External LLM Providers: Trace the SaaS AI Data Chain
Trace SaaS AI model-provider risk by mapping the application, AI feature, external model service, supporting subprocessors and downstream actions. For each boundary, record the information involved, applicable terms and available evidence. Identifying a provider establishes a dependency; it does not prove that a particular tenant sent a specific payload or enabled the feature.
An employee opens a familiar SaaS product and asks it to summarize a document. The interface shows one vendor name. The processing path may involve several organizations. A useful review makes that path visible without pretending that a public subprocessor list is a packet capture or that every available AI feature is in use.

Follow the document through five boundaries
Customer to SaaS feature: identify the selected content, attachments, retrieval sources and user authorization.
SaaS processing: ask what the application stores, transforms or logs before calling an AI service.
Model service: identify the provider, named offering, deployment region and applicable processing arrangement.
Supporting services: check relevant routing, hosting, monitoring and other disclosed processing dependencies.
Output and action: determine where the response is stored and whether it can trigger a message, update or other operation.
This sequence is a review model, not a claim that every product contains all five boundaries. Mark a boundary “not applicable” only after obtaining evidence. Mark it “unknown” when the evidence is missing. That distinction prevents a simple architecture sketch from becoming a falsely complete statement about the deployment.
A diagram is useful only if its edges have evidence
For each connection, record the sender, recipient, information category, processing purpose and source of the claim. Include the source date and tenant scope. A vendor diagram may show the standard architecture while a configuration export shows a particular deployment. Both can be useful, but they answer different questions.
Suppose a public page says the product supports Provider A and Provider B. That is evidence of supported options. It does not identify your selected option. A tenant setting or scoped vendor response may answer that second question. Neither necessarily proves what was transmitted during a historical event; that requires suitable event evidence, if available and authorized.
Read direct-provider promises carefully
OpenAI's enterprise privacy information describes its business offerings. Use it to understand the relevant service context, then ask the SaaS vendor which offering and contractual arrangement it uses. Do not assume your SaaS agreement inherits every control available to a direct provider customer, including a configurable retention option.
Microsoft's documentation on external model provider services illustrates how an application platform can involve external provider services. The lesson for reviewers is to identify the actual route and terms. A recognizable model name is insufficient because commercial service, region, configuration and customer relationship still matter.
Key takeaways
Map processing and action boundaries together.
Separate provider identification from payload evidence.
Track changes that can alter the approved data path.
The request that replaces a sprawling questionnaire
Ask the vendor for a feature-specific data-flow explanation covering inputs, retrieved content, outputs and relevant logs. Request its applicable subprocessor information and describe the tenant, plan and region being reviewed. Ask which dependencies can change without a customer configuration change and how the customer is notified.
Then compare the response with the administrator's view. If the feature offers multiple providers, document who can select one and whether changes are logged. If the vendor does not expose selection to the customer, record that limitation and the alternative evidence. The result should be an answer about your deployment, not a catalogue of every capability the vendor sells.
Example: the unknown at the middle of the chain
In a hypothetical ticketing system, the vendor identifies an external model provider but does not explain whether attachments or only selected ticket text are sent. The dependency can be recorded as identified. The payload scope remains unresolved. Approval for confidential attachments should not follow from the provider name alone.
The application owner can request a feature-specific answer, test an authorized non-sensitive workflow and inspect available settings. Any test should use synthetic information and should avoid claiming visibility the organization does not possess. If transmission evidence is unavailable, document the vendor's scoped statement and the remaining uncertainty rather than fabricating an observation.
Put actions on the map too
A drafting feature and an agent that sends emails create different review questions. Record connector permissions, execution identity, approval points and downstream destinations where relevant. NIST's Generative AI Profile offers wider risk context; this mapping approach focuses the review on the boundaries a specific business workflow actually crosses.
Waldo's AI Risk & Governance and platform overview can support discovery and assessment context across SaaS, AI and identities. Treat discovery as a starting point for the data-chain review. Confirm enablement, selected providers and action settings with the responsible administrator and applicable vendor evidence.
Keep the map maintainable
Assign an owner and list events that reopen the map: a new provider, region change, additional retrieval source or expanded action permission. Store the map beside the application record and date each evidence item. A short, current chain with explicit unknowns is more useful than an elaborate diagram whose assumptions cannot be checked.
Add a change owner to every uncertain edge
Some processing details may remain unavailable to the customer. Assign those unknowns to a vendor-evidence owner and record how changes will be communicated. A missing architectural detail should not become an anonymous permanent question. The business sponsor needs to know whether a future provider substitution, routing change or expanded retrieval source could invalidate the current use decision.
Keep the map at a level the organization can maintain. A concise record of material boundaries, evidence and open questions is usually easier to update than an elaborate diagram copied from a sales presentation. During a handoff, ask the receiving owner to explain the approved data scope from the map. If that explanation depends on undocumented assumptions, revise the record before treating it as a dependable governance artifact.
Frequently asked questions
Does a subprocessor list prove an AI feature is enabled?
No. It describes disclosed dependencies within its stated scope. Verify feature enablement and tenant configuration separately. Preserve the distinction between vendor capability, selected configuration and observed activity.
Must we identify every model used by the vendor?
Focus first on the models and services relevant to the proposed workflow. Ask how routing or fallback can change that set. Record unknown dependencies that could materially affect the use decision.
Can a model name establish a processing region?
No. Obtain evidence about the named service and deployment arrangement. Region statements may apply to storage, inference or other processing differently, so retain the source scope.
What if no payload-level evidence exists?
Use the available architecture, configuration and contractual evidence, label its limits and decide accordingly. Do not turn a provider association into a claim that you observed a particular data transfer.
When should the chain be reviewed again?
Review it after relevant provider, feature, region, retrieval or permission changes. Add a scheduled check as a backstop, and name the owner responsible for monitoring notices and updating evidence.
Next step: Explore AI Risk & Governance. Start with the evidence gap that affects your current workflow and assign its verification to an accountable owner.
Sources and further reading
NIST: Generative Artificial Intelligence Profile, NIST AI 600-1. Published July 26, 2024. Accessed October 3, 2026.
OpenAI: Enterprise privacy. Accessed October 3, 2026.
Microsoft: Data, Privacy, and Security for Microsoft 365 Copilot. Accessed October 3, 2026.
Microsoft: External model provider services. Accessed October 3, 2026.




Comments