ABA Formal Opinion 512, the Illinois AI policy, and how a Clio assistant's controls line up.

What the guidance says

Two sources matter for a firm in Illinois. The American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, "Generative Artificial Intelligence Tools", on July 29, 2024. It forbids nothing; it walks through the Model Rules and says what each one means when a tool drafts, summarizes or suggests: competence, confidentiality, communication with the client, meritorious claims and candor toward the tribunal, supervision of lawyers and non-lawyers, and fees.

The Illinois Supreme Court's Policy on Artificial Intelligence took effect on January 1, 2025. It says the use of AI by attorneys is authorized when it complies with legal and ethical standards, that disclosure of AI use should not be required in a pleading, that the Rules of Professional Conduct apply fully, that whoever signs the work is accountable for it and must review it thoroughly, and that confidential communications and personal information must not be compromised. The Attorney Registration and Disciplinary Commission's contribution is neither a rule nor an opinion: its website publishes a nonbinding companion to the Court's policy, "Implementing Generative AI for Illinois Attorneys", whose release the Illinois State Bar Association reported in October 2025, with a practice kit: a sample notice, a use policy, a client consent form and a terms-of-use checklist.

Below, each duty in the opinion's order, with what the assistant does. None of this is legal advice. How the rules apply to your firm is a judgment for you and your own counsel.

Competence and candor: the output is a proposal, not a record

Opinion 512 says a lawyer must understand what a tool can and cannot do, and may not rely on its output without an appropriate degree of independent review. The Illinois policy says the same in fewer words: you are accountable for the final work product.

The assistant is built so that review is not a habit but a mechanism. Every match it makes - this email belongs to that matter, this meeting has no time entry beside it - is a best guess from evidence it names (an address, a subject line, an attendee), phrased as "possible" or "worth a look", never as a fact. Nothing is written to Clio or sent from a mailbox on that guess. The system refuses to write anything until a proposal for exactly that write has been shown and confirmed, within a few minutes, once; change a detail and you get a fresh proposal.

Where the temptation to invent is greatest, it is designed not to. It never proposes a number of minutes; if you do not say how long, it asks. A draft email takes the facts from the file and leaves a bracketed blank where you have not given it a date, an amount or a position. It does the date arithmetic you ask for, in the open, and never computes a legal deadline from a rule. A conflict check never says "no conflict"; it says what it searched and what matched. The billing review guide and the email drafting guide draw each line exactly.

Confidentiality: where client information goes

The opinion's longest section. Before putting client information into a tool, a lawyer should assess the risk of disclosure, read the terms of use and data policies, and, for a tool that learns from what is typed into it, obtain the client's informed consent first. The Illinois policy adds that confidential communications and personal information must not be compromised.

The assistant runs on each person's own computer, through Claude Code, and signs in to Microsoft 365 and Clio with that person's own logins, so it can see and change only what they can. There is no MG Legal Ops server that receives a firm's email, documents, matters or time entries; the privacy policy lists what we do receive, and it is account details and a daily subscription check, never mail or matters. What reaches Anthropic is what any Claude Code conversation sends, under the firm's own Claude plan and its terms. The confidentiality page says what to check in that plan, and reading those terms is the step the opinion asks of you.

Inside the firm, the assistant keeps its own reach narrow. It ties an email to a matter by who is on it, a matter number in the subject, or a thread you filed before; it never reads the body to work out the matter. It reads only what the signed-in person's own logins allow. Anything written inside an email is treated as information, never as an instruction, and if a message tries to tell it to do something it tells you instead. The problem reports it files about itself carry no names, subjects, addresses or matter details.

Communication, fees and supervision

On communication, Opinion 512 sets no blanket duty to disclose, but says a client must be told when they ask, when their information goes into a tool that needs consent, and when the use bears on the fee; an engagement letter is the natural place. Whether your firm says so, and how, is your call, and the ARDC's kit has a sample notice and consent form to start from. One fact about the assistant helps that conversation: nothing it produces for a client names Clio, Claude or AI, so a bill or a filed letter reads as your own work, and the system refuses a time-entry description that says otherwise.

On fees, the opinion says an hourly bill covers actual time only, that learning a tool is not billable, and that a fee charged for little or no work is an unreasonable fee, however it was set. The assistant's part is that it never supplies a duration: it shows the evidence and suggests wording, and the minutes are yours, on a proposal you read first.

On supervision, the opinion asks managing lawyers for clear policies on permitted use and for training, and supervising lawyers for reasonable efforts to see the policy followed; a vendor is non-lawyer assistance to be vetted. What a firm can put in front of an insurer here: the guides under /docs are the same text the assistant ships with, each with a "what it can't do" section. Every write on every computer is recorded - which tool, which matter, when, whether it worked, never an email's contents - and "what did you file today?" reads it back. Anything the assistant created can be undone, and only that. Whether a computer points at a test Clio account or the firm's real one is set by whoever set it up, not by the assistant. A firm's administrator adds and removes people, each change proposed and applied only on a yes, after a sign-in that proves who is asking.

What it never does

It never writes to Clio, or sends an email, without a proposal and your yes. It never invents a duration, a date, an amount or a deadline. It never states a match or a gap as a fact. It never puts a tool's name into anything a client could see. It never follows an instruction found inside an email. It never sees more than the signed-in person's own logins allow. And it never removes anything it did not create: a matter or an entry made in Clio by other means is changed in Clio.

What it also never does is decide for you. The guidance puts the judgment on the lawyer, and the assistant leaves it there: it finds, proposes and records, and you say yes. Read the opinion, the Court's policy and the ARDC's guide, read your Claude plan's terms, and make the call with your own counsel.

MG Legal Ops runs on each person's own computer, attorney or paralegal, on their own Clio and Microsoft 365 logins, through Claude Code; the confidentiality page says what that means. The guides are the full description. Pricing, and the fourteen-day trial, no card to start.

Also read: Claude Code for law firms and Filing Outlook email to Clio matters, and why it slips. All articles.