# Data Classification & Handling Policy **Purpose.** Defines the sensitivity tiers your data falls into and the handling rules for each — including the rules that decide what may be pasted into an AI tool. **When to use it.** Early. Classification underpins access control, AI input rules, and most security decisions. Review annually. **How to use it.** Use four tiers at most; every tier beyond that is one people will not remember. The AI handling row in section 2 is the part staff will actually look up — make sure it names specific approved tools, not a category. --- ### 1. Classification tiers | Tier | Definition | Examples | Impact if disclosed | |---|---|---|---| | **Public** | Approved for public release. | Published marketing, annual report | None | | **Internal** | Routine business data. | Internal comms, non-sensitive analytics | Minor | | **Confidential** | Sensitive business or personal data. | Customer records, employee data, contracts | Significant | | **Restricted** | Most sensitive; severe impact. | Special category data, credentials, M&A, security detail | Severe | ### 2. Handling rules | Control | Public | Internal | Confidential | Restricted | |---|---|---|---|---| | Storage | Any approved | Approved systems | Approved + access control | Approved + encryption + logging | | Access | Open | All staff | Role-based, owner-approved | Named individuals, owner-approved | | Sharing externally | Free | NDA | Contract + DPA | Owner + Legal approval | | Email | Permitted | Permitted | Encrypted | Not permitted — secure transfer only | | **AI tool input** | **Permitted** | **Approved tools only** | **Approved enterprise tools with no-training guarantee** | **Prohibited unless explicitly approved for that system** | | Retention | Per schedule | Per schedule | Per schedule + review | Per schedule + documented review | | Disposal | Standard | Standard | Verified | Verified + certificate | > **The AI row is the one that gets breached** > Staff paste customer data into public chatbots because it is convenient > and the rule was not to hand. Publish this row separately, name the > approved tools, and put it where people work. ### 3. Applying a classification 1. The Data Owner assigns the classification at creation or acquisition. 2. Where a dataset holds mixed tiers, **the highest tier applies to the whole set** unless it is separated. 3. Classification is recorded in the Data Asset Register. 4. Derived data — including model outputs and embeddings — inherits the classification of its most sensitive input unless a documented assessment lowers it. > **Derived data inherits** > Embeddings and model outputs derived from Confidential data are > Confidential. This is routinely missed, and it is how sensitive data ends > up in an unprotected vector store. ### 4. Special category & regulated data Special category personal data, criminal offence data, children's data, and payment card data carry additional requirements — see Processing & DPIA Log. Default these to **Restricted** unless assessed otherwise. ### 5. Labelling Where the platform supports it, apply the label in the tool. Where it does not, record classification in the register. **[Specify your labelling mechanism.]** ### 6. Review | Version | Date | Owner | Approved by | |---|---|---|---| | 0.1 | [date] | [role] | [committee] | --- ## Adaptation notes - **Small organisations:** Three tiers — Public, Internal, Confidential — is usually enough. Keep the AI input row regardless; it is the highest-value line in the policy. - **Existing security classification:** Do not create a competing scheme. Map onto what security already uses and add only the AI handling row. - **Public sector:** Align to your national scheme (e.g. OFFICIAL / SECRET) rather than these labels, and add the FOI disclosure consideration to each tier. --- *From the [Open Data & AI Governance Kit](https://lsdeva.github.io/governance-kit/). Licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — free to use, adapt, and share with attribution.* ***Not legal advice.** Adapt to your jurisdiction, sector, and risk appetite, and have qualified counsel review anything material.*