Shadow AI is the starting point
Marketers already use consumer AI tools, often on personal accounts, with product images, briefs and sometimes customer data. Size the problem before choosing the replacement: proxy, DNS and CASB logs show traffic to AI services, expense claims show paid subscriptions, and browser management shows extensions, including AI assistants that can read every page a user opens, internal systems included. Blocking alone rarely works, because the work moves to phones. Offer a sanctioned platform that covers the same needs behind company login, and a short rule on what must never go into unapproved tools: unreleased products, customer data and confidential financials.
Identity, roles and external users
Use SSO through your identity provider, over SAML 2.0 or OpenID Connect, so MFA and conditional access apply. Provision and deprovision with SCIM, so leavers lose access the day HR processes them. Content platforms tend to collect agency staff and freelancers, whose accounts are easily forgotten; give them expiry dates and an internal sponsor. Map roles from directory groups (creator, reviewer, approver, brand administrator, workspace administrator) rather than assigning them by hand. Where several brands, markets or agencies share one platform, confirm that workspaces isolate assets, brand rules and saved elements from one another.
The model supply chain
Behind one interface there may be many model providers, each a subprocessor with its own region, retention terms and change schedule. Ask for the list, how new providers are approved and how customers are told. Treat uploaded documents as untrusted input as well: a brief or PDF can carry hidden instructions aimed at a text model, the prompt injection risk that tops the OWASP Top 10 for LLM applications. Synthetic White offers EU or US data residency, zero-retention agreements with model providers, no training on anything customers upload, private or self-hosted models for confidential work, deployment in cloud, private VPC or on-premise, and a log of the model behind every asset.
Vendor risk assessment for AI vendors
Standard questionnaires such as the SIG or the Cloud Security Alliance's CAIQ cover the usual ground: access control, encryption, incident response, business continuity and subprocessors. Add the questions this category needs. Can confidential work be restricted to named models or private deployments? What stops one customer's data from reaching another customer's outputs? How quickly will you hear about an incident at a model provider? Which logs can you export for your own monitoring? Financial entities in the EU must also record the platform in their register of ICT third-party arrangements under DORA (Regulation 2022/2554, applying since 17 January 2025).
Updated 25 September 2026