Can a credit union use an AI chatbot for member service under GLBA?
Yes, but the chatbot vendor becomes a service provider under the Gramm-Leach-Bliley Act, and that triggers obligations the credit union owns — not the vendor.
If the chatbot touches member nonpublic personal information (names tied to account activity, balances, loan status, even the fact that someone is a member), the credit union must:
- Perform vendor due diligence before deployment. NCUA Part 748 and its Appendix A require a documented risk assessment of any third party handling member data. "We tried it and it worked" is not a risk assessment.
- Contractually bind the vendor to safeguards. The GLBA Safeguards Rule requires written contracts obligating service providers to maintain appropriate security. The contract should name encryption at rest and in transit, access controls, breach notification timelines, and what happens to data at termination.
- Confirm the data does not leave the arrangement. The question nobody asks and everybody should: is member data being used to train the underlying model? If the vendor cannot answer in writing, that is the answer.
- Keep the privacy notice accurate. Regulation P governs what the credit union tells members about information sharing. A new data-handling path can make an existing notice wrong.
The practical failure mode is not a breach. It is an examiner asking for the vendor risk assessment on a system that went live eight months ago, and nobody has one.
Does an AI receptionist handling patient calls need a HIPAA Business Associate Agreement?
Almost always yes, and the reason surprises people: appointment scheduling is protected health information.
PHI is not limited to diagnoses and test results. Under HIPAA, individually identifiable health information includes the fact that a named person is a patient of a particular provider, plus anything about the provision of care. A caller's name, phone number, and the fact that they booked with a dermatology practice is PHI. Add the reason for the visit and it is unambiguous.
An AI receptionist that hears, transcribes, stores, or routes that information is a business associate. The covered entity must have a signed BAA in place before the vendor touches PHI.
The conduit exception does not save you. That narrow exception covers entities that merely transmit data without accessing it — the postal service, an ISP. An AI system that transcribes a call, interprets intent, and writes to a calendar is processing the content, not passing a sealed envelope.
Three questions worth asking any AI vendor before signing:
- Will you sign a BAA? If the answer is no, the conversation is over for any PHI-touching use.
- Where are call recordings and transcripts stored, for how long, and who can access them?
- Is any of this data used to train or improve your models? A BAA and a training clause can contradict each other.
AI Genie signs a BAA for healthcare engagements. It is part of onboarding, not an upgrade.
What TCPA rules apply to automated appointment reminders and texts?
The short version: appointment reminders to existing patients usually qualify as informational rather than marketing, which lowers the consent bar. It does not remove it, and the exemption is narrower than most businesses assume.
Two consent standards. Informational messages (reminders, confirmations, care instructions) run on prior express consent, which a patient can give verbally or by providing a mobile number for care. Promotional messages require prior express written consent — a signed, specific agreement. The distinction is the message content, not your intent.
One promotional sentence breaks it. "Your appointment is Tuesday at 2:30" is informational. The same text with "book your next cleaning for 20% off" added is marketing, and it needed written consent you probably do not have.
The healthcare exemption is conditional. The FCC's conditions include that the message be treatment-related, be concise, carry an easy opt-out, stay within frequency limits, and that opt-outs be honored immediately. Miss one condition and the exemption drops for that message. Notably, the FCC extended the exemption to treatment calls but not billing-related calls — a balance-due text sits outside it even though your billing team treats it as routine.
Build on consent anyway. Every practitioner guide reaches the same conclusion: relying on the exemption is legally defensible and operationally fragile. A checkbox on the intake form — "I consent to receiving automated appointment reminders at the number provided" — costs nothing and moots the entire analysis.
Where the risk actually is. Statutory damages run $500 to $1,500 per message, and these cases arrive as class actions. One healthcare system settled for $12.8 million over reminders sent without consent and opt-outs not honored promptly. At a few thousand messages a month, the exposure math gets ugly fast.
This area moves. Confirm current FCC guidance before relying on any specific exemption.
Can a law firm use AI for client intake without violating confidentiality?
Yes, with conditions that most off-the-shelf tools do not meet.
ABA Formal Opinion 512 (2024) addressed generative AI directly. The duties it identified are the ones that govern an intake tool:
- Competence (Rule 1.1). A lawyer must understand the tool well enough to know its limits. "The vendor said it was secure" is not understanding.
- Confidentiality (Rule 1.6). Inputting client information into a tool that may use it for training, or that stores it where third parties can access it, can be an unconsented disclosure. Self-learning tools got specific attention for exactly this reason.
- Supervision (Rules 5.1 and 5.3). An AI handling intake is a nonlawyer assistant. The firm is responsible for its output.
Prospective clients are already covered. Rule 1.18 extends confidentiality duties to people who consult the firm and never retain it. The intake form is the moment duties attach, not the engagement letter. An AI that collects matter details from someone who never becomes a client is holding protected information.
The conflict-check problem. An intake AI that captures adverse-party names creates a record the firm must be able to search. If that data lives only in a vendor's system in an unexportable format, the firm has a conflicts problem it cannot see.
Florida practitioners: the Florida Bar has issued its own guidance on generative AI, and it is not identical to the ABA's. Check both.
Practical minimum before a firm deploys any AI intake tool: a written confirmation that inputs are not used for training, a data-retention and deletion policy, exportable records, and a documented supervision process for what the tool produces.
What should a small business document before deploying AI in a regulated workflow?
Six artifacts. None of them take long to produce, and all of them are painful to reconstruct after the fact.
| Document | What it establishes | Who asks for it |
|---|---|---|
| Vendor risk assessment | You evaluated the provider before go-live | Examiners, auditors, insurers |
| Data flow map | What data goes where, who can see it, how long it is kept | Everyone |
| Written vendor agreement | Security obligations, breach notification, termination handling | Regulators, counsel |
| BAA (healthcare) or equivalent | The legal instrument permitting the vendor to touch protected data | HHS, your own counsel |
| Human review policy | Which AI outputs a person checks before they reach a customer | Plaintiffs' attorneys, regulators |
| Incident and rollback plan | What happens when it is wrong, and how you turn it off | Your board |
The one most people skip is the human review policy, and it is the one that matters most when something goes wrong. "The AI handles it" is not a control. "The AI drafts, a person approves anything involving a dollar amount or a commitment" is.
The second most-skipped is what happens at termination. Most vendor agreements are silent on whether your data is deleted, returned, or retained indefinitely. Ask before signing, not after leaving.
Who is liable when an AI assistant gives a customer wrong information?
You are. Not the vendor, not the model provider.
The clearest decision on this is Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, February 2024). Air Canada's website chatbot told a passenger he could apply for a bereavement discount retroactively. That was wrong — it contradicted the airline's actual policy, published elsewhere on the same site. Air Canada argued the chatbot was a separate entity responsible for its own statements. The tribunal rejected that outright and held the airline liable for information on its own website, regardless of which component produced it.
The principle is not novel. It is ordinary agency law: a business is responsible for what its agents tell customers. The AI does not create a new category.
What follows from it:
- A disclaimer does not transfer the risk. "Responses may be inaccurate" buried in a footer did not help Air Canada and will not help you.
- Your vendor contract will not either. An indemnity clause may let you recover from the vendor afterward. It does not stop the customer, the regulator, or the plaintiff from coming to you first.
- Scope is the real control. The defensible design is an AI that answers from a bounded, reviewed knowledge base and escalates anything outside it — not one that improvises on pricing, eligibility, or policy.
- In regulated contexts the stakes compound. A wrong answer about a loan term, a coverage question, or a medical instruction is not a customer-service problem. It is a regulatory one.
This is why AI Genie builds from a client-approved knowledge base with explicit escalation paths, rather than pointing a general-purpose model at your customers and hoping.