Key takeaways
- Customer-controlled hosting does not automatically keep external model requests inside the same network.
- Not using data for training is different from not retaining it.
- Provider controls depend on the service, endpoint, contract, configuration, and eligibility.
- Deal separation, access restrictions, logs, and exports need implementation testing.
- This article is an engineering review framework, not a determination of NDA or regulatory compliance.
Before an AI workflow receives confidential deal material, the firm should approve its complete data flow. Application hosting is one part of that review. Parsing services, embeddings, model endpoints, telemetry, logs, backups, integrations, and exports may introduce separate processing and storage locations.
Map each processing path
Document the sources, information sent to each service, processing location, retained state, access roles, and deletion rules. A customer-hosted application can still send excerpts to an external model provider. Private networking changes the connection path; it does not by itself establish local inference or zero retention.
Separate model training from retention
OpenAI’s API documentation states that API data is not used for model training by default unless the customer opts in. It separately describes abuse-monitoring logs, application-state storage, and endpoint-specific retention. Zero Data Retention and Modified Abuse Monitoring require approval and have limitations. See the current controls at OpenAI API data controls
Do not extend one provider’s terms to every model service. Review the selected provider, endpoint, feature, configuration, region, and contract with the people responsible for security and data handling.
Test access and deal separation
Define who can upload, retrieve, review, export, and administer material for each deal. Test attempts to access another deal or workstream, including retrieval results, shared caches, logs, background jobs, and exports. Encryption supports the design but does not make cross-deal access technically impossible by itself.
Agree what the system may do
Limit tools and data to the task. Define approval requirements for changes, exports, external messages, and final deliverables. Logs should support investigation without unnecessarily exposing confidential material.
Review requirements for the engagement
The firm’s legal and security owners determine whether the proposed arrangement meets its contracts and applicable requirements. This article does not claim automatic NDA compliance, organization-wide certification, universal zero retention, or that all processing remains inside a VPC.
Dotnitron’s delivery context
Our case records describe specialist contributions within wider delivery teams: seven regional enterprise environments during one engagement, and five customer-controlled private-markets deployments live at the end of another. Those records do not prove a common security configuration across every environment. See Dotnitron client work and review our project-specific approach at Dotnitron security approach.
Use this guide
Turn the article into a working session.
Pick one workflow from the article and map it against your own team. Write down the input sources, current manual steps, reviewer decisions, output format, and the metric that would prove the workflow is worth automating.
- What work should agents prepare before a human reviews it?
- Which documents, data sources, tools, or approved system connections would the workflow need?
- What output would make a reviewer say, this saves real time?