AI email replies for lawyers: ABA Formal Opinion 512 and MailTwin
MailTwin is a Mac app that writes replies in Apple Mail from a few keywords. Lawyers ask the same question before trying any tool like it: what happens to the client’s email? This page sets out what the ABA’s ethics guidance on generative AI asks of you, and which MailTwin settings bear on it. It describes the software. It is not legal advice, and it does not say that any configuration satisfies your obligations.
For how MailTwin fits a solo practice day to day, see Solo lawyers and small firms.
What Formal Opinion 512 says
On 29 July 2024 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, “Generative Artificial Intelligence Tools”. On confidentiality under Model Rule 1.6, the points that matter for email are:
- Consent before input. Because many self-learning AI tools are designed so their output could disclose information relating to a representation, the opinion says the client’s informed consent is required before that information goes into such a tool.
- Boilerplate is not consent. In the opinion’s words, “merely adding general, boiler-plate provisions to engagement letters” is not sufficient.
- Read the terms. As a baseline, lawyers should read and understand a tool’s terms of use and privacy policy to learn who has access to what they input, or consult someone who has.
- No client information, no consent question. Using a tool without inputting information relating to the representation, for example for idea generation, does not need the client’s consent.
The committee notes its conclusions reflect the technology as of publication and may change.
Where the email goes in MailTwin
MailTwin has no server of its own in the AI path. When you ask for a reply, the request goes from your Mac to the provider you chose in Settings → AI → Providers, under that provider’s terms. Which provider you pick decides who sees the text:
- On your Mac. Apple Intelligence (on supported Macs running macOS 26 or later), Ollama, or a local server such as LM Studio or llama.cpp. The model runs on the same Mac, so the email text does not go to an AI company. The local providers accept only a loopback address and refuse any other host before a request is built.
- A cloud provider with your own key. Claude, OpenAI, Gemini, Grok, DeepSeek or Groq. The text needed for the action goes to that provider. This is the case the opinion’s consent and terms-of-use points are about, and reading that provider’s terms is your step, not MailTwin’s.
Name masking
Settings → Privacy → Mask names before cloud requests
(off by default) replaces names, email addresses and phone numbers with
placeholders such as [PERSON_1] on your Mac before a cloud request,
then puts the real values back in the reply. It reduces what the provider sees.
It can miss names, especially unusual ones or names in non-Latin scripts, and the
rest of the message still goes, so it does not make a request anonymous and it is
not a substitute for client consent. On-device providers don’t need it.
Checking what was sent
Settings → Privacy keeps a request log of what each provider received. Use it to confirm which provider answered and, with masking on, what the placeholders replaced, before you rely on a setting for client work.
Style learning
MailTwin can learn your writing style from your sent mail, but only after you start indexing yourself. The index is kept on your Mac. With a cloud provider, a reply in your style includes a few short examples of your sent mail in the request to that provider, so the same consent considerations apply to those examples. With an on-device provider they stay on the Mac.
Frequently asked
Does Formal Opinion 512 allow AI on client email?
It does not forbid it. It requires informed client consent before information relating to a representation goes into a self-learning tool, says boilerplate is not enough, and asks you to read the tool’s terms.
Can MailTwin write replies without sending client email to an AI company?
Yes, with Apple Intelligence, Ollama or a local server. The model runs on your Mac.
Does name masking make cloud AI safe for client confidences?
No. It reduces what is sent but can miss names and leaves the rest of the message intact. It does not replace consent.
Trial
The 30-day trial includes every provider, so you can set up an on-device configuration and check the request log before deciding. Get MailTwin.