# AI Data Classification Handling Matrix Worksheet > Source: https://netray.co/tools/ai-data-classification-handling-matrix-worksheet > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Free Tools ## Short answer Decide per data class whether it may be sent to a hosted LLM, an on-prem model or no model, and who approves exceptions. ## Category compliance-legal ## Audience - CIO - Security officer - IT manager - Compliance officer ## Tool **Kind:** worksheet **Mode:** editable **Columns:** #### Data class **Key:** class **Type:** text #### ERP examples **Key:** examples **Type:** text #### Hosted LLM API **Key:** cloud **Type:** select **Options:** - Allowed - Allowed with masking - Not allowed - Exception only #### On-prem model **Key:** onprem **Type:** select **Options:** - Allowed - Allowed with masking - Not allowed - Exception only #### Masking or control **Key:** masking **Type:** text #### Exception approver **Key:** approver **Type:** text #### Approx. users affected **Key:** users **Type:** number **Rows:** **Class:** Public **Examples:** Published catalogs, public specs, marketing copy **Cloud:** Allowed **Onprem:** Allowed **Masking:** None **Approver:** Not needed **Users:** 120 **Class:** Internal **Examples:** Inventory levels, routings, work instructions **Cloud:** Allowed with masking **Onprem:** Allowed **Masking:** Remove customer names and prices before sending **Approver:** IT manager **Users:** 85 **Class:** Confidential commercial **Examples:** Customer pricing, supplier contracts, margins **Cloud:** Not allowed **Onprem:** Allowed **Masking:** Role-based access and logging **Approver:** CFO **Users:** 30 **Class:** Personal data **Examples:** Employee records, contact details, payroll **Cloud:** Not allowed **Onprem:** Allowed with masking **Masking:** Pseudonymize ids; exclude salary fields **Approver:** Data protection lead **Users:** 12 **Class:** Export controlled **Examples:** ITAR or EAR technical data, controlled drawings **Cloud:** Not allowed **Onprem:** Exception only **Masking:** US-person-only environment, no external network **Approver:** Empowered official **Users:** 8 **Class:** Controlled unclassified (CUI) **Examples:** DoD contract technical data marked CUI **Cloud:** Not allowed **Onprem:** Exception only **Masking:** Authorized enclave and access logging **Approver:** Security officer **Users:** 10 **Totals:** #### Total users in scope **Key:** users **Agg:** sum **Format:** integer #### Data classes **Key:** class **Agg:** count **Format:** integer **Notes:** Replace the sample rows with your own classification scheme and regulatory duties. Defense, export control and personal data rules depend on your jurisdiction and contracts, so have your compliance owner confirm the decisions before publishing the matrix to users. ## Content **Intro:** Most AI data incidents start with a simple question: can I paste this into the chatbot? This worksheet turns your data classification scheme into a clear answer for hosted APIs, on-prem models or neither. **Sections:** #### How to use it **Body:** Start from your existing classification levels or the sample rows. For each class, list real ERP examples so users can recognize them, then set what is allowed for a hosted LLM and for a model you run yourself. Add the masking control and the person who can approve exceptions. Export it to PDF and attach it to your AI acceptable use policy so employees have a one-page reference. #### Reading the matrix **Body:** The usual pattern is that public data may go anywhere, internal data goes to hosted services only after masking, and commercially sensitive or regulated data stays on infrastructure you control. Export-controlled and CUI data usually need an even stricter environment, with access limited to authorized people and no external network path. The exception column keeps decisions accountable rather than ad hoc. **Bullets:** - Name the owner of each class - Give users concrete examples, not definitions - Define masking so it is repeatable - Review after any new tool or contract #### Common mistakes **Body:** Policies often say do not paste sensitive data without defining sensitive. Another gap is ignoring data that leaves indirectly through plugins, browser extensions and meeting transcription tools. Teams also forget that on-prem does not equal permitted: the model still needs access control and logging. Finally, a matrix nobody reviews goes stale when ERP modules or contracts change. **Cta:** Netray runs private models on your own infrastructure so regulated data never needs to reach a hosted API. We can help map your classes to an architecture. **Faq:** #### Can customer pricing go to a hosted LLM if the vendor promises no training? **Answer:** That depends on your contracts and risk appetite. A no-training promise addresses one risk but not retention, access by vendor staff or legal obligations to customers. Many companies still keep it on controlled infrastructure. #### What does masking mean in practice? **Answer:** Replacing identifying values, such as names, ids and prices, with tokens before text leaves your environment, then mapping them back in the answer. It must be automated, tested and logged to be reliable. #### Is on-prem always allowed for personal data? **Answer:** Not automatically. You still need a lawful purpose, minimization, access control and retention rules. On-prem reduces transfer risk but does not remove data protection duties. #### Who should own this matrix? **Answer:** Usually the security or data protection function, with input from IT, legal and the business. A named owner and a review date keep it current. ## Search terms ai data classification matrix, what data can go into chatgpt policy, llm data handling rules template, data classification for generative ai, cui itar data ai policy, on prem vs cloud llm data rules, ai acceptable data table excel --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json