Module 1 · Start safely
Client confidentiality: what must not go into AI
In short
An advocate must keep secret what they learn from a client. Privileged documents and client identities do not go into general AI services. Mask first, check metadata too, and use an office-approved tool for what stays confidential.
After this lesson you can
- Recognise client documents and client identities that must not go into a general AI service
- Mask a case file or contract before asking AI for help without breaking its legal meaning
- Choose an office-approved tool for legal documents that stay confidential even after masking
Terms in this lesson
Understand
A general AI service is one you open with a personal or free account, with no contract between your office and its provider. Anything you paste there leaves your office's control: it may be stored, may be read by the provider's staff for review, and its settings may change without you noticing. For ordinary work, masking keeps that risk manageable. For legal work, some material must not go in at all.
Privileged documents: correspondence between client and advocate, strategy memos, interview notes, unfiled claims or defences, internal legal opinions. Their value lies in being confidential. Client identities: the client's name, the other side's name, witness names, case numbers not yet public, even the fact that someone is seeking legal help from your office. Trade secrets in contracts: prices, payment schemes, acquisition plans, specially negotiated terms. Parties' personal data: addresses, ID numbers, health or financial data, which the PDP Law also protects and some of which is specific personal data.
For material like this there are three options. First, mask it until no outsider could recognise the client or the matter, then use AI for questions that truly do not need identities, such as clause structure or how to lay out a chronology. Second, use an office-approved tool, with a contract covering storage, use of data for training and where processing happens. Third, do it without AI. The third option is not a defeat; for a sensitive case strategy memo, it is often the most sensible choice.
The basics of data that must not go in and how to mask it are covered in Level 1. Reread how to mask data if needed; this lesson only adds what is particular to legal files.
Masking legal files has its own traps. Replacing names alone is rarely enough. A letterhead, a company name in a signature block, a case number, the court and hearing date, the address of the disputed property or one unique fact can make the matter recognisable to anyone who follows the news. Replace them with consistent markers, such as PARTY-A, PARTY-B, [VALUE-1], [DATE-1], so the relationships between parties still read clearly. Keep the key list in the office file, separate from what you send.
Do not let masking break the legal meaning. If the order of dates matters for a deadline, replace dates with day differences rather than deleting them. If the parties' roles matter, such as seller and buyer, keep the roles. And check what you cannot see: comments, tracked changes and document properties often still hold real names.
Before
The prompt
Please tidy this chronology for a memo. [pastes the client's full e-mail with the director's name, the other company's name, the case number and the invoice value]
What came back
A tidy chronology, complete with every name, the case number and the invoice value, now stored in a service your office has no contract with.
After
The prompt
Goal: I need a tidy chronology for an internal memo. Reader: an advocate in my office. Material: event notes I have masked. PARTY-A is the client (buyer), PARTY-B the seller. I replaced dates with D+number from the contract signing. Limits: do not guess who the parties are, do not add events not in the notes. Shape: a table with columns day, event, supporting document. Check: mark events that have no supporting document yet.
What came back
Chronology table: D+0 contract signed (document: contract); D+30 PARTY-B ships part of the goods (document: delivery note); D+45 PARTY-A sends a complaint (no supporting document yet); D+60 invoice [VALUE-1] sent (document: invoice).
What changed
- A chronology does not need real names. Party roles and day differences are enough.
- The key list for PARTY-A and [VALUE-1] stays in the office file, so you can restore the real names yourself.
- The missing-document marker is the next piece of work, not a conclusion.
Common mistake
- Mistake
- Masking the contract text on screen, then uploading the original file that still holds comments, tracked changes and the author's name in its properties.
- Why it hurts
- Tracked changes can show the client's negotiating position, comments can hold the advocate's views and document properties can name the office and the client. All of it is sent even though the page text is clean.
- Do this instead
- Copy the masked text into a fresh, clean file, or paste it as plain text. For documents that stay sensitive, use the office-approved tool.
Try
Sorting card
For each material, which treatment is right if only a general AI service is available?
Statute text you copied from an official source
The indemnity clause from a client contract
The office's blank contract template with no entries
Notes of a witness interview
An unfiled case strategy memo
An event chronology with party roles and day differences
6 cards left
You masked every name in a dispute file that is in the news. The facts are still unique: a building that collapsed in a certain city in a certain week. What is the right step?
Take home
Before pasting a client document
- This is not privileged: not client correspondence, a strategy memo, interview notes or an unfiled draft
- Client, opposing party, witness names and the case number are replaced with consistent markers
- Values, addresses and unique facts that could identify the matter are replaced or removed
- Comments, tracked changes and document properties are not sent along
- The masking key list is kept in the office file, separately
- If it stays confidential after masking, I use the office-approved tool or no AI
What went wrong for us
Identities stay out of general AI services
- Context
- We often need AI help with text that names people and organisations: complaints, contracts, meeting notes. Some of that text is confidential because of its content, not only its names.
- What tends to go wrong
- Replacing names alone is rarely enough. Letterheads, document numbers, unique facts and tracked changes in a file can still show who is being discussed.
- What it costs if ignored
- If identities are sent along, control of that data is lost: we cannot pull it back from a service we have no contract with.
- Root cause
- Masking was done on the visible text, while a file carries more data than shows on screen.
- How we do it
- We mask with consistent markers, keep the key separately and send clean text, not the original file. If something can still be recognised after masking, we use a contracted tool or no AI.
- The rule we keep
- Confidential documents and people's identities do not go into general AI services; only clean, masked text does.
The lessonMask the text, not just the names, and send a clean copy, not the original file.
What it means for your workBefore pasting a work document into AI, copy its text into a new file and ask whether an outsider could still guess who it is about.