Yes, a law firm can use an AI email assistant on client mail without breaching confidentiality. The ABA’s Formal Opinion 512 (July 29, 2024) doesn’t forbid it. It asks you to know three things first: what the tool keeps, who can reach what it keeps, and whether a client has to agree. A private email assistant for law firms is one where you can answer all three from the vendor’s own documents.
- The reading is the product. Any assistant that sorts, summarizes or drafts has to read the mail. The useful questions start after that: what’s kept, for how long, and by whom.
- Training is the consent trigger. Opinion 512 requires informed consent before client information goes into a “self-learning” tool whose output could reveal it. Whether your tool is one is a line in its terms.
- The risk inside the firm counts too. The opinion names colleagues behind an ethical wall, and lawyers who might reuse one client’s facts for another.
- Boilerplate doesn’t carry consent. Where consent is needed, a general clause in the engagement letter isn’t enough.
When a private email assistant for law firms is safe for client mail
The duty underneath it all is Rule 1.6(c). The ABA Model Rule reads:
A lawyer shall make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client.
That’s a standard of reasonable effort, judged against the tool and the matter. Your state’s version is the one that binds you. North Carolina’s Rule 1.6, for example, carries its own 1.6(c). Check your own bar for an AI opinion as well. Opinion 512 itself cites state opinions from Florida, Pennsylvania, West Virginia and the District of Columbia.
Opinion 512 applies that duty to generative AI. It’s the ABA’s first formal opinion on the subject. It works through the existing rules on competence, confidentiality, communication, supervision and fees. On competence it sets a modest bar. Lawyers “need not become GAI experts,” but they must have “a reasonable understanding of the capabilities and limitations” of the tools they use.
Software already reads client mail every time it filters spam or indexes a search. What matters is what this particular tool does with what it reads.
What Opinion 512 asks before client mail goes in
The opinion asks for a risk check in two directions. Here’s the passage:
Before lawyers input information relating to the representation of a client into a GAI tool, they must evaluate the risks that the information will be disclosed to or accessed by others outside the firm. Lawyers must also evaluate the risk that the information will be disclosed to or accessed by others inside the firm who will not adequately protect the information from improper disclosure or use
Outside the firm means the vendor, its AI providers, and other customers of the same model. Inside the firm means your own people.
Then it names the case that needs consent:
Accordingly, because many of today’s self-learning GAI tools are designed so that their output could lead directly or indirectly to the disclosure of information relating to the representation of a client, a client’s informed consent is required prior to inputting information relating to the representation into such a GAI tool.
That turns consent on one fact about the tool. Does what you put in come back out somewhere else? A tool that trains a shared model on your mail can surface one client’s facts in another user’s answer. A tool contractually barred from that is a different risk, and the opinion’s own footnote 34 says its conclusion rests on “the risks and capabilities of GAI tools as of the publication of this opinion.”
The opinion also tells you where to find the answer:
As a baseline, all lawyers should read and understand the Terms of Use, privacy policy, and related contractual terms and policies of any GAI tool they use to learn who has access to the information that the lawyer inputs into the tool or consult with a colleague or external expert who has read and analyzed those terms and policies.
What an email assistant reads
An assistant connected to your mailbox reads the mail in it. That’s how it ranks a pile, pulls a deadline out of the fourth paragraph and drafts a reply. Expect every message the connection reaches to be read, including attachments.
Two kinds of mail stay out of reach:
- Mail sealed in a portal. It arrives as a notice with a link. The assistant sees the notice, and the contents stay in the portal.
- Mail locked end to end. Only the recipient can open it, so no assistant can read, summarize or draft from it.
Which encryption leaves what in your mailbox is covered in encrypted email for attorneys. For most firms, the honest picture is this: the assistant reads nearly everything, and the few messages that warrant sealing are the ones it can’t.
What it keeps, and where
Reading is brief. Keeping is what lasts, and it’s what Opinion 512’s vendor questions are really about. An AI email assistant usually holds four kinds of thing.
| What’s kept | What it is | What to ask |
|---|---|---|
| The correspondence | Messages, attachments, contacts, calendar | Is it copied, or read in place? How long is it kept? |
| What the tool made from it | Summaries, extracted tasks, ranking signals, a search index | Does deleting a message delete these too? |
| Excerpts sent to AI providers | Pieces of mail sent to a model to produce the above | Which providers? Can they train on it? What do they log? |
| Logs | Security, abuse and system records | How long, and what’s in them? |
The second row is the one firms forget. A search index built from your mail is a copy of its meaning. If the vendor’s deletion clause only covers “messages,” the index may outlive them.
The third row is the one a no-training promise usually leaves out. A vendor can bar training and still let its AI provider keep prompts for abuse monitoring under the provider’s own terms. Ask which applies.
Opinion 512 lists the vendor checks it borrows from earlier cloud-computing opinions. In plain terms, it asks you to:
- make sure confidentiality is enforceable, with notice of a breach or a subpoena for client data
- look at the tool’s security, its policies and the limits on its liability
- find out whether it keeps your data after you stop using it, or claims rights in it
- understand that its servers can fail and are a target for attack
Where your mail goes when software reads it, with the vendor questions in full, is in how AI tools handle your data and what to check before connecting your inbox.
What Microsoft and Google say about their own assistants
Most firms run on Microsoft 365 or Google Workspace, so their built-in assistants are the first comparison. Both publish a no-training position and a permissions position. Checked October 4, 2026:
- Microsoft Copilot. Microsoft’s data, privacy and security page says “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs.” It also says Copilot “only surfaces organizational data to which individual users have at least view permissions.”
- Gemini in Google Workspace. Google’s privacy hub says “Workspace does not use customer data for training models without customer’s prior permission or instruction.” It adds that the AI follows existing Workspace permissions.
Both are vendor statements, and both can change. Read them the way Opinion 512 asks you to read any vendor’s terms, and save a dated copy.
Who inside the firm can see it
This is the half of Opinion 512 that firms skip. The opinion describes a self-learning tool revealing information to people in the firm “who either are prohibited from access to said information because of an ethical wall or” who “could inadvertently use the information from one client to help another client.”
Two things set your exposure here, and neither is the vendor’s.
Who has a seat on which mailbox. A seat on an assistant sees the mailbox it’s connected to, including every summary and search result built from it. A shared intake address or a partner’s delegated inbox carries every client in it. If a lateral hire is screened from a matter, they shouldn’t have a seat on any mailbox that holds that matter.
What a draft can draw on. An assistant that drafts from your own correspondence can, in principle, carry a fact from one client’s thread into another client’s reply. The name is right and the tone is right, and one detail came from somewhere else. Review before sending is the control that catches it.
Opinion 512 puts the policy on the firm’s managers:
Managerial lawyers must establish clear policies regarding the law firm’s permissible use of GAI, and supervisory lawyers must make reasonable efforts to ensure that the firm’s lawyers and nonlawyers comply with their professional obligations when using GAI tools.
For a small firm that policy can be one page. Decide who connects a mailbox and which mailboxes. Decide which actions need a lawyer’s review. Write down who may accept a vendor’s terms, since a staff member clicking “agree” on an AI feature is the likeliest way client mail ends up somewhere you didn’t choose.
What the client is owed
There are two separate duties here, and the opinion keeps them apart.
Consent under Rule 1.6. If the tool is self-learning in the opinion’s sense, you need informed consent first. The opinion says what makes it informed. The client must have:
- your best judgment about why the tool is being used
- specific information about the risk, including the kinds of client information disclosed
- the ways others might use that information against the client
- a clear explanation of the tool’s benefits to the representation
And it rules out the shortcut:
To obtain informed consent when using a GAI tool, merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient.
Communication under Rule 1.4. Where consent isn’t required, you still have to consider whether to tell the client. The opinion says the facts of each case decide. Some cases are settled. You must disclose “if asked by a client how they conducted their work, or whether GAI technologies were employed,” or if the engagement agreement or the client’s outside counsel guidelines require it.
Corporate clients often do require it. Read the guidelines before you connect the mailbox their matters run through.
Even where neither rule requires it, the opinion says you may tell clients how you use AI, and that “the engagement agreement is a logical place to make such disclosures.” Plain disclosure is usually the calmer path. A client who learns it later tends to ask why nobody said so.
Mail that needs more than the standard terms
Some client mail carries obligations beyond Rule 1.6. The common one for law firms is health information. A firm handling protected health information for a hospital or health plan can be a HIPAA business associate. That needs a business associate agreement with anyone who stores that information for you, and that includes an email assistant. The HIPAA side is explained in are encrypted emails HIPAA compliant?
Read each vendor’s terms for a regulated-data clause. Many exclude HIPAA data from their standard service, and a firm that connects a health-law mailbox anyway has a gap no setting closes.
How Point handles client mail
Point is an AI email client that sits on the Gmail or Microsoft 365 account your firm already has. Point reads the mail, sorts it, briefs you each morning and drafts replies. Here’s what Point keeps, with the document behind each line.
- Training. Point’s terms, version 1.0 effective July 7, 2026, say at section 6.8 that Point won’t use your content to train generalized third-party or foundation models unless you expressly authorize it. Point does use your content to run and improve the service for your firm. That covers ranking that learns which senders matter to you, and a search index of your own mail.
- AI providers. Anthropic and OpenAI process excerpts to produce summaries and drafts, and Google’s embedding API builds the search index. All three are named on the dated subprocessor list, used under business terms with no-training controls where those exist. The addendum’s section 7.2 says providers may still process prompts and logs for safety, abuse prevention and reliability under their own terms.
- People at Point. The privacy policy at section 6.8 limits human access to support you asked for, security and abuse investigation, legal compliance, troubleshooting, service operation, or your consent.
- Retention. Mail is kept while your account is active. Summaries and index entries generally live as long as the mail they came from. On termination, the addendum commits to deleting or returning your data within 60 days at your choice, where that’s specified in your agreement. Backups turn over on their own cycle.
- Sealed mail. A message you lock end to end in Point can be opened by its recipient and nobody else, Point included.
Two lines in the terms matter most for a law firm. Section 5.1 leaves attorney-client privilege and client confidentiality with you. The same section excludes HIPAA health information, and data under professional-secrecy rules Point hasn’t expressly accepted, unless a separate written supplement covers it. Ask Point in writing whether your practice’s mail needs one before you connect a mailbox.
Inside one mailbox, Point draws no wall between one client and another. Point does keep separate businesses you operate apart from each other. That line is between organizations, and it isn’t an ethical wall. The difference is set out in keeping one business isolated from another.
How far Point acts on its own is set per kind of action. Every kind starts on review, so a draft waits for you. You can move a setting up later, and at the top of the dial Point acts without asking. A sent email can’t be pulled back off someone else’s server. For client mail, review is the setting to keep. Everything Point did is in the activity log, with a time on each line.
The rest of what Point does is on the benefits page.
Common questions
Does ABA Opinion 512 require client consent to use an AI email assistant?
Only in some cases. Opinion 512 requires informed consent before client information goes into a “self-learning” tool whose output could disclose it to others. Whether a tool is self-learning in that sense depends on its terms, especially whether it trains shared models on your input. Where consent isn’t required, Rule 1.4 may still require you to tell the client, and you must disclose if the client asks.
Is a clause in our engagement letter enough?
Not for consent. Opinion 512 says general, boilerplate provisions authorizing AI use are “not sufficient” to obtain informed consent. An engagement letter is still a good place to describe how the firm uses AI and to record a client’s instructions. Where consent is needed, the client needs the specific risks, the kinds of information involved and the benefits.
Is it safe to use Microsoft Copilot on privileged email?
Microsoft says Copilot doesn’t train foundation models on prompts, responses or Microsoft Graph data, and that it only surfaces data a user can already view. Those are Microsoft’s own statements, so check them against your agreement. The inside-the-firm risk depends on your permissions. If a lawyer can open a mailbox, Copilot can summarize it for them.
Is emailing a lawyer confidential?
The lawyer’s duty of confidentiality covers it. Privilege is narrower, and a client can lose it by writing from a work account, sharing a password or forwarding the advice. The client’s side is set out in is emailing a lawyer confidential?
Does an AI assistant read encrypted email?
It reads what arrives readable. Ordinary mail is encrypted only in transit, so the assistant reads it like any message. A portal message shows only its notice, and a message locked end to end can’t be read by anyone but the recipient. The full picture is in encrypted email for attorneys.
What should a small firm do first?
Read the vendor’s terms for training, retention and human access, and save a dated copy. Decide who gets a seat on which mailbox. Keep drafts on review. Check whether any client’s outside counsel guidelines speak to AI. Then decide, client by client, whether consent or a disclosure is owed.