The assistant explains the options. Deciding stays with your attorney.
The four mechanisms, and how we know they hold: outcome language banned at the prompt firewall, your disclaimer on every answer verbatim, written rules an attorney at your firm approves before launch, and every memo in your inbox.
The assistant stays out of the practice of law by explaining rather than deciding. It describes the options in front of a visitor, never says whether they qualify or how their matter will turn out, closes every answer with your disclaimer verbatim, and runs on written rules an attorney at your firm approved before launch. The phrases that predict a result are banned at the prompt firewall, not discouraged in a style guide.
What keeps it out of the practice of law
The law firms page lists the guardrails in a sentence each. These are the same guardrails, with what each one does in practice.
- It describes the options in front of a visitor and never says whether they qualify or how their matter will turn out. Telling a member of the public either of those is the practice of law, and everything in the build follows from that.
- The phrases that predict a result are banned at the prompt firewall, not discouraged in a style guide. "You'll get custody" is the shape of the thing it cannot say.
- When a visitor discloses a complication, the answer names it plainly and early instead of softening it into something more comfortable to read.
- Your disclaimer closes every answer, verbatim, and is translated when the answer is not in English.
- An attorney at your firm reads the written rules and approves them before launch. Nothing goes live that a lawyer at the firm has not read.
- You see every memo it produces, so nothing goes out under your letterhead unread.
- When a question does not carry enough to answer, it says so and routes the person to a consultation instead of assuming facts and building on them.
How we know the rules hold
A rule like "never predict an outcome" is easy to put in a prompt and impossible to verify by reading code. In the build behind the case study's numbers, this is how it was verified.
- The rules run as checks in the pipeline, rather than as instructions the model is asked to remember.
- A separate test tier grades every prompt change against ideal responses the attorney wrote: deterministic checks for the rules he made non-negotiable, and an LLM judge scoring the rest against his version.
- Each run is logged against the exact prompt it tested.
- The model is never retrained. The attorney's judgment lives in curated example answers the system retrieves, which keeps it readable rather than baked into weights nobody can inspect.
Read the full AI intake page for law firms
Who this answer is for
The guardrails hold when someone at the firm owns them. Where nobody can, we would rather say so now.
A good fit
- An attorney at the firm is willing to read the written rules and approve them.
- The objection to AI at your firm is exactly this one: that it might practice law.
- You want to see every memo that goes out under your letterhead.
Not a fit
- You want it to tell a visitor whether they qualify. It will not, by design.
- You want the rules written and launched without anyone at the firm reading them.
- You want us to make legal judgments. Those stay with the firm.
What we will and will not claim about compliance
The list a firm should hold any AI vendor to, with our answer on each line.
What we can claim
- Outcome language banned at the prompt firewall and enforced in the system, not in a style guide.
- Your disclaimer on every answer, verbatim, translated when the answer is not in English.
- An attorney at your firm approves the written rules before launch.
- Every memo in your inbox, so nothing leaves under your letterhead unread.
What we will not claim
- Bar compliance certification. We build the guardrails and your firm approves them; we are not the body that issues one.
- That the assistant gives legal advice or determines eligibility. It does neither, by design.
- Any prediction about a legal outcome.
- A guaranteed lift in signed matters, guaranteed revenue or guaranteed staff savings.
Questions about AI intake
How does it stay out of the practice of law?
It explains and it does not decide. Explaining what the options are is information, and software can do it. Deciding which one applies to a person is legal judgment, and the system is built so it cannot drift into that seat: it never determines eligibility and never predicts an outcome.
Which phrases are banned?
The ones that predict a result. "You'll get custody" is the shape of it. The list sits in the written rules an attorney at your firm approves, so you read it before launch rather than discovering it afterwards.
What happens to the disclaimer when the answer is in Spanish?
It is translated and it still closes the answer. Every answer carries your disclaimer verbatim in English, and translated when the answer is not.
Who writes the rules and who approves them?
We tune the prompts for the practice areas you take. An attorney at your firm reads every one and approves it before launch, and nothing goes live otherwise.
How do you know a prompt change did not break a rule?
In the build behind the case study's numbers, a separate test tier graded every change against ideal responses the attorney wrote: deterministic checks for the non-negotiable rules, an LLM judge for the rest, and a log of each run against the prompt it tested. Reading the code cannot tell you an answer was safe; that tier could.
Bring your attorney's questions to the call
Fifteen minutes on how the guardrails above answer the questions your attorney would ask. If your firm needs something we do not build, we will say so on the call.