Where Does Your Data Go?
Every prompt your team submits travels through an infrastructure path. For an Australian or New Zealand organisation with regulatory duties, understanding that journey is how you separate genuine enterprise isolation from consumer defaults.
Start with the podcast discussion.
Listen to Steve Wilson explain how to trace data pathways across commercial AI subscriptions, what enterprise data protection terms mean in practice, and how to answer tough privacy questions.
Watch the data flow walkthrough.
A short overview of the route a prompt takes from user input to secure processing and back, highlighting where encryption boundaries exist and how logging works.
Enterprise AI does not train on your business prompts. Your task is verifying that promise.
Consumer accounts retain conversations to improve future systems. Commercial contracts guarantee that your prompts and retrieved files stay isolated within dedicated tenant boundaries.
Where does your data go when you hit enter?
When staff members start pasting project background into an AI assistant, senior management often worries about corporate information leaking into public knowledge bases. That concern is sensible when people use free consumer tools. Those consumer versions routinely log inputs to train future models.
Commercial products work under different rules. When you deploy Microsoft 365 Copilot, Claude Teams, or paid enterprise workspaces, the supplier commits to legal boundaries. Your team needs to understand the mechanics so you can explain them clearly to risk committees, clients, and privacy officers.
Understanding these flows builds directly on our earlier work analysing the Microsoft Copilot landscape across core workplace tools. It also directly influences who can see what in your systems when AI connects directly to organisational drives.
The three states of business information
To evaluate vendor security statements properly, you must follow your information through three distinct stages of handling.
Transit security
Information moving between your browser and the cloud processor travels through encrypted channels using standard transport security protocols.
Processing runtime
The foundation model evaluates your text in temporary memory to calculate a response, clearing that working memory once the output returns.
Storage and retention
Prompt history, user logs, and cached context remain stored inside your tenant boundary subject to your internal document retention rules.
Model isolation
Commercial vendors contractually exclude business queries from being recycled into public foundation model training sets.
Under the hood: the technical path of an enterprise prompt
When an employee in Wellington or Brisbane types a prompt into a commercial system, what actually happens on the network? Following this technical sequence allows non-developers to speak with technical authority.
Consider an analyst using Microsoft Copilot to summarise a confidential tender proposal stored in SharePoint. The user clicks submit. The prompt travels through an encrypted session to the organisation tenant boundary. Before the text ever reaches an inference engine, an orchestration service queries the enterprise graph to confirm the user possesses legitimate file read permissions.
The orchestration engine pulls excerpts from the approved document and wraps them with the user query into a single bundled prompt. This expanded payload travels over private backbone connections to the model hosting enclave. The model reads the context, generates the summary text, and immediately purges the active prompt from its working memory. At no point does the model retain the confidential file excerpts to retrain its general weights.
Understanding this architecture helps when planning practical AI strategy and consulting across regulated industries. When meeting formal governance needs, teams often pair these internal controls with independent verification tools such as contract obligation intelligence to confirm compliance across suppliers.
Meeting regional privacy expectations
In New Zealand, agencies must comply with the Privacy Act 2020 and guidance issued by the Office of the Privacy Commissioner regarding cloud services and automated processing. The primary questions focus on cross-border data transfers, secure handling, and ensuring personal information is retained only as long as necessary.
- Trace the processing boundary Identify which geographical cloud regions process your prompts and whether data sovereignty clauses apply to your organisation.
- Classify allowed information Establish clear categories for public, internal, sensitive, and restricted operational documents before enabling team access.
- Verify commercial agreements Check that user accounts sit under signed enterprise agreements rather than individual consumer payment terms.
- Audit active logging Confirm who inside your organisation has access to employee prompt audit trails and security monitoring records.
- Document human sign-off Record the verification steps completed by subject matter experts before generated material enters official reports.
Audit your team data path.
Take one frequent team task involving client text. Trace where that text is pasted, which subscription tier covers the user, and verify whether public model training is turned off in the account settings.
Evaluate vendor data terms.
Copy this prompt into your analysis workspace to extract key governance clauses from vendor service terms.
Act as an enterprise technology risk reviewer for an Australian and New Zealand professional services business. Review the attached vendor terms of service for an AI productivity tool. Identify the specific clauses that address: 1. Model training on customer data; 2. Data retention periods for prompt text; 3. Encryption standards in transit and at rest; 4. Geographical regions used for computation and storage; 5. Third-party subprocessors. Summarise findings in a clean table and highlight any points requiring clarification from IT security.
Secure your team workflows.
Moving from individual experiments to embedded team capability requires clear operational guardrails and informed governance.