L1Lesson 1Free

Your first real legal AI win, set up safely

Get a real draft in minutes, with the safe-tool, de-identify, and verify habits built in from the first task.

After this, you'll be able to read an AI tool's terms of service for what matters, pick a zero-retention tool over a consumer one for work data, strip the client out of every input, and verify every fact before anything leaves your desk.

Before you start

Complete Privilege and confidentiality 101 first; this lesson turns the duty it explains into the concrete tool-and-habit setup that honors it.

The idea

By the end of this lesson you will have a working legal AI assistant that drafts a real letter, summary, or memo in minutes, and you will have set it up so it never costs you a client or a sanction. That is the payoff: not a toy, a tool you use on Monday. If you are thinking "AI is more trouble than it is worth once you account for the risk," this lesson is the direct answer, because the safety is three quick habits, not a second job.

The headline failures, the lawyers who filed AI-invented cases, did not have a tool problem; they had a verify problem. A model will hand you a case name, a citation, and a holding that all look real and none of which exist. That is hallucination, and the fix is not a better tool. It is you, reading the source. So one rule sits under everything below: a human checks every citation and every fact before anything leaves your desk.

Before you verify anything, you choose what to type into. You met this choice in the privilege lesson; here you actually make it, once, for real work. A consumer tool is the free or personal version you sign up for in a browser, like a standard ChatGPT or Claude account. An enterprise tool is the paid, firm-controlled version with a business contract behind it.

The difference is not the quality of the answers. It is what the company is allowed to do with what you type. A consumer tool's terms of service, the contract you agree to by clicking "I accept", often let the provider keep your inputs and use them to train future models.

Why this breaks privilege: if your client's facts can be retained, reused, or disclosed by an outside company, a court may decide you had no reasonable expectation of privacy, and the protection is gone. The Clio guidance frames the safe path as zero-retention: the provider keeps no record of what you typed once it has answered you. No stored copy means nothing for the provider to leak, hand over, or train on.

So the first real skill is reading the terms of service for two words: retention and training. You are looking for a tool that says it does not retain your inputs and does not train on them, ideally in a written business agreement, not a marketing page. If a tool cannot tell you that plainly, it is a consumer tool, and consumer tools are for public, non-client questions only.

Reading the terms for two wordsBefore you type anything real, the terms tell you one thing: is this a tool you can put client work into, or only public questions.

Open the termsThe actual terms of service, not the 'we value privacy' marketing page.
Does it keep your inputs?Look for the word retention. You want a clear no, in a written business agreement.
Does it learn from them?Look for the word training. You want a clear no here too.
Public questions onlyAny yes, or no clear answer, makes this a consumer tool. Never client data.
Safe for client workKeeps nothing, trains on nothing, in writing. This is the tool you set up.

The safe setup is not one precaution, it is three stacked on top of each other, and they each catch what the others miss. A safe tool keeps the provider from holding your data. De-identifying keeps the client out of the box even if the tool is wrong about its own terms. And the human verify step catches the made-up case the first two layers do nothing about. Picture them as three layers, each guarding against a different failure.

The three layers of a safe setupEach layer guards against a different failure. Stack all three and one slip does not cost you a client or a sanction.

A zero-retention toolThe provider keeps nothing once it answers, so there is nothing to leak or train on.
De-identified inputBracketed placeholders, so even if the tool is wrong about its terms, the client never enters the box.
A human verifies every factCatches the made-up case the first two layers do nothing about. The drafter is never the signer.

Be firm about where reading the terms gets you: it tells you what a tool claims, not whether it is safe to bet a client's privilege on. The practical rule (the one you met in the privilege lesson) is harder-edged. A consumer tool is unsafe for client data no matter how its terms read, only a firm-procured enterprise tool under a signed data processing agreement is safe for it, and any client-identifiable input also needs the client's informed consent, not a boilerplate engagement-letter line.

Here is the before and after: the unsafe way is to paste "Draft a demand letter for my client Jane Doe, account 4471, re: her dispute with Acme Corp" into a free browser chatbot. The safe way is to use an enterprise, zero-retention tool and paste "Draft a demand letter for [CLIENT], [ACCOUNT NUMBER], re: a billing dispute with [COUNTERPARTY]." De-identifying means swapping every name, number, and identifying fact for a bracketed placeholder, so even on a safe tool, the raw client never enters the box. Two layers, not one.

De-identifying is swapping every name, number, and identifying fact for a bracket. The request still works; the client never enters the box.

	Raw client paste	De-identified paste
The client's name	"my client Jane Doe"	[CLIENT]
The account number	"account 4471"	[ACCOUNT NUMBER]
The other side	"her dispute with Acme Corp"	"a billing dispute with [COUNTERPARTY]"
What the AI is asked	Draft a demand letter	Draft the exact same demand letter
What the tool sees	A named, traceable client	A request with nobody identifiable in it

The placeholders do not weaken the draft. You fill the real names back in yourself, after, by hand.

The last layer is the one that keeps your name out of the sanctions headline. Every citation, every quote, every factual claim the AI gives you gets checked by a person against the real source before it goes into a filing, an email, or a client's hands. The AI drafts; a qualified human signs off. Always, with no exception for "it looked right."

Now try it. Open the tool you are considering, find its terms of service, and answer two questions in writing: does it retain my inputs, and does it train on them? Then give it the paste-ready ground-rules prompt below so it knows how to behave.

A safe tool, a de-identified input, and a human who verifies every fact: that is the entire setup. Get it in place once and it runs quietly in the background on every task after this. Next, you point this assistant at its first real job: turning a long contract or case into one clear page you can trust.

The unsafe path skips the tool check, keeps the client's name in, and trusts the output. The safe path does all three.

	Unsafe first use	Safe first use
The tool	A free consumer account you never checked	An enterprise or zero-retention tool whose terms you read
What you type	The real client name, account, and facts	Bracketed placeholders like [CLIENT] and [COUNTERPARTY]
What happens to it	May be retained and used to train future models	Kept nowhere once you have your answer
Before it leaves your desk	You trust the draft and send it	A human verifies every citation and fact against the source

The safe path is three layers. Skipping any one is how careful lawyers end up sanctioned.

Try it (16 min)

Watch out for

Trusting a 'we take privacy seriously' marketing page instead of the actual terms of service. The words that matter are retention and training, in the binding agreement, not the sales copy
Treating de-identifying as optional once you are on a safe tool. It is the second layer: use a zero-retention tool AND strip the client out, because layers fail and you want a backup
Pasting a real client name 'just this once' because the matter feels low-stakes. There is no low-stakes exception to confidentiality; the habit is the protection
Asking the AI to confirm that its own citations are real. It will happily agree to a case it invented; verification means you opening the source, not the model reassuring you
Assuming your firm has no rule, so anything goes. The absence of a written AI policy is not permission; check whether your firm has an approved-tool list before you pick one

Paste this into Claude

Then, in your chosen tool, paste this safe-use ground-rules prompt so the assistant knows the rules of engagement:

"You are my legal drafting assistant. Follow these ground rules for everything we do together:
1. I will give you facts with placeholders like [CLIENT], [COUNTERPARTY], [DATE], [AMOUNT]. Never ask me to fill in real client names or identifying details, and if I paste a real one by mistake, stop and warn me.
2. Treat every case name, statute, citation, quote, and number you produce as UNVERIFIED. Label them clearly so I know to check each one against the original source before I rely on it.
3. If you are not sure a fact or citation is real, say 'I am not certain this exists, verify it' rather than presenting it confidently.
4. You assist my judgment, you do not replace it. A qualified person reviews and signs off on everything before it leaves my desk.
Confirm you understand these four rules, then wait for my first task."

After it confirms, give it one small, fully de-identified task (for example, "Draft a two-sentence acknowledgment that we received [COUNTERPARTY]'s letter dated [DATE]") and watch whether it respects the rules.

What good looks like

You can state, in writing, whether your chosen tool retains inputs and whether it trains on them
You chose an enterprise or zero-retention tool for anything involving client facts, and reserved any consumer tool for public, non-client questions
Every input you sent used bracketed placeholders, so no real client name, number, or identifying fact entered the box
The assistant confirmed the four ground rules before doing any task
You verified at least one fact or citation in its output against the real source before treating it as true

When this breaks

"The answers are good, so the tool is fine" breaks because answer quality and data safety are unrelated; a consumer tool can give brilliant drafts while retaining and training on every client fact you paste, which is exactly what costs you privilege
"I de-identified, so I do not need to verify" breaks because de-identifying protects the client's confidentiality, not the truth of the output; a de-identified prompt can still return a fabricated case, so the human verify step is a separate, non-skippable layer

You can now

Set up one work-ready AI session: confirm in writing that your tool does not retain or train on inputs, send only a bracketed, de-identified prompt, and verify one citation or fact in the output against its real source before trusting it. If you cannot confirm the retention-and-training answer from the tool's terms, the one thing that is wrong is the tool choice, not the prompt: switch to an enterprise or zero-retention tool, or keep this session to public, non-client questions only.

Key takeaways

Safe first use is three layers, not one: a zero-retention tool so the provider keeps nothing, de-identified inputs so the client never enters the box, and a human who verifies every fact before it leaves your desk. With the assistant set up, the next lesson points it at its first real job: turning a long contract or case into one clear page you can trust.

Read the terms of service for two words before you type anything real: retention (does it keep your inputs) and training (does it learn from them)
Use an enterprise or zero-retention tool for any client work; a consumer tool that retains and trains on inputs can cost you privilege
De-identify every input with bracketed placeholders, even on a safe tool, so the raw client never enters the box; two layers beat one
The single rule that prevents the sanctions headline: a human verifies every citation, quote, and fact against the real source before it leaves your desk
The absence of a firm AI policy is not permission; check for an approved-tool list, because someone may have already vetted the safe options for you

Was this helpful?

Up nextSummarize a contract or case in plain English

← Back to AI, Law and Compliance

Your first real legal AI win, set up safely

Get a real draft in minutes, with the safe-tool, de-identify, and verify habits built in from the first task.

Raw client paste

De-identified paste

The client's name

"my client Jane Doe"

[CLIENT]

The account number

"account 4471"

[ACCOUNT NUMBER]

The other side

"her dispute with Acme Corp"

"a billing dispute with [COUNTERPARTY]"

What the AI is asked

Draft a demand letter

Draft the exact same demand letter

What the tool sees

A named, traceable client

A request with nobody identifiable in it

Unsafe first use

Safe first use

The tool

A free consumer account you never checked

An enterprise or zero-retention tool whose terms you read

What you type

The real client name, account, and facts

Bracketed placeholders like [CLIENT] and [COUNTERPARTY]

What happens to it

May be retained and used to train future models

Kept nowhere once you have your answer

Before it leaves your desk

You trust the draft and send it

A human verifies every citation and fact against the source

First, do the tool check (no AI needed for this part). Open the AI tool you are thinking of using for work and find its terms of service or privacy policy. Write down the answer to two questions: 1. Does it RETAIN my inputs after answering? (Look for "retention", "we store", "we keep".) 2. Does it TRAIN on my inputs? (Look for "train", "improve our models", "your content may be used".) If the answer to either is yes, or you cannot find a clear answer, that tool is for public, non-client questions only. Then, in your chosen tool, paste this safe-use ground-rules prompt so the assistant knows the rules of engagement: "You are my legal drafting assistant. Follow these ground rules for everything we do together: 1. I will give you facts with placeholders like [CLIENT], [COUNTERPARTY], [DATE], [AMOUNT]. Never ask me to fill in real client names or identifying details, and if I paste a real one by mistake, stop and warn me. 2. Treat every case name, statute, citation, quote, and number you produce as UNVERIFIED. Label them clearly so I know to check each one against the original source before I rely on it. 3. If you are not sure a fact or citation is real, say 'I am not certain this exists, verify it' rather than presenting it confidently. 4. You assist my judgment, you do not replace it. A qualified person reviews and signs off on everything before it leaves my desk. Confirm you understand these four rules, then wait for my first task." After it confirms, give it one small, fully de-identified task (for example, "Draft a two-sentence acknowledgment that we received [COUNTERPARTY]'s letter dated [DATE]") and watch whether it respects the rules.