Secure PII analytics under GDPR, CCPA/CPRA, and PCI DSS comes down to four questions: which regulations the exact service covers, whether a signed DPA or BAA exists, whether PII is copied or queried in place, and what an AI assistant actually sends to a model.
Quick Summary (TL;DR)
- GDPR requires a binding processor contract under Article 28, not just a “GDPR ready” marketing page.
- CCPA/CPRA obligations differ depending on whether the vendor is a business, service provider, contractor, or other recipient.
- PCI DSS posture varies sharply: Tableau Cloud and Amazon QuickSight/Quick publish PCI DSS support, while Domo and Qlik Cloud both state they are not PCI DSS compliant/certified.
- “HIPAA compliant” and “HIPAA certified” are not real product properties. HHS does not certify products, so the accurate claim is whether a vendor will sign a BAA.
- NLQ assistants often send more than the question itself: Power BI/Fabric Copilot grounding can include schema data and query results sent to Azure OpenAI; Qlik sends dataset metadata to Anthropic via AWS Bedrock.
- “Data stays in your VPC” and “the AI runs in your VPC” are different claims, private connectivity does not mean inference moved into the customer’s environment.
- In a self-hosted, on-premises, or private-cloud deployment, Knowi states its Private AI processes prompts and queries inside the customer’s environment without an external AI API call.
Why “Secure Analytics Platform” Is Not a Single Checkbox
A BI platform can hold a current SOC 2 report, offer a signed DPA, and let a customer pick a data region, and still send part of a query to a separate AI service the buyer never explicitly selected. A self-hosted product can keep the core dataset inside a customer’s network and still require an external LLM API for its AI assistant. Treating “secure PII analytics” as one property misses both cases.
The Four Layers Buyers Actually Need to Separate
Regulatory coverage: does the exact service, not the vendor’s broader cloud portfolio, document GDPR, CCPA/CPRA, and PCI DSS separately. Contractual instrument: whether a Data Processing Agreement (DPA) is automatically incorporated, and whether a Business Associate Agreement (BAA) is available where health data is in scope.
Data movement: whether PII must be copied into a warehouse or cache before analysis, or the platform can query the source directly. AI/NLQ payload: whether a question, schema, sample values, retrieved rows, or generated answers reach a separate model, and who operates it.
Where This Article Stops: HIPAA and Healthcare PHI
This piece focuses on general enterprise PII under GDPR, CCPA/CPRA, and PCI DSS for fintech, ecommerce, and B2B SaaS teams, not protected health information. For architecture questions specific to PHI, a dedicated comparison of AI architectures for PHI covers that ground already.
GDPR, CCPA/CPRA, and PCI DSS: What Each Actually Requires
Buyers often ask whether a platform is “GDPR compliant” as if that were one yes-or-no fact. The regulations ask more specific questions, each applying to a different part of the vendor relationship.
GDPR: Processor Contracts, Not Just a Privacy Page
Where a vendor processes personal data on a customer’s behalf, GDPR Article 28 requires a binding contract covering that processing, and Chapter V regulates transfers outside the EU. Publishing a “GDPR ready” page does not substitute for that processor arrangement, or answer where the specific service transfers data during query execution or AI processing.
CCPA/CPRA: Service Provider and Contractor Obligations
California law defines distinct roles, business, service provider, contractor, and other recipient, each carrying different obligations. Recent California Privacy Protection Agency enforcement has focused on missing contractual privacy protections, making contract language, not a trust-page badge, the thing worth reviewing before purchase.
PCI DSS: Card Data Scope Varies Sharply by Vendor
PCI DSS requirements apply based on a system’s actual role in the cardholder-data environment, and its future-dated requirements became effective March 31, 2025. Vendor posture here is unusually inconsistent and worth checking individually, not assumed from a vendor’s other certifications.
| Vendor | Public PCI DSS posture (documentation reviewed August 2026) |
|---|---|
| Tableau Cloud | Current security material states PCI DSS 4.0 support. |
| Amazon QuickSight / Quick | AWS lists PCI DSS among QuickSight’s compliance posture, carried forward into Quick Suite. |
| Domo | States Domo is not PCI DSS compliant; customers should not upload payment-card or financial-account information. |
| Qlik Cloud | Not PCI DSS certified; customers should not store PCI DSS data there, per Qlik’s privacy FAQ. |
| ThoughtSpot | States it does not store or process cardholder data, so it does not qualify as a PCI processor, merchant, or service provider. |
| Power BI, Looker, Sisense, Knowi | No specific PCI DSS attestation surfaced for the platform itself. Unverified until confirmed directly with the vendor. |
For a payments team, that table does more work than a single “compliant” badge: three of the six positions above are explicit exclusions or non-certifications, not research gaps.
Where NLQ Prompts and Query Results Actually Travel
The sharpest 2026 differentiator for PII analytics is not the deployment diagram, it is what a natural-language question actually sends to a model and who operates that model.
Vendor-Managed Cloud Models
Microsoft documents that Power BI/Fabric Copilot builds grounding from relevant service data, sends the final prompt plus that grounding to Microsoft-managed Azure OpenAI, and in some summarization flows sends evaluated query results back to Azure OpenAI. Microsoft says this is Azure-hosted OpenAI rather than the public OpenAI or ChatGPT service, is not used to train models, and stays in the capacity’s region unless an administrator enables cross-region processing. Gemini in Looker similarly processes prompts and governed Looker context under Google Cloud contract terms, with outputs excluded from training.
Bring-Your-Own and Regional Model Routing
Qlik documents that its GenAI assistant sends prompts plus dataset names, column types, and relationships to an Anthropic model hosted through AWS Bedrock in the tenant’s region, not used for training. Sisense is unusually explicit that payloads vary by feature: its Assistant sends prompts and semantic metadata, narrative features can send aggregated results, and enrichment features can send sample values, depending on the managed Azure-hosted LLM or a customer-configured provider in use. Tableau Server, from version 2025.3 with expanded provider support in 2026.2, lets a customer connect Tableau Agent to a provider of its choosing, routing around the Einstein Trust Layer used on Tableau Cloud.
Self-Hosted and On-Prem Model Paths
Fewer products document a path where AI processing never leaves the customer’s own infrastructure. In a self-hosted, on-premises, or private-cloud deployment, Knowi states its Private AI processes prompts and query execution inside the customer’s environment without calling an external AI API. Its managed-cloud offering instead runs the vendor’s own models on Knowi-managed infrastructure with no third-party LLM seeing queries or results, a narrower assurance than customer-environment processing. A closer look at self-hosted LLMs for regulated teams, including a dedicated fintech and banking section, is already published on this site. Sisense also documents a self-hosted GenAI option, described as beta, and its legacy Simply Ask feature does not send data to an LLM at all.
Five Architectural Patterns for Querying PII Safely
Rather than a vendor-versus-vendor security table, which an earlier deployment-by-deployment comparison of VPC and public-cloud analytics already covers for Tableau, Power BI, Looker, and Knowi, the more useful comparison for PII buyers groups by architectural pattern: how the AI and query path is built, independent of which brand ships it. Most vendors offer more than one pattern depending on the plan or deployment selected.
| Pattern | Where query context travels | Regulatory exposure | Operational cost |
|---|---|---|---|
| Public cloud LLM, standard terms | Prompt, often schema or evaluated results, reach the vendor’s managed model (Azure OpenAI behind Power BI Copilot; Anthropic via Bedrock behind Qlik’s assistant). | Model operator is a subprocessor under the vendor’s DPA; transfer terms and training-use language need review. | Lowest, usually included in the subscription. |
| Public cloud LLM, zero-retention/no-training terms | Same payload, but the vendor contractually states prompts and outputs are not used for training (Microsoft, AWS). | Narrower than an undocumented arrangement, but still a live subprocessor; no-training is a data-use limit, not residency. | Low, often gated behind a specific tier. |
| Bring-your-own LLM (BYO) | Whatever the customer configures; prompts and, for some features, metadata move to the customer-selected provider (Sisense; Tableau Server 2025.3+/2026.2). | Customer takes on that provider’s DPA and region; the vendor’s own posture no longer fully describes the AI leg. | Moderate, a second vendor relationship to manage. |
| Tokenization or masking before the AI layer | Only masked or tokenized values reach the model; raw PII stays behind the masking layer. | Reduces how much of the payload qualifies as personal data, but masking coverage and re-identification risk become the real question. | Engineering to build and maintain masking rules, plus accuracy tradeoffs. |
| In-VPC or self-hosted inference | Stays inside the customer’s network per the vendor’s stated architecture, the pattern Knowi’s self-hosted/on-prem Private AI and Sisense’s beta self-hosted GenAI both target. | Removes the AI step as a separate subprocessor, since inference runs on infrastructure the customer already controls. | Highest, customer owns model hosting and capacity. |
| Direct query, no data movement | The query executes against the source, returning results live rather than duplicating records first (Power BI DirectQuery; direct SQL/NoSQL/API querying). | Reduces persistent PII copies in vendor storage, but doesn’t answer the AI-payload question above, the two axes are independent. | Low storage cost; performance depends on source capacity. |
A platform’s AI pattern and its data-movement pattern are separate decisions. A source-direct product can still route AI through a public cloud LLM, and an Import-based product can still ship its AI assistant off by default.
A Procurement Checklist for PII Analytics Vendors
These distinctions convert into direct questions worth putting to any vendor in writing, since verbal reassurance on a call is not a substitute for the contract.
Questions to Put to Every Vendor in Writing
- Will you sign your standard DPA for the exact service and plan we are purchasing, with a current subprocessor list?
- Is this exact service currently in your PCI DSS scope, or do your terms prohibit cardholder data?
- For each NLQ or generative-AI feature, which fields transmit: prompt, schema, sample values, retrieved records, or final results?
- Who operates that model, and is it a subprocessor under our DPA?
- Can generative processing leave our selected data region, and what setting controls that?
- Can we disable generative AI while keeping conventional dashboards and query functionality?
- What is persistently copied into your platform, what is only cached, and how are deletion requests handled?
- Do row-level security and column permissions apply identically to AI-generated queries, not just dashboards?
Under CCPA/CPRA, ask whether the vendor’s role is business, service provider, contractor, or another category, since the obligations differ and recent enforcement has focused on missing contract language, not missing marketing claims.
Where Knowi Fits Best
Knowi’s clearest architectural difference for this buyer is source-direct querying: current documentation describes native SQL, MongoDB, Elasticsearch, and REST API connectivity with cross-source joins, without a proprietary warehouse as a prerequisite, and results held in a configurable temporary cache rather than a permanent copy. That fits fintech, ecommerce, and B2B SaaS teams whose customer PII sits across a relational database, a document store such as MongoDB, and internal or partner APIs, the combination covered in a comparison of BI tools for MongoDB and in analytics over REST API data.
On compliance, Knowi currently publishes SOC 2 Type II status and states a BAA is available, narrower and more accurate than “HIPAA certified,” since HHS does not certify products. GDPR controls such as data minimization and right-to-erasure support are documented, but a public DPA execution page, CCPA/CPRA posture, and PCI DSS status were not established in current research and should be raised directly rather than assumed. On AI, the self-hosted/on-prem/private-cloud Private AI claim, no external AI API call, applies specifically to that deployment; the managed-cloud offering is a narrower, separate assurance, that Knowi’s own models run on Knowi-managed infrastructure with no third-party LLM seeing queries or results.
Teams already standardized on a hyperscaler’s compliance program, an AWS shop on QuickSight/Quick or a Microsoft shop that has reviewed Azure OpenAI’s Copilot terms, may find that vendor’s documentation more complete on PCI DSS and CCPA/CPRA today, where Knowi’s public material is thinner. The trade-off: connector-level access to NoSQL and API sources without ETL is documented in more detail than non-HIPAA regulatory posture, so a payments or EU-data-heavy buyer should verify GDPR terms, CCPA role, and PCI scope directly before treating any pattern above as settled.
Analyze Regulated PII Without Routing It Through a Third-Party LLM.
Knowi queries your SQL, NoSQL, and API sources directly, so PII doesn’t have to move into a separate warehouse to be analyzed. Deploy in the cloud, or self-host so query processing and Private AI stay inside your network.
What you can do with Knowi:
- Query SQL, MongoDB, Elasticsearch, and REST APIs directly, without a mandatory ETL warehouse.
- Deploy self-hosted or on-premises so Private AI processes prompts and queries inside your environment.
- Enforce row-level security, RBAC, and SSO consistently across dashboards and AI-generated queries.
- Keep query results in a configurable temporary cache instead of a permanent copy.
- Review SOC 2 Type II reporting and BAA availability directly with our team.
- Ask questions in natural language across every connected source.
Used by fintech, ecommerce, and B2B SaaS teams that need to query customer PII where it lives, under the deployment model their compliance review requires.
Frequently Asked Questions
Does my analytics vendor need to sign a DPA for GDPR?
Where a vendor processes personal data on your behalf, GDPR Article 28 requires a binding contract governing that processing, so a signed DPA is the legal instrument to check for, not a general privacy statement. Ask for the DPA covering the exact service and plan you are purchasing, plus its current subprocessor list.
Is Power BI Copilot GDPR compliant, and does it send my data to OpenAI?
Microsoft documents that Power BI/Fabric Copilot sends the final prompt plus grounding data to Microsoft-managed Azure OpenAI, not the public OpenAI or ChatGPT service, and that this is not used to train underlying models. Its Products and Services DPA covers the contractual GDPR terms; whether that satisfies your requirements depends on the data involved and the region setting in use.
Is Tableau Cloud or Domo PCI DSS compliant for payment card data?
Tableau Cloud’s current security material states PCI DSS 4.0 support, while Domo’s general Service Agreement states Domo is not PCI DSS compliant and customers should not upload payment-card or financial-account information. Qlik Cloud takes a similar position, stating it is not PCI DSS certified, so posture should be checked vendor by vendor.
Does CCPA require my analytics vendor to be a service provider or contractor?
California law defines distinct roles, business, service provider, contractor, and other recipient, each with different obligations, so the correct question is which role the vendor occupies in your contract. Recent California Privacy Protection Agency enforcement has focused specifically on missing contractual privacy protections rather than missing public statements.
What PII does a natural-language BI assistant actually send to an LLM?
It depends on the vendor and feature: Qlik documents sending dataset names, column types, and relationship metadata to Anthropic through AWS Bedrock, while Sisense documents that some features send aggregated results or sample column values depending on which GenAI feature is used. Ask each vendor for the exact field list per AI feature rather than a general assurance.
Does keeping my database in a VPC stop my BI tool from sending query results to an AI service?
Not by itself. Private connectivity, such as AWS PrivateLink or a VPC connection into a source database, can keep network traffic off the public internet while the analytics service and its AI features remain vendor-managed and still call an external model, different from an in-VPC or self-hosted deployment where inference runs on infrastructure the customer controls.