Skip to content
tightlip
Menu

For law firms

Use AI on client work without pasting in the client.

Tightlip swaps client names, matter facts, case numbers and account details for placeholders before a prompt reaches ChatGPT, Claude or Copilot. The draft comes back with the real names put in on the lawyer’s own computer.

1. Your textOn your device
Privileged and confidential. Memo to file. Matter: Delgado v. Brennan Freight LLC, Case No. 1:24-cv-03187. Our client Maria Delgado (DOB 03/14/1981) says Brennan Freight LLC held back her final commission of $48,200 after she resigned. Opposing counsel is Daniel Okafor (dokafor@okaforlegal.com, 212-555-0143). Ms. Delgado wants any settlement wired to her account at Harbor Point Bank, routing 021000021, account number 4471009233.
2. What the AI seesSent to the AI
Privileged and confidential. Memo to file. Matter: [PERSON_1_LAST] v. [ORG_1], Case No. [CASE_NUMBER_1]. Our client [PERSON_1] (DOB [DOB_1]) says [ORG_1] held back her final commission of $48,200 after she resigned. Opposing counsel is [PERSON_2] ([EMAIL_1], [PHONE_1]). Ms. [PERSON_1_LAST] wants any settlement wired to her account at [ORG_2], routing [ROUTING_NUMBER_1], account number [BANK_ACCOUNT_1].
3. The answer you readRestored on your device
Asked: Draft a short demand letter to opposing counsel.Dear Daniel Okafor, We act for Maria Delgado in Delgado v. Brennan Freight LLC, 1:24-cv-03187. Your client held back her final commission of $48,200. Please confirm within 14 days that Brennan Freight LLC will pay it in full.
Real output from Tightlip on a made-up memo to file: 10 items protected (2 names, 2 organisations, 1 case number, 1 date of birth, 1 email address and 3 more). The swap table that links each placeholder to its real value stayed on the device.

What bar opinions say

Three state bars on putting client information into generative AI, quoted word for word.

“If client confidences and secrets are not protected by the GAI tool, then the lawyer must anonymize their inputs to protect client details.”
Alaska Bar Association, Ethics Opinion 2025-1, 23 April 2025

Tightlip does the anonymizing step on your computer, before the prompt goes out.

“When using an open model without client consent, lawyers need to anonymize or redact information that the client considers sensitive.”
Oregon State Bar, Formal Opinion 2025-205, February 2025

Tightlip swaps names and identifiers for placeholders, and your always-protect list covers what your client names as sensitive.

“Hence, an attorney should take care that any information inputted into a generative AI product does not identify the client or the nature of the representation.”
Kentucky Bar Association, Ethics Opinion KBA E-457, 15 March 2024

Tightlip replaces client names, case numbers and the matter names you list. The facts you leave in are still yours to judge.

Bar opinions are guidance. Your jurisdiction’s rules and your client’s instructions govern.

What Tightlip catches in legal work

Every value is swapped for a numbered placeholder. The same party keeps the same placeholder for the whole conversation, so the AI can still reason about who owes what to whom.

Parties and people[PERSON_1]
Names after cues such as “our client”, “plaintiff”, “defendant” or “opposing counsel”, names in case captions, honorifics (Ms., Judge, Hon.), greetings and signatures, and names from a first-name list followed by a surname.
Case and docket numbers[CASE_NUMBER_1]
Federal numbers such as 1:24-cv-03187, state docket styles, and anything after “Case No.” or “Index No.”.
Companies and firms[ORG_1]
Capitalised names ending in LLC, Inc., Corp., LLP, PLLC, P.C., Bank, Holdings, Partners and similar.
Settlement and payment details[BANK_ACCOUNT_1]
Routing numbers confirmed by their check digit, account numbers after an account keyword, IBANs, card numbers.
Contact details[EMAIL_1]
Email addresses, US and international phone numbers, street addresses with city, state and ZIP.
Identity numbers[SSN_1]
Social Security numbers, tax IDs, dates of birth, passport and driver licence numbers after their labels.

Your always-protect list

Detection finds names and numbers by pattern. It cannot know that “Project Heron” is your client’s acquisition or that “Northgate” is a matter name. Add those words to your always-protect list and they are swapped every time, in every tool.

The list lives in your browser on your computer. Tightlip’s servers never receive it.

How a firm starts

  1. 1

    Check a real memo

    Paste one into the free leak check. It runs in the browser and sends nothing, so you can use a live document.
  2. 2

    Pilot the workspace

    Each lawyer who signs up gets USD 20 of credits, about 400 protected prompts, with the AI model included.
  3. 3

    Add the extension

    For lawyers who already have ChatGPT, Claude or Copilot, the browser extension protects the message box on send.
  4. 4

    Roll it out

    For firm-wide use with an annual invoice, email hello@tightlipai.com.

Questions partners ask

Does this meet our duty of confidentiality?
That judgement belongs to the firm and your bar’s guidance on AI. What Tightlip gives you is a concrete control: client names, numbers and your listed terms are replaced before the text leaves the lawyer’s computer, and the lawyer sees the exact text the AI will receive.
Can the AI still follow the facts?
Yes. Each party keeps one placeholder for the whole conversation. If Maria Delgado is [PERSON_1], a later “Ms. Delgado” becomes [PERSON_1_LAST], so the model still knows it is the same person.
What if it misses a name?
It can. Detection is automatic and imperfect. The workspace shows what the AI will see before you send, the extension lets you see what was sent, and the always-protect list covers the names that matter most to you.
Does Tightlip keep a copy of our documents?
No. With the extension, the protected text goes straight from the lawyer’s browser to the AI tool and Tightlip receives only a credit count. In the web workspace, protected text passes through our server to the model and is not stored or logged. The full flow is on the security page.
What does it cost?
USD 0.05 per protected prompt, paid in advance as credits. No seats. Firm-wide annual invoices are available. See pricing.

Paste a real memo. See what the AI would have read.