PII handling — redaction, PDPA-2023, and citizen dignity

Three ways to see it

1. 1

   PII, personally identifiable information, is anything that lets someone work out who an individual is. The obvious: full name, CNIC, passport number, mobile, email, bank account, address. The less obvious: date of birth combined with city, biometric template, IP address, browsing pattern, photograph. The much less obvious: voice recording, gait pattern, handwriting sample. Under PDPA-2023, anything that identifies a Pakistani citizen, alone or in combination with other data, is regulated. Sending it to a third party without lawful basis or consent is a violation. The chat with an AI vendor is sending it to a third party. There is no informal exemption for 'I was just asking'.

2. 2

   Three redaction strategies, in order of strength. Strategy one: tokenisation. Before the prompt leaves your machine, every CNIC is replaced with a stable token like CNIC_A1B2, every name with PERSON_001, every phone with PHONE_X. The AI sees only tokens. Whatever the AI returns is then de-tokenised inside your environment before showing the user. The AI never saw real PII. Strategy two: aggregation. Instead of asking 'is this customer creditworthy', ask 'is a customer with income X, age Y, employment type Z creditworthy', dropping name and CNIC entirely. The decision logic stays useful, the identity disappears. Strategy three: on-premise or in-country processing. Run the model inside your own data centre (or a Pakistani cloud region with data-residency guarantees) so the data physically never leaves jurisdiction.

3. 3

   Two practical tools that do most of the work. First, Microsoft Presidio (an open-source library) auto-detects PII patterns across English and many other languages and replaces them with placeholders. Plug it in before every prompt; it adds about 50 milliseconds. For Urdu and Pakistani-specific patterns like CNIC format (5-digit, 7-digit, 1-digit), you write a small extension. Second, the major vendors now offer enterprise tiers (Anthropic Enterprise, OpenAI Enterprise) with contractual no-training-on-your-data, encryption-at-rest, and data-residency options. For most regulated workloads in Pakistan, paying for the enterprise tier is a fraction of the cost of one breach.

The why-tree

Try this with Claude

AI-edge prompt: 'Here is a sample customer loan application [paste with synthetic data]. Identify every field that would be PII or sensitive personal data under Pakistan's PDPA-2023. For each, recommend either redact-and-tokenise, aggregate-and-drop, or keep-with-consent. Output as a markdown table. Then write the SQL or Python snippet that performs the redaction before the data hits any external API.'