Broker guide
Zero Data Retention: Check the AI Data Chain
Does zero data retention cover your whole AI workflow? Check provider API terms, vendor storage and the records your brokerage still keeps.
- Published
- Updated
Zero data retention limits what an AI service keeps after processing, within the scope of its agreement and settings. Before approving an upload, check whether the promise covers the exact model request and find out what the broker tool keeps before and after that request.
Define What Zero Data Retention Covers
Each provider specifies what its zero data retention terms cover. Get the definition in writing for the service you intend to use and check these four questions separately.
| Question | What to look for |
|---|---|
| Can the provider train on the data? | The terms for using inputs and outputs to train or improve models |
| Can abuse monitoring keep the data? | Whether safety logs contain prompts or responses, plus any exceptions for flagged content or legal obligations |
| Can a feature store the data to work? | Saved files, conversation state, background jobs and caches that let a service resume work or retrieve a result |
| What account data remains? | The separate rules for billing and operational metadata, such as request timestamps and usage counts |
An eligible application programming interface (API) setting applies to the software’s calls to that provider. A chat application’s training and history controls govern that application. For example, OpenAI’s ChatGPT data controls let someone turn off model improvement while regular conversations remain in history.
Trace a Loan Document Through Every Service
Imagine a fictional broker tool that receives a sample payslip and sends its text to a model to extract the gross pay and pay period. It saves the result for a broker to compare with the payslip. The model provider might discard the covered content after processing while the vendor keeps the uploaded PDF and the extracted result.
| Where the data goes | What might remain | Who supplies the evidence |
|---|---|---|
| Upload and text extraction | Original PDF and an extracted text copy if the tool reads a scanned page | Vendor, including any outsourced text extraction |
| Background processing | A queued job with the text or a link to the file, plus a failed job waiting to retry | Vendor or queue operator through the vendor |
| Model API | Request, model output and any state or records covered by an exception | Model provider’s terms and documentation, plus the vendor’s account and configuration evidence |
| Monitoring and support | Error traces with snippets of the payslip or a file attached to a support ticket | Vendor, including any logging and support subprocessors |
| Results and backups | Extracted figures in the vendor database and copies of that database or uploaded files in backups | Vendor and its storage providers |
| Brokerage loan file | Original payslip, checked income figure and the reviewer’s note | Brokerage or licensee, under its own access rules and records schedule |
Check the destinations on failed jobs too. A successful request can leave a short-lived copy while a retry or diagnostic trace keeps the same content for longer.
Check the Provider Contract and Endpoint
- Get the agreement that covers the vendor’s production account. Record the contracting entity, service and applicable order form or written approval. Match the account or organisation and project in the approval to the ones making the request.
- Ask for a dated export or screenshot of the retention setting on that project. Check whether the project inherits the organisation’s setting or overrides it. Use the project’s identifier and hide secrets in the evidence.
- Compare the exact endpoint, model and enabled features with the provider’s current documentation. Map attachments, file stores, search tools and background processing to their own rules. Ask for the timing and trigger for deleting any persisted file.
- Match the written exceptions to the intended upload. Record how the provider handles abuse monitoring, flagged content, legal retention and any notices that change model eligibility for that account.
The OpenAI API data controls illustrate why that check goes beyond the account setting. OpenAI requires prior approval and acceptance of additional requirements. It lists eligible uses of /v1/responses with limitations, while /v1/files is ineligible and files can persist until deletion or configured expiry.
The documentation also describes when flagged image inputs can be retained for safety review. Ask which exceptions apply to the model and files you plan to send.
Check the services agreement and its data processing addendum against the signed terms. The published agreement separately limits use of customer content for model improvement unless the customer explicitly agrees.
Ask the Vendor for Its Own Retention Evidence
Request the vendor’s agreement and data-flow evidence even if it already sent the model provider’s documentation.
For example, Adobe describes its own Acrobat AI Assistant storage separately from the restrictions it says apply to its model provider. It documents a 12-hour cloud retention period for session content, with exceptions for retained chat history and user-reported content. Check how the product documentation and the vendor’s data processing terms apply to your account.
Copy and adapt this request for the workflow your brokerage wants to approve.
For the payslip extraction workflow in our proposed plan, please send the applicable agreement and retention schedule. Attach a data-flow diagram. Identify where you hold the uploaded file, extracted text and returned income figure. Distinguish your own controls from the model provider’s controls.
For each copy, identify the operator and any subprocessor. Name the regions for processing, storage and support access. Explain which staff roles can see the content and how access is authorised and recorded.
Include retries, operational logs, monitoring traces, support attachments and backups. For each copy, give the retention period, the deletion trigger and the time until the backup copy expires. Explain how you confirm deletion or export records at contract exit and what happens if a backup is restored.
Provide the incident notification deadline and the information you commit to sharing with us. Attach the provider documentation and the account-specific approval or configuration evidence that supports any zero data retention claim.
Save the signed terms, technical documentation and vendor answers against the claims they support. If a settings export shows a different project from the one in the vendor’s data flow, resolve the mismatch before approving any client uploads.
Record the Permitted Use and Remaining Copies
Record the following decisions in your AI risk assessment. Give each approval an owner, identify the applicable production configuration and reassess when the vendor changes a provider, model, feature or storage location.
| Task and data | Provider assurance | Vendor assurance | Loan-file evidence | Decision or upload block |
|---|---|---|---|---|
| Test extraction with a fully fictional payslip | Document the tested feature and account | Confirm the sample stays out of the production loan file | Keep the test result with the tool assessment | Synthetic test only while live-data assurances are missing |
| Extract income from a real payslip | Signed terms, approved project and compatible endpoint with acceptable exceptions | Documented uploads, logs, support access and deletion on approved terms | Original payslip, checked figure and reviewer note | Hold upload until both assurances are evidenced and the licensee approves this use |
| Verify income manually in the approved loan system | No model request | Apply the existing controls for the loan system | Original payslip and verified figure | Use this route while an AI upload is on hold |
For entities covered by the Australian Privacy Principles, APP 11 governs how you secure and dispose of personal information. Take reasonable steps to destroy or de-identify it when you no longer need it for a permitted purpose. Exceptions include records an Australian law or court or tribunal order requires you to retain.
Under section 120, a credit licensee must provide the preliminary assessment on eligible requests within seven years of the credit assistance quote. Keep the required loan evidence available under your licensee’s records schedule. Track deletion of the vendor’s working copies separately.