On this page
What can AI actually do for a law or accounting firm?Where does AI get it wrong, and how do you contain it?How is client information handled at each level?Who can get into my systems, and what happens if we part ways?What about insurance and continuity?Which professional-responsibility questions are yours to answer?What does ABA Formal Opinion 512 say?What should you ask any AI vendor?What can Primus answer today?How does a Blueprint Session work for a firm?How broad is Primus's experience with professional firms?Questions people askI am Glenn Murano. I am not a lawyer, an accountant or a compliance expert, and Primus AI Consulting is not a compliance firm. I help owners understand AI, decide whether it fits, and put it to work. For a firm, the hard part is not the technology. It is confidentiality, and being wrong in front of a client or a court.
This page says what AI is good for in a firm, where it fails, what each of the three levels does with your data, which professional-responsibility questions you need to answer with your own counsel, and what to ask any AI vendor, including me. I will also be straight about breadth: ClosingCT, a real-estate closing practice, is the professional-services client on my list.
What can AI actually do for a law or accounting firm?
Five jobs come up most. In each one, the AI prepares and a person decides. See AI agent use cases by industry for examples.
| Job | What the AI does | Who decides |
|---|---|---|
| Intake follow-up | Drafts a reply to a new inquiry and gathers the basic details a person will need. | A staff member reviews it before it goes out. |
| Scheduling | Proposes times and drafts the confirmation. | Staff approve the booking. |
| First drafts | Turns your notes into a first draft of a letter, a summary or a checklist. | The attorney or accountant reads every word and owns the result. |
| Billing reminders | Drafts a polite note for an unpaid invoice. | You approve it before it is sent. |
| Internal knowledge search | Finds and quotes from your own documents and procedures when staff ask a question. | Staff check the original document it points to. |
The workflow agent level (Level 2) is built on one rule: the agent drafts, and you approve.
Where does AI get it wrong, and how do you contain it?
Three failures matter most to a professional firm. None of them can be designed away. The design goal is that a mistake lands on a person's desk before it lands in front of a client or a court.
| Failure | What it looks like | What contains it |
|---|---|---|
| Made-up citations | A filing cites authority that does not exist. In Mata v. Avianca (S.D.N.Y., June 22, 2023), the court's opinion records that the respondents acknowledged six cited decisions were generated by ChatGPT and do not exist. The court imposed a $5,000 penalty. | A person checks every citation against the original source before anything is filed or sent. Nothing leaves the firm without approval. |
| Wrong summaries | A summary reads smoothly but changes or leaves out a term, a date or a dollar amount. | Treat the AI's summary as a draft. A person compares it to the source document before relying on it. |
| Prompt injection | Hostile text hidden in an email or document tells the AI to do something its owner never asked for. OWASP lists prompt injection as LLM01 in its risk list for large language model applications. | OWASP's guidance includes requiring human approval for high-risk actions and limiting what the model can access. A Level 2 workflow agent asks before it acts, and the Blueprint marks the work AI should not touch. |
Ready to put real numbers on it?
A Blueprint Session gives you the written plan and the fixed price for your business. The session itself is free.
How is client information handled at each level?
This is what I can state about each level, and nothing beyond it.
For a firm I recommend that approval stays on for everything the agent drafts. My Terms let you change that in writing, but I would not.
| Level | Price | What happens to the text you give it |
|---|---|---|
| Level 1: Assistant | $3,000 to $6,000, hardware included | Runs through a commercial AI provider you choose and pay directly. The provider processes the text you send under its own terms. |
| Level 2: Workflow agent | $6,000 to $15,000, hardware included | Same arrangement: a commercial AI provider you choose and pay directly, processing what you send under its own terms. |
| Level 3: Private on-site agent | $15,000 to $30,000, computer extra | Runs on a computer you own, on your premises. A local model may be added so sensitive text is not sent out. What is sent out is decided in the Blueprint. |
Before you put client information into any provider-based setup, read the provider's terms and check its retention and training settings. Whether a given use is acceptable is for your firm to decide with its own counsel. There is more on Level 3 on the private on-site AI page.
Who can get into my systems, and what happens if we part ways?
You hold the logins. I ask for access only to the systems a build needs, and you can remove that access at any time.
- The Care plan includes remote monitoring and daily backups of the agent.
- You keep the agent, the data in it and the computer it runs on. There is no license fee and no lock-in.
- If you leave the Care plan, what ends is support: monitoring, backups, repairs, updates and coaching. Your setup keeps running as you own it.
- Under my Terms, my total liability for any claim is limited to the fees you paid me in the 12 months before the claim. For a client who has had only the free Blueprint Session, no fees have been paid, so that limit is $0. Ask your counsel whether that works for your firm.
The full wording is in the Terms. Ask me anything it leaves unclear before you book.
What about insurance and continuity?
I do not publish insurance details on this site. If your firm needs to see them, ask me in writing before you book and I will answer in writing.
Because you own the agent, the data and the computer, you are not dependent on me to reach your own information. That is the practical answer to a "what if you are unavailable" question, and the Terms describe the rest.
Which professional-responsibility questions are yours to answer?
I am not a lawyer, so I will not tell you what is permitted. I can name the rules lawyers are already weighing and point you to the text, so you and your own counsel can read it. These are the ABA Model Rules; Connecticut adopts its own versions in the Practice Book. If you are an accountant, closing agent or consultant, your profession has its own standards. I have not reviewed those for this page, so ask your professional body.
- Competence (Rule 1.1). The comment says a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology. Check the Connecticut version of this rule in the Practice Book.
- Communication (Rule 1.4). A lawyer must reasonably consult with the client about the means by which the client's objectives are to be accomplished. The ABA singled out Rule 1.4(a)(2) as particularly relevant to generative AI.
- Fees (Rule 1.5). Fees and expenses must be reasonable. As the ABA describes it, a lawyer may charge for the time spent entering information into an AI tool and reviewing the result, but in most circumstances cannot charge a client for learning how to use the tool.
- Confidentiality (Rule 1.6). A lawyer may not reveal information relating to the representation of a client without informed consent or another listed exception, and must make reasonable efforts to prevent inadvertent or unauthorized disclosure of, or access to, that information. This is Model Rule 1.6(c) and Connecticut Rule 1.6(e).
- Candor to the tribunal (Rule 3.3). A lawyer shall not knowingly make a false statement of fact or law to a tribunal.
- Supervision (Rules 5.1 and 5.3). Partners and managers must make reasonable efforts to ensure the firm has measures giving reasonable assurance that lawyers and nonlawyers act compatibly with the lawyer's obligations. Connecticut's Rule 5.3 and its commentary also address services a lawyer uses outside the firm, so read the current text with your counsel.
Rules and opinions summarised here were last checked in November 2026. Check the current text before you rely on them.
What does ABA Formal Opinion 512 say?
ABA Formal Opinion 512, titled Generative Artificial Intelligence Tools, was released on July 29, 2024 by the ABA Standing Committee on Ethics and Professional Responsibility. The ABA announced it as the committee's first formal opinion on generative AI and said the model rules on competency, informed consent, confidentiality and fees principally apply. The opinion describes itself as general guidance and says updated guidance will likely follow from the ABA committee and from state and local bar ethics committees.
I could open the ABA's announcement and the rule texts. I could not open the 15-page opinion itself, so read the full opinion (PDF) with your counsel before relying on my summary.
On Connecticut: I searched the full text of the 2026 Connecticut Practice Book for "artificial intelligence" and "generative" and found no rule that uses either term. Guidance can change, so check with the Connecticut Judicial Branch and the Connecticut Bar Association for what is current.
Your own counsel decides what is acceptable for your firm. Primus is not a law firm, an accounting firm or a compliance firm.
What should you ask any AI vendor?
Use this checklist with me and with anyone else who sells you AI. A vendor who cannot answer these plainly is telling you something.
- Where does the text I send go, and who processes it?
- What are the AI provider's retention and training settings, and who sets them?
- Will you sign an NDA before we talk about specifics?
- What does the system do on its own, and what waits for a person to approve?
- What happens when the AI is wrong, and how will I find out?
- Who can access my systems, and how do I revoke that access?
- Do I own what you build, and can I take it with me if we part ways?
- Are you a compliance firm, and will you promise that I am compliant? (An honest vendor says no.)
- What is the full price, including the AI provider's bill?
What can Primus answer today?
Here are my answers, where I can give them from facts rather than hope.
- Levels 1 and 2 use a commercial AI provider you choose and pay directly, at cost. The text you send is processed under that provider's terms.
- Level 3 runs on a computer you own on your premises, and a local model may be added so sensitive text is not sent out. What is sent out is decided in the Blueprint.
- I sign an NDA on request before the Blueprint Session.
- A Level 2 workflow agent drafts and you approve. The Blueprint also marks the work AI should not touch.
- You keep the agent. There is no license fee and no lock-in.
- I am not a compliance firm, and I certify nothing.
- Builds are priced on the services and prices page, and the AI provider bills you directly.
- Please do not put client names or privileged detail in your first note or the booking form. Describe the type of work instead.
How does a Blueprint Session work for a firm?
The Blueprint Session is a 60-minute working session with me. It is free. Within five business days you receive a written plan that says what to automate, which level fits, and a fixed price for the recommended work. The plan is yours to keep whether or not you build with me, and sometimes the honest answer in it is that AI is not the right fix.
If you want an NDA, ask and I will sign one before the session. After you pick a time, you get a confirmation by email. If you build, a build typically takes 2 to 8 weeks, and the how it works page shows each step. To begin, book a Blueprint Session, or ask me first if you are not sure.
How broad is Primus's experience with professional firms?
Not broad, and I would rather you hear it from me. ClosingCT, a real-estate closing practice, is the professional-services business I have worked with. The others I have worked with are other kinds of small business. I have 32 AI agents live in Connecticut and Colorado, and that is a count of agents, not a measure of results.
If you want to understand AI before you decide anything, start with how an AI agent works, then the readiness checklist. If you want an AI-use policy for your staff, how to train your team to use AI is a place to begin.
Questions people ask
Can a law firm use AI without breaking confidentiality rules?
That depends on the tool, its settings, your firm's policies and your clients' consent, and your own counsel decides. The ABA's announcement of Formal Opinion 512 says a lawyer using generative AI must be cognizant of the duty to keep client information confidential unless the client gives informed consent. I can describe exactly what a system does with the text you give it. I cannot tell you that a use is permitted.
Do we have to tell clients we use AI?
I cannot answer that for you. Rule 1.4(a)(2) requires reasonably consulting with the client about the means by which the client's objectives are to be accomplished, and the ABA singled it out as relevant to generative AI. Whether and how to disclose, including in your engagement letter, is a question for your firm and its counsel.
Will AI make up cases or facts?
Yes, it can. In Mata v. Avianca, the court's opinion records that the respondents acknowledged six cited decisions were generated by ChatGPT and do not exist, and the court imposed a $5,000 penalty. Anything an AI drafts should be checked against the original source by a person before it is filed, sent or relied on.
Will you sign an NDA?
Yes, on request, before the Blueprint Session.
What does AI cost for a firm?
A Blueprint Session is free. Builds run $3,000 to $6,000 for an assistant, $6,000 to $15,000 for a workflow agent, and $15,000 to $30,000 for a private on-site agent, with the computer extra at Level 3. The AI provider bills you directly at cost. The Care plan is optional and starts at $500 a month. The written plan gives you a fixed price for the work it recommends.
Can you make us compliant or certify our AI use?
No. Primus AI Consulting is not a compliance firm, a law firm or an accounting firm, and I certify nothing. I can explain what a system does and where information goes. Whether that meets your obligations is for you and your own counsel.
Sources
- ABA, ABA issues first ethics guidance on a lawyer's use of AI tools (news release, July 29, 2024)
- ABA Model Rule 1.1: Competence, and its comment
- ABA Model Rule 1.4: Communications
- ABA Model Rule 1.6: Confidentiality of Information
- ABA Model Rule 3.3: Candor Toward the Tribunal
- ABA Model Rule 5.1: Responsibilities of a Partner or Supervisory Lawyer
- ABA Model Rule 5.3: Responsibilities Regarding Nonlawyer Assistance
- Connecticut Practice Book 2026, Rules of Professional Conduct, Connecticut Judicial Branch
- Mata v. Avianca, Inc., Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023)
- OWASP Gen AI Security Project, LLM01: Prompt Injection
Ready to decide?
The Blueprint Session is how you turn this into a plan and a price. The session itself is free. Or ask Glenn first.
