Follow one ordinary request
Suppose a colleague asks for a briefing on an open case. The configuration may provide case fields, conversation history and selected supporting material. That context becomes part of the model request. The answer may return to a chat window, a record view or another configured process.
For REDA AI, Salesforce runs the configured workflow while external AI providers process information needed for reasoning. A diagram that shows only Salesforce screens can miss this boundary. Our Governance & Trust page explains the product’s data flow and channel distinctions.
- Salesforce context
- AI provider processes
- Answer or proposed action
- Salesforce workflow review
Ask four separate data questions
| Question | Evidence to request |
|---|---|
| What is sent? | A representative request showing the fields, documents and conversation context involved. |
| Who can access it? | The running identity, object/field access, sharing and tool behaviour for this channel. |
| What is retained? | Provider, host and application retention terms, including exceptions. |
| Is it used for training? | The applicable product/account terms; this is separate from retention. |
A promise not to train on a request does not, by itself, mean the request is never retained. Likewise, a deletion window at one provider says nothing about a separate assistant host or your own Salesforce records.
Read the provider terms for your actual setup
Anthropic’s commercial retention page describes deletion of API inputs and outputs within 30 days under its standard terms, with stated exceptions and possible contractual arrangements. That is not a universal “nothing is stored” promise. Confirm the endpoint, feature and account agreement being used; other providers and assistant hosts have their own terms. [1]
Keep the review dated. If the provider, model or enabled feature changes, revisit the relevant terms and the information included in the request. An old architecture approval should not silently cover a materially different data path.
Test access with the people who will use it
Illustrative check. A service user can see their own queue but not a restricted account. Ask for a briefing involving both. The test should establish what the implementation actually returns and whether the action route respects its intended scope.
Salesforce documents that sharing, object permissions and field access need appropriate enforcement in Apex. Do not infer the whole access model from a class name or a prompt that says “only show permitted data”. [2]
Give the review a useful finish
Record the approved information categories, providers, channel, permitted actions and owner. Agree how configuration changes are reviewed and what would pause the workflow. Start with representative synthetic data where possible, then use real information only under the appropriate organisational approvals.
REDA can walk your administrators and security reviewers through the intended configuration. The goal is a decision grounded in the actual data path and evidence, rather than a blanket assurance that AI is safe.
Sources
- Anthropic — How long do you store my organization’s data?. Accessed 22 September 2026.
- Salesforce Developers — Secure Apex classes. Accessed 22 September 2026.