. Run the set before deployment and after every model or prompt change. Keep the results and the reviewer names. A public benchmark can help you shortlist, but it cannot tell you whether the system is safe for your patients or your matters.

What should we pilot first?

Choose a read-only, reversible workflow with a clear reviewer and a measurable outcome: discharge-summary drafting, referral triage support, document summarization, or contract review with a lawyer checking every clause. Avoid diagnosis, treatment recommendations, patient-facing advice, or autonomous filing in the first pilot. Define the data, the inclusion and exclusion criteria, the success metric, the harm metric and the stop rule. Set a deletion date. The pilot should produce evidence for governance, not a demo for the board.

What to verify in a healthcare or legal contract

These clauses decide whether the workflow can be defended to a regulator, a client or a court.

  • BAA scope. Covered services, permitted uses, safeguards, breach notification, subcontractor flow-down and termination.
  • Confidentiality. No-training commitment, access limits, support access, telemetry and evaluation data.
  • Data path. Regions for prompts, logs, backups and support access, with the transfer mechanism for each.
  • Retention. Separate periods for prompts, outputs, logs, embeddings and backups, plus legal hold and deletion proof.
  • Subprocessors. Current list, notice, objection rights and equivalent obligations for any model provider.
  • Incident response. Notification clock, severity, updates, root-cause report and cooperation with regulators and clients.
  • Audit rights. Report review, bridge letter, questionnaire and an agreed path to a remote or on-site assessment.
  • Model change. Engine and endpoint change notice, evaluation rights and deprecation windows.
  • Intended use. Clear limits on clinical or legal reliance, and your obligations for human review and escalation.
  • Remedies. Liability, indemnity, insurance, termination for cause and transition assistance.

What we did not verify

We did not review a signed BAA, a DPA, a clinical validation or a legal opinion for this page. We did not test the model on PHI or privileged material, and we did not assess Arabic OCR or generation quality for patient records. Plugsky makes no HIPAA certification claim, and this page makes no claim that any workflow is clinically or legally safe. The ABA opinion is guidance, not a safe harbor, and state rules may differ. Every regulatory statement here needs review by counsel for your entity, license and jurisdiction. Nothing on this page is legal or medical advice.

Frequently asked questions

Does Plugsky have a HIPAA certification?

No. HHS does not issue HIPAA certifications, and Plugsky makes no such claim. The relevant evidence is a signed BAA and controls that meet the Security Rule.

Will Plugsky sign a BAA for any workload?

A BAA is available by contract, subject to review. The scope, covered services and permitted uses must be agreed before PHI is processed.

Can we use the API without a BAA?

Do not send PHI without a signed BAA and a documented minimum-necessary workflow. De-identified data may fall outside HIPAA if it meets the standard.

Who is the covered entity?

Your hospital, clinic, insurer or clearinghouse is usually the covered entity. Plugsky's role depends on the contract and the service, so confirm it in the BAA.

Is de-identified data always safe?

No. Safe Harbor and Expert Determination have specific requirements, and re-identification risk can change with the dataset and the model. Document the method.

Can a lawyer rely on the output?

No. A lawyer must verify citations, analysis and facts, and remains responsible for the work. ABA Formal Opinion 512 addresses competence and supervision.

Does AI use waive privilege?

Not automatically, but disclosure to a third party can create risk. The answer depends on the workflow, the vendor terms and the jurisdiction; ask counsel.

Can we prevent training on client data?

Ask for a contractual commitment covering prompts, completions, embeddings and feedback. Do not infer it from a “private” or “enterprise” label.

How do we isolate law-firm matters?

Use separate workspaces or keys per matter, role-based access tied to the matter team, and tested ethical walls. Do not rely on a prompt instruction.

What is the breach notification deadline?

HIPAA generally requires individual notice without unreasonable delay and no later than 60 days after discovery. The BAA should define the vendor's notice to you.

Is the output a medical device?

Not by default. If you intend the output for diagnosis or treatment, check the regulatory classification and clearance requirements before deployment.

Does this page constitute legal advice?

No. It is a procurement and workflow guide. Verify HIPAA, privilege, consent and state bar requirements with qualified counsel.

Review the workflow before the first record

Start with the BAA, the data path and the reviewer. Then test the model on de-identified cases.

See BAA terms → Security review

Sources

Primary documents we checked. They describe the regulatory floor, not your specific obligations.

Last checked 20 November 2026. This is a 2027 planning guide, not legal or medical advice.