Module 1 · Start safely
Not legal advice: AI assists, the advocate answers
In short
In legal work, AI prepares material and drafts; the advocate assesses, decides and answers the client. Write that limit into every request and label every output a draft until reviewed. This is not legal advice.
After this lesson you can
- Explain why AI output in legal work stays a draft until an advocate reviews it
- Separate legal tasks AI may assist with from legal services that remain an advocate's job
- Write the AI's role limits into every request for legal and compliance work
Terms in this lesson
Understand
This track is written for people who deal with legal documents every day: paralegals, in-house legal staff, compliance officers, secretaries at law firms, and junior advocates who want to work faster without lowering quality. The first question is always the same: where may AI help, and where must the work stay with the person in charge?
A large language model (LLM) is good at reading fast, tidying and rewriting. It can list the clauses of a forty-page contract, build a chronology summary from a stack of letters, or prepare a list of questions for a client meeting. Work like this saves hours, and the result is easy to check because the material is in front of you.
What a model cannot carry is responsibility. It does not know facts you have not told it, does not know which rule applies on the date of the events, and can make up article numbers that look right. This is called hallucination, and in legal work it most often takes the form of an article, a judgment number or an effective date that never existed. It is also not bound by an oath, a code of ethics or a duty to keep client secrets. An advocate is bound by all of these. That is why legal opinions, case strategy and answers to clients are an advocate's work, not a machine's output.
The simplest way to hold this line is to give every AI output one status: draft. A draft may be neat, complete and good. Its status changes only after the person in charge has read it, checked its sources and decided to use it. This is the principle you know from Level 1 as human in control, except that the stakes here are higher: one wrong sentence in a letter to a client can become the basis of a harmful decision.
This line also protects you. If you are not an advocate, passing AI output to a client as a legal answer means providing a service that is not your role. If you are an advocate, signing a draft you have not checked means carrying the machine's mistake under your own name.
To make this easier to use, split legal work into three layers. The material layer: tidying files, indexing documents, building a chronology from existing letters, translating drafts for internal reading. Here AI helps most safely, as long as the material is masked where needed. The draft analysis layer: mapping issues, comparing clauses with the office's standard list, suggesting questions not yet answered. AI may help, but the result is only a map for the advocate to read, not a conclusion. The decision layer: legal opinions, choice of strategy, advice to the client, signatures, filings with a court or agency. This layer belongs entirely to the person in charge. AI may help tidy the wording of a letter whose content the advocate has decided, but it may not decide that content.
Before
The prompt
Our client is being charged a late-payment penalty of five percent a day by its supplier. Is that penalty valid? Keep it short so I can forward it to the client.
What came back
A penalty that large is invalid because it is excessive and against the principle of fairness. The client does not have to pay and may refuse the invoice.
After
The prompt
Goal: I am a paralegal preparing material for the advocate handling a client's late-payment penalty. I am not asking for a legal opinion. Reader: the responsible advocate, not the client. Material: the penalty clause from the contract, already masked, below. Limits: do not conclude whether it is valid. Do not name articles except from text I paste. Shape: a list of issues, then a list of documents and facts to find. Check: mark everything the advocate must decide. [penalty clause]
What came back
Draft for advocate review. Issues: (1) how the clause calculates the penalty and whether it has a cap; (2) whether other contract terms cover delay or force majeure; (3) how the penalty was invoiced or paid before. To find: the full contract with its amendments, the invoices, payment records, correspondence about the delay. For the advocate to decide: the assessment of the penalty's size and the step to advise the client.
What changed
- The first answer gives a legal conclusion without the full facts, ready to forward to the client. That is the decision layer taken by a machine.
- The second request names the advocate as reader, limits the AI's role and asks for issues and missing material.
- The draft label on the first line makes its status clear to anyone who opens the file later.
Common mistake
- Mistake
- Forwarding an AI answer to a client, a manager or the other side as a legal answer, only because it is well written and confident.
- Why it hurts
- The client decides on that answer: to pay, refuse, sign or wait. If the answer is wrong or incomplete, the harm is real and no advocate ever checked it. The person who forwarded it shares the blame.
- Do this instead
- Keep AI output as a labelled internal draft. Send it to the responsible advocate with its material and sources, and let the advocate decide what is told to the client.
Try
Sorting card
For each task, decide: may AI help as a draft, or is it the advocate's work from the start?
Signing a demand letter on the client's behalf
Answering a client's message about the chances of winning
Deciding whether the client should settle or sue
Building a chronology from correspondence in the file
Indexing three hundred masked evidence documents
Preparing a list of questions for a witness interview
6 cards left
You asked AI to summarise a judgment for tomorrow's team meeting. The summary is good. What is the right step before uploading it to the case folder?
Take home
Role limits in every legal request
- I state who the reader is: an advocate, the legal team or myself, not the client
- I state that I am not asking for a legal opinion or conclusion
- I ask for issues, questions or missing material, not a final answer
- I label every output a draft until an advocate reviews it
- I hand decision-layer tasks to the person in charge
What went wrong for us
AI output in legal work is always a draft
- Context
- When we use AI for work that touches law, such as tidying terms of service or answering users' questions about their data, we separate who prepares from who decides.
- What tends to go wrong
- A neat draft easily looks like a final document. Without a firm rule, sentences from a model can move into a letter, a page or a message without ever being read by the person in charge.
- What it costs if ignored
- A wrong legal sentence in a document outsiders read can become a promise, an admission or advice we never meant to give.
- Root cause
- Quality of language does not show whether the content was checked. A text's status must be written down, not guessed from how neat it is.
- How we do it
- Every AI output for legal work is labelled a draft and comes with the sources it used. The label is removed only after the person in charge reads and approves it.
- The rule we keep
- No legal text from AI goes out before a legal expert has reviewed it, and the reviewer is recorded.
The lessonLegal text from AI stays a draft until the person in charge reads and approves it.
What it means for your workWrite "draft" on the first line of every AI output and name who must review it.