// AI GOVERNANCE / VENDOR REVIEW
AI Vendor Review: Where Your Data Actually Goes
Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, and Claude Mythos 5 require 30-day retention and are not ZDR-eligible unless expressly authorized by Anthropic. Its model-specific retention rules belong in the review of the exact model you intend to call.
This closes the six-part AI governance series that opened with the case against treating governance as a policy document. The first five pieces covered access, review, logging, and audit scope, the surfaces a company controls internally. This one is the outward-facing surface: what you are actually buying when you sign with an AI vendor, and where your data goes after you hit send.
Zero data retention is not one setting
Every major AI vendor now advertises some version of zero data retention. Treated as a single feature you either have or do not, the phrase hides three different gates, and I checked each one directly against the vendor's own current documentation rather than a compliance blog's summary of it.
OpenAI requires prior approval and additional requirements for ZDR; contact sales about eligibility. Approved customers configure it at organization and project level. Endpoint and feature eligibility still constrain that setting: Batch is ineligible, and some application state may persist. OpenAI's data-control table distinguishes abuse-monitoring logs from application state. Default abuse logs may retain content up to 30 days, with stated safety or legal exceptions.
Default conversation-content handling is separate from an enabled ZDR arrangement. Anthropic requires sales or account-team enablement per organization. Eligible features are covered; Covered Models require an exception, and flagged-content or legal-hold retention can apply. Check the product, model, feature, and arrangement together, including workspace controls where offered.
Google's paid Gemini API services do not use prompts or responses to improve its products. Its retention guide separately requires disabling Interactions state storage and Live session resumption, avoiding Search and Maps grounding, deleting uploaded files, and avoiding explicit context caching for a zero-data footprint. Implicit caching is project-isolated RAM with a 24-hour TTL, which Google treats differently from storage at rest. Paid services still have limited abuse-monitoring logging; Google directs guaranteed ZDR or enterprise-DPA workloads to Vertex AI.
Training and retention are two different promises
"We don't train on your data" answers one question. It does not answer whether your prompts are stored, for how long, or who can read them while they sit there. Ask both questions for the service you plan to use: whether the provider uses content for training and what it retains for processing, safety, or legal obligations.
Check the service, account, billing status, and region used in the pilot and in production. Under Google's Gemini API terms, Unpaid Services content may be used to improve products and may receive human review. Paid Services prompts and responses are not used to improve products. Google applies Paid Services data handling in the EEA, Switzerland, and the UK even for unpaid use; AI Studio can also qualify through a billing-enabled project or a Workspace enterprise account. Anthropic's commercial terms prohibit training on customer content without express permission, while its consumer policy has separate training choices. A change of tier can change the terms that govern the work. Verify them rather than inferring the privacy posture from the API key alone.
| API service | Training terms | Retention default | Zero retention gate |
|---|---|---|---|
| OpenAI API | No training since March 2023, unless opted in | Up to 30 days, abuse monitoring | Prior approval; organization/project configuration; endpoint and feature limits |
| Claude API | No training without express permission | Conversation content not retained by default on the Claude API; feature, model, safety, and legal exceptions | Enablement per organization; eligible features; model, safety, and legal exceptions |
| Google (Gemini API) | Paid Services prompts/responses not used to improve products; account and region rules apply | Limited abuse logs; separate feature and file storage | Feature controls, file deletion; Vertex AI for guaranteed ZDR/enterprise DPA |
Anthropic documents separate retention for claude.ai chat, files and projects, and for local and remote session transcripts. This table covers the API.
Where your data actually goes: subprocessors
Your vendor may rely on other processors for hosting, search, monitoring, or other functions. A subprocessor list identifies those relationships; it does not mean every listed company receives every customer's data. Map the relevant service, enabled feature, processing purpose, and region to the processors that can handle your workload.
OpenAI publishes a subprocessor list with processing purposes and locations, plus email change notifications. Anthropic publishes its list through its Trust Center and lists TurboPuffer for web search. Preserve the applicable list and notifications with your review so a later change can be assessed against the workload you approved.
The relationship can run deeper than a single hop. Microsoft says OpenAI became a listed subprocessor on June 23, 2026 and available for use on July 9. That documentation concerns models operated by OpenAI, separate from models operated by Microsoft through Azure OpenAI. Availability and activation depend on product, tenant eligibility, and administrative settings. Check the operator and active configuration for the Copilot experience you use before concluding which processing relationship applies.
What to actually ask a vendor
Five questions, answered in writing, not inferred from a trust badge or a blog post summarizing one.
- What is the training default for the specific product and tier I am buying? Not the company's general policy. The product, the tier, the account type, since free and paid frequently diverge and consumer and commercial almost always do.
- What is the retention default, and is my chosen model excluded from any zero-retention option you advertise? A vendor's homepage claim about zero retention is a starting point. The model-specific exception list is the actual answer.
- Can I see your current subprocessor list, and how do I get notified when it changes? A list with no update mechanism is a snapshot, not a control.
- Will you put the training default and retention default in the DPA, not just the privacy page? Check the contract's change provisions, notices, exception clauses, and obligations rather than assuming the webpage fixes them permanently.
- Are you a subprocessor to anything else I already use? The Microsoft and OpenAI example shows why you should inspect dependencies behind the product you buy.
None of this requires distrusting OpenAI, Anthropic, or Google. Every fact in this piece came from each vendor's own current documentation, checked directly rather than taken from a summary, which is exactly the discipline this series has argued for since the audit-scope piece: a vendor's policy page is a claim. Keep a dated evidence record of the terms and configuration you checked. Connector controls also need to constrain where tool calls send data.
That is the same test this whole series has run against a policy document, a review gate, a connector, a log, and now a vendor. A claim is not a control until someone can point to the mechanism behind it and the date it was last checked.
Frequently asked questions
Does OpenAI train on my API data?
OpenAI does not train on API data by default unless you opt in. Retention is separate: default abuse logs may keep content up to 30 days, subject to stated exceptions. ZDR needs prior approval; approved customers configure organization and project controls, with endpoint and feature limits. Batch is not ZDR-eligible. Check the current data-control table for the actual workload.
What is zero data retention and does every AI vendor offer it the same way?
ZDR concerns storage of customer prompts and responses, with provider-specific limits. Enablement is separate from eligibility: OpenAI uses approved organization/project controls, Anthropic enables eligible arrangements per organization, and Google documents feature and file controls while directing guaranteed-ZDR workloads to Vertex AI. Check model exclusions, application state, safety or legal exceptions, and third-party destinations.
What should an AI vendor due diligence checklist include?
Record five things for the workload: the training terms for the exact service, account, tier, and region; retention and model or feature exceptions; relevant subprocessors, purposes, and locations; the DPA's commitments, exceptions, change provisions, and notices; and dependencies behind other products you already use. Confirm the active configuration as well as the contract. A public list or policy alone does not establish the path your data took.
Is the free tier of an AI tool different from the paid tier for data privacy?
The terms can differ, but check the actual service, account, and region. Google's Unpaid Services content may be used to improve products and reviewed by humans; Paid Services prompts and responses are not used to improve products. Paid Services data handling also applies to unpaid use in the EEA, Switzerland, and UK. AI Studio has additional billing-project and Workspace-enterprise rules. Anthropic's consumer and commercial products have separate terms. Verify training, retention, and exceptions for the account used in production.
What is a subprocessor and why does it matter for AI vendor review?
A subprocessor handles data on the vendor's behalf, for example hosting or monitoring. The vendor's DPA can impose obligations downstream: OpenAI's DPA requires comparable subprocessor obligations and retains liability subject to the agreement. Check the actual contract, processing purposes, locations, change notices, and objection rights. Those protections do not remove the need to review the chain.
Sources
OpenAI API training and Zero Data Retention eligibility, endpoints, and abuse-monitoring retention window: OpenAI, "Data controls in the OpenAI platform" (checked October 6, 2026). OpenAI's August 19, 2026 announcement continuing Zero Data Retention for frontier models and previewing Private Safety Processing: OpenAI, "Offering Zero Data Retention for frontier models", corroborated by Techstrong.ai and OpenAI's own August 19, 2026 announcement post on X. OpenAI subprocessor list and June 2, 2026 update: OpenAI, Sub-processor list. OpenAI's addition to Microsoft's subprocessor list: Microsoft Learn, "OpenAI as a subprocessor in Microsoft Online Services". Anthropic's retention defaults, Zero Data Retention scope, and the Covered Models list requiring 30-day retention: Anthropic, "API and data retention" (Claude Platform Docs, checked October 6, 2026). Anthropic commercial no-training terms: Anthropic, Commercial Terms of Service. Anthropic Trust Center launch and subprocessor updates: Anthropic Trust Center. Google Gemini API training rules by service, account, and region: Google, Gemini API Additional Terms of Service (checked October 6, 2026). Google Zero Data Retention configuration requirements: Google AI for Developers, "Zero data retention in the Gemini Developer API" (checked October 6, 2026).
About the author
Jeff Brokaw is a sitting CMO and Certified Chief AI Officer who ships AI in production, not slideware. He has been building AI systems commercially since 2016. He built the commercial engine behind $185M in new-business revenue for a defense manufacturer, and authored the go-to-market behind a $114M institutional raise that came together in under 30 days.