Banning AI tools does not work. It moves usage onto personal accounts where you have no visibility at all. This policy assumes people will use AI and makes the boundary explicit: which tools, which data, and how to get something new approved.
Publish an approved list, not a prohibition. Name the tools people may use and what data each may receive. A blanket ban drives usage to personal accounts, which removes your logging, your data processing terms, and your ability to respond to an incident.
Draw the line on data classification, not on task. Trying to enumerate permitted tasks is endless. Map the approved tools to your classification tiers instead: this tool may receive Internal data, that one may not receive Confidential, nothing may receive Restricted.
Give a fast route to approve new tools. AI tooling changes monthly. If requesting approval takes six weeks, people will not ask. A one-page intake and a one-week turnaround keeps the policy connected to reality.
Here's a preview. Unlock the free download to get all 4 sections (3 more below).
Maintain this actively. An out-of-date list is the main reason AI policies get ignored.
| Tool | Approved use | Max data classification | Account type | Training on our data | Owner | Reviewed |
|---|---|---|---|---|---|---|
| Enterprise AI assistant | Drafting, summarising, code assistance | Confidential | Company SSO only | No (contractual opt-out) | IT Director | 2026-09-01 |
| Code completion tool | Code suggestions in the IDE | Internal | Company SSO only | No (enterprise tier) | Engineering lead | 2026-09-01 |
More free templates IT teams use alongside this one.
You cannot govern AI usage you cannot see, and most AI adoption happens through free tiers and browser extensions that never touch procurement. This inventory is built to surface those, which a record of what you bought will not show.
AI risk assessments often reduce to a data protection review, which misses the failure modes specific to these systems: confidently wrong output, drift, prompt injection, and decisions affecting individuals. This template covers those explicitly.
Classification schemes fail when they produce labels without consequences. This one attaches concrete handling rules to each tier (where it may be stored, who may access it, whether it can leave the country), so the label changes what happens.
An approved tools list with the maximum data classification each may handle, concretely prohibited uses, the route to request new tools, human review and disclosure requirements for AI-assisted output, and a named owner with a short review cycle.
Blanket bans generally relocate usage to personal accounts rather than stopping it, which removes logging, contractual protections, and incident response capability. Providing sanctioned tools with clear data boundaries gives better control in practice.
Only into tools covered by an appropriate data processing agreement, with training on your data contractually excluded, and only up to the data classification you have approved for that tool. Restricted data should not go into AI tools at all.
