Privacy Impact Assessment (PIA): AI Transaction Categorization
Project Name: Orbiq
AI Service Provider: Groq Inc. (US-based)
Document Status: Version 1.0 (Official)
Date: April 14, 2026
Primary Reviewer: Privacy Officer (Sumeet Chaurasia)
1. Executive Summary
This Privacy Impact Assessment (PIA) evaluates the privacy risks associated with Orbiq’s use of Groq’s Large Language Models (LLMs) for automated financial transaction categorization. The goal of this system is to enhance user productivity by automatically tagging merchant strings into relevant budget categories (e.g., "Groceries," "Entertainment"). Orbiq employs a "Privacy by Design" approach, utilizing local sanitization and cloud-based Zero Data Retention (ZDR) to mitigate the risks of cross-border data transfer to the United States.
2. System Description and Purpose
Orbiq provides an AI-driven categorization engine that processes transaction data imported by users via CSV or Excel files.
- Purpose: To provide real-time spending insights and pacing metrics.
- Mechanism: Transactions not matched by local user rules are sent to Groq’s API to receive a category suggestion and confidence score.
- Necessity: AI categorization is a core feature required to handle the high volume and variability of merchant descriptions across different banks (BMO, CIBC, RBC, etc.).
3. Data Flow and Data Minimization (PIPEDA Principle 4)
Orbiq implements a strict "local-first" data minimization procedure before any data crosses the Canadian border.
3.1 Local Sanitization Layer
Before the API call to Groq, the Orbiq backend executes a sanitization class that strips the following from the merchant string:
- Direct Identifiers: Full names and specific account numbers.
- Location Data: Specific street addresses or postal codes.
- Exact Financials: Amounts are generalized or rounded to provide context without precision.
3.2 Data Elements Sent to Groq
The final prompt sent to Groq contains only:
- Sanitized merchant keywords (e.g., "WAL-MART" instead of "WAL-MART STORE #1234 TORONTO").
- Transaction type (e.g., "Debit" or "Credit").
- Generalized amount range.
4. Third-Party Risk Assessment (Groq Inc.)
Groq is a US-based entity, which introduces specific regulatory considerations under PIPEDA Section 4.1.3 and the US CLOUD Act.
4.1 Zero Data Retention (ZDR)
Orbiq has explicitly enabled ZDR mode. This ensures that:
- Prompts and outputs are processed in volatile memory (RAM).
- No data is persisted on Groq’s US servers for "abuse monitoring" or "reliability."
- Data is not used to train Groq’s global models.
4.2 Legal Basis and Contractual Protection
- Consent: Users are provided with "Meaningful Disclosure" during onboarding that AI processing occurs in the US.
- DPA: Orbiq has executed a Data Processing Addendum (DPA) with Groq, incorporating Standard Contractual Clauses (SCCs) to ensure a level of protection comparable to Canadian law.
5. Risk Analysis and Mitigations
| Identified Risk | Impact | Mitigation Strategy | Residual Risk |
| Cross-Border Access | US authorities may compel access via the CLOUD Act. | Data Minimization: Sanitized data is de-identified and of low value to foreign authorities. | Low |
| Model Leakage | User spending patterns could enter global AI training sets. | ZDR Enforcement: Contractual and technical toggle to prevent data persistence and training. | Low |
| Inaccurate Categorization | Users may rely on incorrect financial projections. | Disclosure & UI: Clear UI warnings stating AI outputs are "informational only" and allowing manual overrides. | Moderate |
6. Accountability and Governance (PIPEDA Principle 1)
Orbiq’s Privacy Officer remains responsible for this AI integration. This includes:
- Annual Audits: Reviewing Groq’s SOC 2 reports and verifying the ZDR toggle status.
- Individual Access: Users can request an export of the data categories they have accepted or rejected via the standard DSAR channel.
- Withdrawal of Consent: Users may disable AI features at any time, reverting to manual categorization.
7. Conclusion
The AI Transaction Categorization system, as implemented with Groq and local sanitization, provides significant utility while maintaining a robust "Privacy Citadel." By stripping PII locally and enforcing Zero Data Retention in the cloud, Orbiq minimizes the risk of significant harm to users. This system is deemed compliant with PIPEDA’s fair information principles and the OPC’s 2023 guidance on AI.
Approved By: Sumeet Chaurasia — Privacy Officer, FINARO CAPITAL SERVICES INC. (Orbiq)
April 14, 2026