“Is our data used to train the model?” is an important question—but it is not a complete AI security assessment. Even when training is excluded, sensitive information may still pass through conversation history, retrieved documents, application logs, connected tools, and service-specific retention mechanisms.

This session follows a request through a cloud-hosted AI application to examine where data travels, which identities can access it, and what may remain after the response is returned. We’ll distinguish model training from inference, service retention from application logging, and conversation history from persistent agent memory.

Using worked examples, we’ll compare a direct model interaction with a retrieval-augmented application and an agent that invokes tools. We’ll examine how each introduces different access and data-handling requirements, and why private networking alone cannot establish that every document, response, or downstream action is authorised.

Rather than treating all AI services as having the same behaviour, we’ll identify which questions require provider documentation, configuration evidence, contractual assurance, or application-level testing.

Attendees will leave able to:
- Map the data paths and access boundaries of an AI application.
- Distinguish training, inference, retention, logging, and memory when assessing sensitive-data use.
- Identify the evidence needed to evaluate retrieval permissions, connected tools, and service-specific data handling.

Audience: Cloud architects, security professionals, platform engineers, and data owners.
Level: Intermediate.
Duration: 45 minutes; expandable to 60 minutes with deeper scenario discussion.
Prerequisites: Basic understanding of cloud identity and application architecture; no machine-learning knowledge required.
Format: Slides, data-flow diagrams, and worked examples.
Scope: Azure-centred examples, with principles applicable to other cloud and SaaS AI services. Provider-specific claims will be checked against current documentation before delivery.