Customer identifiers stripped before the request reaches the internet.
Your team writes naturally in plain names and real details. On the device, Marling strips the details that identify a customer before a request reaches ChatGPT, Claude or Gemini, and restores them in the reply.
What happens to a message.
-
You write it as you would anyway
“Draft a letter for Matthew Sorensen, DOB 04/17/1986, confirming the new arrangement.”
-
The device finds the identifiers
Names, dates of birth, phone numbers, addresses, Social Security numbers, record and matter numbers, and a customer’s name inside a case caption. This runs on the device, before anything is sent.
-
Placeholders go out, not the real values
The model works on “Draft a letter for <FULL_NAME>, DOB <DOB> …”. The substance of the request stays intact, so the answer is still useful.
-
The answer comes back complete
The device swaps the real values back in. You see a finished letter with the right name on it.
What leaves, what stays.
| Item | Where it goes |
|---|---|
| Names, dates of birth, contact details, Social Security numbers, record and matter numbers | Stripped on the device before anything goes out, and restored in the reply |
| The substance of a request | Leaves, identifiers removed |
| A passage from one of your documents used in an answer | Leaves, identifiers removed |
| Your full documents, your saved conversations, and your search index | Held on the device in your office, with an encrypted backup to our backup server |
| Dictation audio | Never leaves, never stored |
If masking can’t run, nothing sends
If the masking step on the device is unavailable, the message is refused and stays on the device.
How often it catches a name
In our tests on deliberately awkward text, such as names in file names, tables and logins, Marling hid about 96% of names it had never seen as standard, and about 99% once a business’s customer list was loaded. It is most likely to miss a surname standing on its own, for example in a file name, a spreadsheet row or a line copied from another system, so check text like that before you paste it. Marling Support can add your customer list so those names are always hidden.
You decide what counts as sensitive
Tell Marling Support your record or matter number format from two or three examples, or mark your business’s own name as fine to show, and we set it up with you. Every change is previewed and checked before it takes effect.
Questions people ask before they start.
So the model still sees my customer’s situation?
Yes. It sees “customer is behind on payments and asking for a plan”, not whose payments. That is what makes the answer useful and the exposure small.
Does the AI learn from our conversations?
No. On both Hosted Marling and the office device, Marling uses the providers’ commercial services, whose terms exclude training models on your requests. Separately, Shared Memory uses approved preferences and rules about your business to guide future replies. Its suggestions wait for an Admin or Marling Support to approve them.
Will it hide the case law I’m citing?
No. Published case names, reporter citations, court rules and statutes reach the model exactly as you wrote them. Your customer’s own name is removed, including inside a case caption, and Marling tells the two apart by the citation: a published authority carries a reporter cite, a live matter carries a docket number.
Is this enough for regulated work?
Masking alone does not make your work compliant. On the office device with the Mental Health or Legal plan, Marling replaces identifying details it recognizes before a request reaches an AI model over the internet. Some details can be missed. If the masking step is unavailable, the message is refused and stays on the device. Hosted Marling and the general small-business plan do not apply this masking step. Before you use real customer information, we walk through the setup, safeguards, and agreements your work requires with you.