How to prepare AI for sensitive data

Start with the data flow

A local model addresses one part of the architecture. The application also receives files, retrieves data, builds context, stores responses and records events. Each step can create another copy of the content.

Before integration, record what data enters, which services receive it and who can access it. Define retention, deletion and responsibility for each step with the company.

Authorisation belongs in the backend

A hidden interface option does not protect a record. The backend must check the user, workspace and requested action before reading data or starting processing.

The same rule must cover files, queues and streamed responses. A task identifier does not replace authorisation to view its content.

Separate content from observability

Operations must know whether a task failed, which version ran and how long it took. Identifiers, failure codes, usage and duration can provide that information. Define which records actually need the full content.

Keep media in protected storage and serve it through authorised routes. Process files through streams or bounded batches so that upload size does not determine total memory use.

Check the source before showing a suggestion

In a conversation analysis implementation, I used code to check whether a cited excerpt appeared in the attributed person's speech. The system rejects signals without that evidence before displaying them.

That check establishes that the source is present. It does not validate the model's full interpretation. The interface should show the source and leave the decision with the responsible person.

What to bring to an assessment

• Data types, authorised users and retention requirements defined by the company.
• A service diagram and the contracts of providers that receive content.
• Response review criteria and handling of failures, unavailability and deletion.