Can AI Train on Your Contract Data? Check the Exact Upload Route
By Anirudh Menon, Viaclause editor
Separate an AI no-training promise from storage and deletion. Map each document copy to its product, account, terms and evidence before approving an upload.
AI-generated editorial illustration by Viaclause. Person and setting are fictional.
Whether an AI vendor can train on your contract data depends on the terms governing the exact service you use. A no-training commitment is important, but it does not answer how long files remain stored or what happens to copies held by a connected application.
Use the AI contract review evaluation guide to define the buying task and acceptance checks for the wider workflow.
Before approving an upload, procurement and legal teams need an answer tied to a product, account and workflow. “We use this model” is not enough detail. This guide proposes a record for that decision, using Anthropic's published materials as a concrete example rather than a ranking of vendors.
Find the promise and its scope
Anthropic's Commercial Terms, section B, prohibit training models on Customer Content from the covered services. That defined content includes inputs and outputs. The terms identify which offerings they govern and distinguish consumer offerings. Section E.4 separately addresses destruction requests and exceptions for legal compliance and automated backups.
Those provisions show why a review needs separate answers for use, storage and deletion. Do not convert the training restriction into a statement that no copy exists. Nor should you apply a commercial commitment to a different account or service without checking its governing terms.
The Claude Platform retention documentation adds feature-level detail. It describes zero data retention, or ZDR, as an arrangement covering eligible routes and features. Its table excludes the Files API from ZDR and says files remain until deletion or configured expiration. It also says requests using features marked ineligible for ZDR are not automatically blocked by that arrangement; the feature's retention policy applies. Check current eligibility and your actual contract before relying on it.
Ask the person configuring the tool to draw the route
Use one test document made entirely of invented information. Give it a distinctive harmless filename so the team can find it without putting client material into an unapproved system. Ask the administrator to identify every step from selection to final output.
Hypothetical route: an employee selects a file in a contract-review application. That application extracts text, sends it to a model service, saves the answer, and lets the employee download a report. We are not claiming a particular vendor uses this design. It is a scenario for exposing questions that a model-only review could miss.
For this scenario, create separate rows for the original file, extracted text, submitted prompt, model response and downloaded report. Add other copies only when the architecture or observed workflow supports them. Do not invent hidden storage simply because it is possible.
Complete a record for each copy
- Holder and route: which organization holds the copy, in which named product, account or workspace?
- Purpose: processing this request, product improvement, model training, support or another specifically documented use?
- Permission: the governing clause, applicable product tier, relevant setting and any separate opt-in. Record the date checked.
- Storage: whether the copy persists, the documented duration or trigger, and qualifications such as backups or legal holds.
- Deletion: who can request it, which copy the action targets, what confirmation is available and what remains excluded?
- Evidence type: contract text, product documentation, administrator configuration, observed interface result or vendor confirmation.
- Open question: the exact missing fact, its owner and the date needed for the upload decision.
Keep those evidence types separate. A screenshot of a disabled training setting records a configuration. It does not prove the full contractual commitment. Conversely, a contractual promise does not demonstrate that the application was configured to use the account covered by it.
Run a limited deletion check without overstating it
For the synthetic document, follow the documented deletion process. Check the locations the team can legitimately observe: the application's file list, saved conversation, exported report and local download folder, where applicable. Record what disappeared, what remained and what could not be checked.
Do not write “all data deleted” based on a disappearing file tile. The defensible result is narrower: the identified item is no longer visible in the checked interface after the stated action. Backend handling, exceptions and time limits still depend on the governing commitments and appropriate vendor evidence.
Do not test model training by repeatedly asking whether the model remembers your document. That is not a reliable substitute for the contractual and configuration evidence this review requires. The proposed exercise checks a workflow, not the vendor's training pipeline.
Make the decision specific enough to revisit
Record approval for a defined route and data category, with any unresolved restrictions. For example: “Synthetic evaluation documents only in the named workspace until the application's own retention and deletion terms are confirmed.” That is a proposed internal decision statement, not a conclusion that any real product is unsafe.
Revisit the record when the application changes its model provider, upload method, feature set or account tier. For connected practice-management tools, the MyCase capability review shows how to keep new actions outside an earlier pilot's approval until reviewed.
Once the route is approved, the content still needs review. Use the contract evidence prompt for source checking, and use the export-before-exit checklist when planning a transition. A satisfactory no-training answer is one part of the decision, not the whole file-handling plan.
Sources
- Anthropic Commercial Terms of Service, effective June 17, 2025, introductory scope, section B and section E.4; checked September 27, 2026. Contractual training restriction and distinct destruction provisions.
- Claude Platform: API and data retention, checked September 27, 2026, ZDR coverage, feature eligibility table and ineligible-feature explanation. Product-specific example, not a universal retention schedule.