Personal Data (PII)

Personally identifiable information (PII) can identify a person directly, or combined with other data: a birth date, a postcode and a gender together often single out one person.

Common fields and whether they count as PII
Field	PII?	Why
Name	Usually	Identifies with context
Email address	Yes	Often unique
Phone number	Yes	Reaches the person
Home address	Yes	Locates the person
Passport number	Yes	Government identifier
IC/NRIC number	Yes	Also encodes birth date
Customer ID	Potentially	Joins back to the person
IP address	Context-dependent	Linkable to a household
Date of birth	Potentially	Strong quasi-identifier
Bank account details	Yes, sensitive	Financial data
Password	Credential	Store only a salted hash

Three techniques reduce the risk:

Masking hides part of a value: john.smith@example.com becomes j***@example.com.
Pseudonymization replaces an identity with an artificial one: John Smith becomes CUSTOMER_18372. A separately kept key or mapping can reverse it, so GDPR still treats it as personal data.
Anonymization removes identifying information so the person can no longer reasonably be identified, for example by aggregating orders into monthly totals. Only truly anonymous data falls outside GDPR.

Apply them before data goes to an external large language model (LLM): identify the PII, drop what the model does not need, mask or pseudonymize the rest, and review what comes back.

Protecting PII before data reaches an external LLM

Here is that pipeline for one BookNest support ticket (sample data, not a real person). The pseudonym is a keyed hash (HMAC): the same customer always gets the same ID, and nobody without the key can reverse it.

pii_before_llm.py: pseudonymize, mask and scrub a ticket before it leavesPython

import hashlib, hmac, json, os, re
KEY = os.environ.get("PSEUDONYM_KEY", "demo-key-keep-in-a-vault").encode()
ticket = {"name": "John Smith", "email": "john.smith@example.com", "ic": "900101-12-5678",
          "total": 150.00, "text": "Refund please, call me on +60 12-345 6789. Book damaged."}
def pseudonym(value):          # same input + same key -> same ID; no mapping table to leak
    digest = hmac.new(KEY, value.encode(), hashlib.sha256).hexdigest()
    return "CUSTOMER_" + str(int(digest[:8], 16) % 100000).zfill(5)
def scrub(text):               # PII hides in free text too (IC first: PHONE would match it)
    for label, pattern in [("IC", r"\b\d{6}-\d{2}-\d{4}\b"), ("EMAIL", r"\S+@\S+"),
                           ("PHONE", r"\+?\d[\d -]{7,}\d")]:
        text = re.sub(pattern, f"[{label}]", text)
    return text
user, domain = ticket["email"].split("@")
payload = {"customer": pseudonym(ticket["email"]), "email": user[0] + "***@" + domain,
           "total": ticket["total"], "text": scrub(ticket["text"])}
print(json.dumps(payload, indent=1))      # name, IC and phone never leave

Output

{
 "customer": "CUSTOMER_63904",
 "email": "j***@example.com",
 "total": 150.0,
 "text": "Refund please, call me on [PHONE]. Book damaged."
}

Regular expressions miss things, such as a name typed into the text. Production pipelines add a trained detector such as Microsoft Presidio (PII Classification and Masking), keep the key in a secrets manager, and send highly sensitive data only to a model deployed inside the organization.

MENU