Check Data Privacy, Compliance, and Access
Key Takeaways
- Domain III Task 6 is the per-dataset clearance inside CPMAI Phase 2; the project's overall privacy and security plan belongs to Domain I Task 1.
- Vendor licenses that permit internal analytics frequently forbid using the same data to train or fine-tune a model — analytics rights and training rights are separate rights.
- Least privilege must be applied to the assembled dataset in the AI workspace, because an extract leaves the access controls of the system of record behind.
- A privacy impact assessment is scoped to the specific use; repointing a cleared dataset at a different model requires a new assessment, not a footnote.
- Lineage must let an auditor trace one training record back to its source, its transformations, and the permission that allowed its use.
Permission Is a Separate Question from Availability
Domain III, Identify Data Needs, carries 26% of the PMI-CPMAI exam, and its sixth task — check data privacy, compliance, and access — decides whether the data you found in Task 3 and gathered in Task 5 may lawfully become training data. This work sits in CPMAI Phase 2, Data Understanding. Domain I Task 1 owns the project's overall privacy and security plan; this task is the per-dataset clearance that applies that plan to the exact records, for the exact purpose, in this project.
The distinction the exam tests is easy to state and easy to lose under time pressure: having data is not the same as being permitted to use it for this purpose. A customer record collected lawfully, retained lawfully, and used every day for billing may still be off limits as training data for a churn model, because the basis that justified its collection never extended to that use.
Verify Data Usage Rights and Licensing Agreements
Work through four sources of restriction for every dataset in the inventory.
- Consent scope and purpose limitation. Read what the customer was actually told. Consent given for service delivery does not automatically cover profiling, model training, or secondary analytics. Purpose limitation means the permitted purpose travels with the record; a new purpose needs a new basis.
- Third-party license terms. Vendor feeds routinely permit internal reporting and analysis while explicitly forbidding use of the data to train, fine-tune, or improve a machine learning model, or forbidding any derived work that outlives the subscription. Analytics rights and training rights are different rights.
- Contractual restrictions on customer data. Master service agreements and data processing agreements often bind you as a processor acting only on documented instructions. Enterprise customers frequently negotiate clauses that bar their data from any shared or cross-client model.
- Open-data license conditions. Open is not unconditional. Attribution, share-alike, non-commercial, and no-redistribution terms all attach obligations to the model or product built on that data.
Ensure Compliance with Data Protection Regulations and Policies
The Exam Content Outline names GDPR (General Data Protection Regulation), CCPA (California Consumer Privacy Act), and evolving AI regulations and industry standards, plus your organization's own policies — which are often stricter than the statute. Three checks bite at the data-acquisition stage rather than at deployment:
- Lawful basis and sensitive categories. Establish which basis covers this processing and whether the set contains special-category data (health, biometric, racial or ethnic origin, union membership) that carries extra conditions. Fields you do not need are a liability, not an asset.
- Cross-border transfer limits. Records collected in one jurisdiction may not be freely copied to a cloud region in another. Establish, per dataset, whether a transfer is permitted at all and under which approved mechanism, and record that answer against the source in the inventory before the extract is scheduled.
- Retention limits. Data scheduled for deletion after a set period cannot quietly live forever inside a training corpus, and deletion requests must reach every copy, including extracts held for modeling.
Implement Access Controls and User Permissions for Data Resources
Apply least privilege to the assembled dataset, not only to the source systems. The moment an extract lands in the AI workspace it leaves the protection of the system of record: the claims platform may have restricted a table to six adjusters, while the extract is readable by fourteen people including two contractors. Controls that belong on the working copy include a named access list with a documented business need, role-based permissions separating read from export, masking or tokenization of direct identifiers that modeling does not require, separation of raw and prepared zones, time-boxed access that expires with the phase, and access logging you can produce on request.
Conduct Privacy Impact Assessments for Data Usage
What a privacy impact assessment (PIA) — a DPIA (data protection impact assessment) under GDPR — must contain and who signs it is set out under Domain I Task 1, with the project's privacy and security plan. What Task 6 adds is scope discipline. The assessment is written for this dataset and this purpose, so it is re-run whenever either changes: repointing a cleared fraud dataset at a marketing model is a new assessment, not a footnote on the old one. Until it is signed, the dataset stays outside the training set.
Track that outcome as a per-dataset clearance status in the source inventory built in Task 3 — cleared, cleared with conditions, pending determination, or refused — each with a date and the name of the person who decided. A dataset carrying no recorded status is not cleared; it is unexamined, and from inside a pipeline the two look identical.
Document Data Lineage and Usage for Audit Purposes
Record source, transformations, and destination for every field that reaches the training set, alongside the permission basis for each source. The test is whether an auditor — internal, regulatory, or a customer exercising a contractual right — can take one training record and trace it back to the system it came from, the rule that allowed its use, and every step that changed it. Provenance answers whose data is this and under what right; lineage answers what happened to it on the way here. Both belong in the Domain I audit trail, and both are cheap to capture during Phase 2 and expensive to reconstruct in Phase 6.
| Check | Question you must answer | Evidence | Who signs off |
|---|---|---|---|
| Usage rights | Does the consent or contract permit this purpose, including model training? | Consent text, MSA/DPA clause, vendor license | Legal counsel or contract owner |
| Regulatory basis | What lawful basis covers this processing, and does the set contain special-category data? | Record of processing, lawful-basis note | Privacy lead or data protection officer |
| Cross-border transfer | May these records leave the jurisdiction where they were collected? | Residency rule, approved transfer mechanism | Privacy lead plus infrastructure owner |
| Retention | Is any record past its retention limit or subject to a deletion request? | Retention schedule, deletion log | Data owner for the system of record |
| Access control | Who can read the assembled dataset, and is that the minimum needed? | Named access list, role definitions, masking spec | Data owner plus workspace administrator |
| Privacy impact | What is the residual privacy risk of this specific use, and what mitigates it? | Completed PIA or DPIA with mitigations | Privacy lead; sponsor accepts residual risk |
| Lineage | Can an auditor trace a training record to its source and its permission basis? | Lineage record: source, transformations, destination | Project manager as accountable owner |
The Exam Behavior: Stop and Get It in Writing
When data is available but its permitted use is unclear, the credited action is to hold that dataset at the boundary of the training set and obtain a written determination — from legal, the contract owner, the vendor, or the privacy lead — before it is ingested. Not a verbal reassurance, not a sponsor's risk acceptance (a sponsor cannot grant rights a contract withholds), and never the reasoning the exam punishes most: it is already in our data lake, so we may use it. Removing a source later means retraining, and in some cases retiring, a model built on it.
A purchased market-data feed is licensed for “internal reporting and analysis” and says nothing about machine learning. The data science team wants to include it in the training set for a demand-forecasting model, and the sprint starts Monday. What should the project manager do first?
A claims dataset has been extracted into the project's AI workspace. In the source platform the table was restricted to six named adjusters; in the workspace, all fourteen project members, including two contractors, can read it in full. What is the project manager's most appropriate action?
During an internal audit, the reviewer selects one record from the model's training set and asks which system it came from, what was done to it, and under what permission it was used. The team can produce the final training table but cannot answer any of the three questions. What should the project manager do?