Data, privacy, and PDPA-2023: what stays in, what stays out

Three ways to see it

1. 1

   Pakistan's Personal Data Protection Act 2023 (PDPA-2023) is the binding legal frame in 2026. The Act distinguishes a data controller (the organisation that decides why data is collected) from a data processor (the organisation that handles data on the controller's behalf). When you call OpenAI, Anthropic, or any hosted LLM with personal data in the prompt, you are using that vendor as a processor. The Act requires a written processor agreement, a data protection impact assessment for high-risk processing, retention limits, and notification within 72 hours of a breach. None of this is optional. None of this is well known yet. Knowing it gives you a permanent advantage.

2. 2

   Three data categories to classify on day one. Public data: data already on the public internet, like company names from SECP or active taxpayer status from FBR's portal. Safe to send anywhere. Internal but non-personal: internal SOPs, vendor lists without contact details, anonymised aggregates. Safe to send to vendors with a processor agreement. Personal or regulated: CNICs, salary data, biometric data, financial transactions, medical records. Most of this should never leave your perimeter in raw form. If it must, the legal basis, the scope, and the retention must be documented before the first prompt is sent. The classification is the policy. The classification done in advance is what makes the AI build defensible.

3. 3

   PII redaction patterns. Before sending a prompt that contains text from your corpus, run a deterministic redactor that removes or masks CNICs, NTNs, account numbers, phone numbers, postal addresses, dates of birth, and biometric references. The redactor is a small piece of code, not an AI. Regex is enough for most cases. The pattern is: the corpus you store keeps the full text, the prompt you send is the redacted text, the answer that comes back is overlaid against the original on display. The model never sees the personal identifiers, the user sees only what they are entitled to see, and the audit log records both states.

The why-tree

Try this with Claude

AI-edge prompt to try with Claude or ChatGPT: 'Draft a DPIA, data protection impact assessment, for a Pakistani bank deploying an LLM-based vendor compliance assistant. Use PDPA-2023 categories. Output as a structured table with risk, mitigation, residual risk, and sign-off owner. Identify the three highest residual risks.' Then sit with your CISO and tear it apart.