Florida Bar Ethics Opinion 24-1 requires a lawyer to investigate an AI provider before entrusting client information to it. These are the questions that follow from it, each anchored to the rule that requires it, with room to record what the vendor actually said.
← Back to overviewOpinion 24-1 permits Florida lawyers to use generative AI, subject to existing duties of confidentiality, competence, and supervision. On where the data sits, it is direct:
“Confidentiality concerns may be mitigated by use of an in-house generative AI rather than an outside generative AI where the data is hosted and stored by a third-party.”
Florida Bar Ethics Opinion 24-1
On consent, the opinion draws a line that has practical consequences for how a firm operates day to day. Where a tool discloses confidential information to a third party, it recommends obtaining the client’s informed consent. Where it does not:
“If the use of a generative AI program does not involve the disclosure of confidential information to a third-party, a lawyer is not required to obtain a client’s informed consent.”
Florida Bar Ethics Opinion 24-1
ABA Formal Opinion 512 takes a firmer national position, treating client consent as required rather than recommended, and stating that boilerplate consent language in an engagement letter is not sufficient for this purpose. Read together, the two opinions mean that a cloud-based tool can create a per-client, per-matter consent obligation that a self-hosted one does not.
Opinion 24-1 requires a lawyer to establish whether a provider retains what is submitted to it — both during the engagement and after it ends.
The opinion directs lawyers to determine whether the provider retains information submitted before and after the discontinuation of services. A retention answer that only covers the active subscription period is incomplete.
An assurance of deletion that cannot be evidenced is not something a firm can rely on when answering its own client.
Storage location determines which laws and which government access regimes apply to material the firm holds in confidence.
The opinion singles out self-learning systems as a specific confidentiality risk, because inputs can persist in the model itself rather than in a database a vendor can point to.
Ask separately about your own systems and about any underlying model provider. The two answers are often different, and only one of them is usually volunteered.
Opinion 24-1 states that this raises the possibility that a client's information may be stored within the program and revealed in response to future inquiries by third parties.
This is the question the previous two are really driving at. Ask it plainly and note whether the answer is contractual or merely reassuring.
Opinion 24-1 adapts the standard applied to cloud-computing providers: confidentiality obligations must be enforceable, and breach notification must exist.
A privacy policy is a statement of current practice. A contractual term is a promise. The opinion asks for enforceable obligations.
The firm's own notification duties to clients and regulators start when it learns of a breach. A vague vendor timeline becomes the firm's delay.
The opinion directs lawyers to investigate limitations on the provider's liability. A cap set at fees paid means the firm carries essentially all of the exposure.
Diligence on a vendor is incomplete if that vendor passes the data to someone else. Most AI products are assembled from several providers.
Model providers, hosting, logging, analytics, and support tooling may each touch client material. Ask for the list, not a summary.
A vendor's zero-retention claim is only as durable as the agreement behind it. Ask whether that agreement is in place today and what happens if it changes.
Without notice, the diligence the firm performed at purchase can quietly stop describing reality.
Conflicts screens and ethical walls are only meaningful if the tool enforces them at the point of search, not after results are returned.
If the system retrieves everything and then hides what a user is not entitled to, the material has already been read into the model's context.
Ask who can change a wall, whether the change is logged, and whether the firm can produce that log later.
Opinion 24-1 treats AI output like the work of a nonlawyer assistant: it must be reviewed, and the lawyer remains responsible for it.
Review is only practical if verification is cheap. If checking an answer means re-reading the file, the tool has moved work rather than saved it.
A citation the model composes can be wrong in exactly the way that is hardest to notice — plausible, formatted correctly, and pointing nowhere.
Supervision under 4-5.3 requires the firm to be able to demonstrate oversight after the fact, not only to intend it.
This question determines whether adopting the tool creates a recurring obligation on every matter, or none at all.
Opinion 24-1 recommends obtaining the client's informed consent where it does, and states that consent is not required where it does not. ABA Formal Opinion 512 adds that boilerplate consent in an engagement letter is not adequate — so a yes here means a real conversation with each affected client.
None of these is disqualifying on its own. Each is a point where the first answer is often narrower than it sounds, and a second question is usually worth asking.
LexDeploy builds private AI infrastructure that runs inside a firm’s own environment, so we have an obvious interest in the answers to section G. That is exactly why this checklist is written as questions rather than conclusions, and why we would rather you put the same seventeen to us as to anyone else. If our answers are not better on their face, we are not the right choice.
LexDeploy is not a law firm and does not provide legal advice. This document summarises published ethics opinions as a practical aid for vendor evaluation; it is not a substitute for reading them or for your own professional judgment. Read the sources: Florida Bar Ethics Opinion 24-1 and ABA Formal Opinion 512.
Section G is only worth as much as the person on the other side of it. One engineer specifies the system, holds administrative access, and answers when something breaks.
About LexDeploy →