# Startup AI Governance Checklist and Policy Templates The lightweight set for a company under 50 people: a simple AI inventory, a one-page acceptable use policy, a three-clause vendor check, a data flow summary, and a template for answering a customer's AI governance questionnaire. **Who this is for:** The founder, ops lead, or first compliance hire at a small company that needs enough governance to pass diligence and not get burned. Source playbook: https://aigovernance.com/playbook/ai-governance-checklist-startups --- ## Startup AI system inventory _The minimum viable inventory. Five columns, kept in a shared sheet._ ### Template | System | Purpose | Vendor | Data processed | Owner | |---|---|---|---|---| | | | | | | ### Worked example | System | Purpose | Vendor | Data processed | Owner | |---|---|---|---|---| | Support assistant | draft support replies | Vendor API | customer emails, order info | Head of Support | | Code assistant | dev productivity | Vendor IDE | source code (no prod secrets) | CTO | | Sales email tool | outreach drafting | Vendor SaaS | prospect names, public company info | Head of Sales | | Analytics summaries | dashboard narratives | embedded in BI tool | aggregated product metrics | Data | ### Acceptance criteria - Every AI tool in use has a row, including embedded features and free-tier use. - Each row names an owner. - The sheet is reviewed at least quarterly, even if briefly. --- ## One-page acceptable use policy _The whole AI policy for a small company. One page._ ### Template > One page. Everyone reads it at onboarding. - **Approved tools:** [list] - **Never put in an AI tool:** customer personal data, secrets or credentials, unreleased financials, anything a customer contract prohibits - **Always:** treat output as a draft and check it; do not paste large blocks of proprietary code into third-party tools - **New tool:** ask [person] before using it for work - **Problems:** tell [person] if an AI tool causes an error or exposes data; no blame for speaking up ### Worked example "Approved: [Assistant], [Code tool] (enterprise plan). Never enter customer personal data, passwords, API keys, or unreleased financials into any AI tool, including personal accounts. Treat AI output as a draft and verify it. Want a new tool? Ask [CTO] first. If an AI tool causes a mistake or leaks something, tell [CTO] right away; you will not be blamed for reporting it." ### Acceptance criteria - It fits on one page and lists approved tools and forbidden data categories. - Everyone acknowledges it at onboarding. - It names one person as the approval and escalation point. --- ## Vendor contract AI review checklist _The three clauses a small company should not sign an AI vendor without._ ### Template | Clause | What to look for | Present? | Action if missing | |---|---|---|---| | Training-data use | our data is not used to train the vendor's models, or opt-out is available and we take it | Y / N | request the enterprise/opt-out plan; get it in writing | | Data handling and deletion | where data is processed, retention period, deletion on request | Y / N | ask for the DPA; confirm region and retention | | Liability and IP | we own outputs; some indemnity for IP claims on outputs | Y / N | push for output ownership at minimum; note residual risk | ### Worked example | Clause | Vendor status | Action | |---|---|---| | Training-data use | opt-out available on the Team plan, not default | upgraded to Team; opt-out confirmed by email | | Data handling / deletion | DPA covers 30-day retention, EU processing | acceptable | | Liability / IP | we own outputs; no IP indemnity | accepted; low exposure for our use, noted in the inventory | ### Acceptance criteria - All three clauses are checked before signing any AI vendor. - Missing clauses have an action and, if accepted, a noted residual risk. - The result is recorded next to the vendor in the inventory. --- ## AI data flow summary _One diagram-in-words: what personal data goes to which AI tool, and where it ends up._ ### Template | Data category | Source | AI tool it reaches | Sent how | Stored where / how long | Basis | |---|---|---|---|---|---| | | | | API / SaaS / embedded | | consent / contract / legitimate interest | ### Worked example | Data category | Source | AI tool | Sent how | Stored / how long | Basis | |---|---|---|---|---|---| | Customer email content | support inbox | Support assistant (Vendor API) | API per ticket | vendor: 30 days; ours: with the ticket | contract performance | | Prospect name + company | CRM | Sales email tool | SaaS sync | vendor: account lifetime | legitimate interest | | Source code | repo | Code assistant | IDE plugin | vendor: 0 days (zero-retention) | n/a (not personal data) | ### Acceptance criteria - Every flow of personal data into an AI tool is listed with its basis. - Vendor and internal retention are both recorded. - The summary is something you could hand to a customer's security reviewer. --- ## Due diligence AI governance questionnaire response _A reusable set of answers for the AI governance section of a customer or investor security questionnaire._ ### Template > Keep current. Reuse across deals. - **Do you maintain an AI system inventory?** yes / how often reviewed - **Do you have an AI acceptable use policy?** yes / acknowledged at onboarding - **How is customer data protected when AI tools are used?** the data flow summary in brief; opt-outs taken - **Do you use customer data to train models?** no / with the opt-outs in place - **Human oversight of AI decisions affecting individuals?** where applicable, describe it - **Incident process for AI failures?** brief description - **Sub-processors that are AI vendors:** list, with links to their compliance pages ### Worked example "We maintain an inventory of all AI tools in use, reviewed quarterly. All staff acknowledge a one-page AI acceptable use policy at onboarding. Customer data sent to AI tools is limited to what the feature requires; we use enterprise plans with training opt-outs on every AI vendor that offers one, and we do not permit customer personal data in tools without a DPA. We do not use customer data to train models. AI-assisted outputs to customers are reviewed by a person before sending. AI-related errors are handled under our incident process [link]. Our AI sub-processors are: [list with compliance-page links]." ### Acceptance criteria - Answers are true and match the inventory, policy, and data flow summary. - The AI sub-processor list is current. - The response is version-controlled and reused rather than rewritten per deal. --- ## Governance controls this kit produces evidence for - **MGV-002**: The inventory and vendor checklist are a lightweight intake and approval process. - **SCT-007**: The one-page policy is the acceptable use policy for external AI tools. - **PRC-001**: The three-clause vendor check is proportionate vendor due diligence. - **DGC-002**: The data flow summary documents personal data handling across AI tools. - **HOC-001**: Even the lightweight inventory records enough to reason about each system's risk.