ISO 27001:2022 Annex A 5.34 Privacy and protection of PII for AI Companies

ISO 27001 Annex A 5.34 is a security control that mandates the identification and implementation of requirements for personal data protection. This preventive implementation requirement ensures legal compliance through rigorous data mapping, providing the business benefit of enhanced stakeholder trust and protection against membership inference attacks in AI models.

For an innovative AI company like yours, managing Personally Identifiable Information (PII) is not just a standard compliance task. It is a core component of building trust with your customers and partners, enabling you to innovate responsibly. The international standard for information security, ISO 27001, provides a clear framework for this crucial activity in ISO 27001 Annex A 5.34 Privacy and protection of PII.

The primary purpose of Control 5.34 is to ensure your organisation meets its legal, statutory, regulatory, and contractual requirements for the preservation, privacy, and protection of PII. It is a preventive control, designed to help you create clear guidelines and procedures that maintain risk at an acceptable level and safeguard the personal data you handle.

The “No-BS” Translation: Decoding the Requirement

Let’s strip away the consultant-speak. Annex A 5.34 is about ensuring you don’t accidentally train your model on your customer’s credit card numbers.

The Auditor’s View (ISO 27001)	The AI Company View (Reality)
“The organisation shall identify and meet the requirements regarding the preservation of privacy and protection of PII according to applicable laws and regulations and contractual clauses.”	Know where the PII is. Don’t just dump a CSV into S3 and forget about it. Scan it for emails, names, and phone numbers. If you find them, either delete them or encrypt them.
“Privacy impact assessments.”	Check before you train. Before you start a new training run, ask: “Is there any personal data in this dataset?” If yes, anonymise it.

Shutterstock

The Business Case: Why This Actually Matters for AI Companies

Why should a founder care about “PII Protection”? Because regurgitating PII is the fastest way to get sued.

The Sales Angle

Enterprise clients will ask: “How do you prevent our customer PII from leaking into your model responses?” If your answer is “We rely on the model’s safety filters,” you lose. If your answer is “We use a dedicated PII Redaction Layer (e.g., Presidio) before data hits the model, and we contractually guarantee zero-retention for inference,” you win the deal. A 5.34 is your privacy architecture.

The Risk Angle

The “Membership Inference” Attack: Researchers can query your model to find out if a specific person was in the training set (e.g., “Was John Smith treated for cancer?”). If your model reveals this, you have breached GDPR/HIPAA. A 5.34 forces you to de-identify data before training to prevent this.

DORA, NIS2 and AI Regulation: Privacy By Design

Regulators demand “Privacy by Design,” not just a privacy policy.

• GDPR (Article 25): Requires “Data Protection by Design and Default.” You must implement technical measures (like pseudonymisation) to strip PII from training data.
• EU AI Act: High-risk AI systems must have “data governance” that includes examination for possible biases. Often, bias correction requires processing sensitive data (e.g., race). You need a specific legal basis and strict controls (A 5.34) to do this lawfully.
• DORA: Financial entities must protect client data. If your AI processes transaction data, you must treat it with the same security as a bank vault.

ISO 27001 Toolkit vs SaaS Platforms: The Privacy Trap

SaaS platforms help you generate a privacy policy, but they don’t help you engineer privacy. Here is why the ISO 27001 Toolkit is superior.

Feature	ISO 27001 Toolkit (Hightable.io)	Online SaaS Platform
The Method	DPIA Templates. A rigorous “Data Protection Impact Assessment” template that forces you to analyse AI specific risks.	Checkbox Policy. Platforms give you a generic “Privacy Policy” for your website, but not the technical procedure for sanitising training data.
Ownership	Your Assessment. You keep the DPIA document. It proves you thought about the risk.	Black Box. The platform says “Compliant” because you ticked a box, but you have no evidence of the actual risk analysis to show a regulator.
Simplicity	Data Mapping. A spreadsheet to map PII flows. “Data enters via API -> Redacted by Lambda -> Stored in S3.”	Automated Noise. Data discovery tools often flag every string of 16 digits as a credit card, drowning you in false positives.
Cost	One-off fee. Pay once. Protect privacy forever.	Per-Data Source Cost. Privacy modules often charge by the number of databases connected.

The AI Magnifying Glass: Analysing Unique PII Risks in Your Workflow

While Control 5.34 applies to all businesses, your AI-driven workflows introduce unique risks.

Exposure of Sensitive Training Datasets

A breach of training data could lead to massive unauthorised disclosure. More advanced threats include model inversion attacks, where attackers reconstruct sensitive PII from a model’s outputs. Protecting the training set is as important as protecting the production database.

Disruption of Algorithmic Processes

If the integrity of PII is compromised (e.g., medical records altered), your diagnostic AI could make life-threatening errors. A failure to protect this data directly impacts business continuity and stakeholder trust.

Vulnerabilities in the AI Supply Chain

Your AI supply chain includes data labellers and cloud providers. You must ensure every partner handling PII complies with the same rigorous data protection standards. A weak link in the labelling team exposes your data.

Your Compliance Blueprint: Actionable Steps for PII Protection

Complying with Control 5.34 requires a structured approach.

Establish Clear Governance and Responsibility

Appoint a Privacy Officer (this can be a shared role in startups). Identify all applicable data protection laws (GDPR, CCPA) and document them in your Legal Register (A 5.31).

Develop Topic-Specific Policies and Procedures

Create a Privacy Policy that covers PII Classification, Secure Handling, and Data Retention. Specifically address PII in MLOps: how do you manage PII across development, testing, and production environments?

Implement Technical and Organisational Measures

Implement strong encryption (AES-256) for PII at rest. Use data masking or synthetic data in non-production environments. Never use live customer PII for testing code.

The Evidence Locker: What the Auditor Needs to See

When the audit comes, prepare these artifacts:

• Data Processing Agreement (DPA): Signed agreements with your customers and your vendors.
• Record of Processing Activities (ROPA): An Excel sheet listing what PII you have, why you have it, and where it goes.
• DPIA Reports (PDF): Completed impact assessments for high-risk processing (e.g., training a new model on user data).
• Subject Access Request Log: A log of user requests (e.g., “Delete my data”) and evidence that you fulfilled them within 30 days.

Common Pitfalls & Auditor Traps

Here are the top 3 ways AI companies fail this control:

• The “Dev DB” Leak: Developers took a dump of the production database (with real PII) to test a new feature on their laptop. Instant GDPR violation. Use synthetic data.
• The “Log File” Oversight: You encrypt the database, but your application logs print out the PII in plain text for debugging. The auditor checks the logs and finds email addresses. Fail.
• The “Forever” Retention: You keep user data indefinitely “to improve the model.” You have no legal basis for this. You must delete data when the user leaves.

Handling Exceptions: The “Break Glass” Protocol

Sometimes you need to access raw PII to debug a critical error.

The PII Access Workflow:

• Trigger: Production incident requiring inspection of specific user record.
• Approval: DPO / Privacy Lead approves temporary access.
• Access: Access granted via Just-In-Time (JIT) privilege for 1 hour.
• Log: Access is logged and reviewed. Any data downloaded is deleted immediately after debugging.

The Process Layer: “The Standard Operating Procedure (SOP)”

How to operationalise A 5.34 using your existing stack (Linear, Excel).

• Step 1: Mapping (Manual). Use the High Table ROPA Template to map data flows.
• Step 2: Assessment (Manual). Before a new feature launch, Product Manager completes a “Privacy Checklist” ticket in Linear.
• Step 3: Redaction (Automated). Configure libraries (e.g., Microsoft Presidio) to automatically redact PII from logs and training pipelines.
• Step 4: Deletion (Automated). Set up cron jobs to delete user data 30 days after account termination.

By implementing a structured compliance programme, you can turn a complex obligation into a source of competitive advantage. Leveraging expert tools like the High Table ISO 27001 Toolkit provides the definitive path to achieving this.

ISO 27001:2022 Annex A 5.34 for AI Companies FAQ

About the author