When your company connects Claude or ChatGPT Enterprise, the Pragmatico Platform reads free text from those conversations (message wording, chat names, project names, and attachment filenames) to manage your proficiency scoring and coaching. Since we cannot control what customers put into their AI tenant as they operate their business, Pragmatico has implemented enterprise-grade LLM filtering and hygiene to ensure sensitive data never enters our systems.
This means every message is evaluated for sensitive data and is either omitted, redacted, or accepted in a separate, isolated enviroment that Pragmatico owns and operates outside our production data warehouse. This ensures all customer data is protected, isolated, and redacted or omitted where relevant.
What We Keep
Ordinary work text that doesn't match a redact or omit rule is stored as written in our database to support proficiency management and coaching. This is the overwhelming majority of the safe-for-work AI usage that we see from our customers.
The type of data Pragmatico keeps and stores
Names, locations, dates, times, and links stay in the message. This level of data allows for specific coaching without crossing the threshold into PII or identifiable, sensitive data.
Below is an example of a chat with real turns of work and organizational context that is safe to store.
Scenario 1: Using AI to review and iterate on artifacts:
Redact (identifier replaced, sentence kept)
In some cases, a message may still need to be stored with context, as it is part of everyday work. Examples include judging or improving email threads, or business transcripts. Transcripts have value for coaching and reporting, but Pragmatico refuses to store the specifics of the data because of compliance requirements. In this case, we redact the message and replace it with a placeholder variable.
Types of messages we redact
What we may find in your messages | What our system replaces it with |
| |
Phone |
|
Payment card (with payment words nearby) |
|
IBAN |
|
Crypto wallet |
|
US Social Security number |
|
ITIN or driver’s license |
|
Passport |
|
HR/case / ticket ID |
|
Scenario 2: A message with redacted messages
Omit text (record kept, wording not stored)
Some messages contain sensitive topics that shouldn't be stored, even if you redact the words. In this case, even redacting variables would still show a sensitive conversation. When this happens, the Pragmatico platform records that a message happened (when, who, which chat, whether files were present), but it omits the turn of the message and its full content. The product has no way to read the original.
Types of messages we omit
Categories | Types of Content |
Pay | You disclose an amount (I make $95,000) |
Performance | PIP, written warning, I was fired, let me go |
Protected class | Stated pregnancy, disability, religion, orientation, age, genetic or family medical history |
Health | Named conditions, medical leave, rehab / sober / therapist in a health frame |
Grievance | Harassment, discrimination, investigation |
Family leave | FMLA, Parental Leave, Divorce as listed |
Immigration | H-1B, green card, visa sponsorships |
Background | Criminal-record frame, probation officer, bankruptcy |
Never stored
File contents from Claude or ChatGPT chats. The Pragmatico Platform may keep the typed prompt and the filename (the filename follows Keep / Redact / Omit above). For Claude, it may record that files were present. The file itself is not stored.
A copy of the omitted wording, the matched phrase, or where it sat in the sentence. The Platform keeps the category (for example, health or pay), not the sentence.
Network TLS fingerprints.