Blog. Sylox perspective.
Adding a new AI data source changes the review
Co-founder and CEO · November 6, 2026 · 3 min read

On this page
When a team connects a new data source to an internal AI assistant, the assistant keeps its name. What it can see, and what it can say, changes. Adding a source is a governance decision, and it deserves the review the assistant had when it was first switched on.
Support tickets change the audience for the same assistant
An internal assistant answers employee questions from public product documentation. Everything it can read is already on the company website. The support team asks for more: connect the assistant to the support platform, so it can answer from past tickets.
Nothing has gone wrong, but the questions have changed. Tickets hold customer names, email addresses and account details, and some carry files customers uploaded. The assistant’s audience is every employee, including people who would never get access to the support platform itself.
Write the boundary down before and after
If you can’t describe the old boundary, you can’t say what the change adds. NIST’s Generative AI Profile (AI 600-1) lists what an inventory entry for such a system can include. Two items are “data provenance information” and “human oversight roles and responsibilities,” and the profile adds special considerations for personal, privileged, proprietary or sensitive data. That entry is your before page.
| Item | Before | After |
|---|---|---|
| Sources | Public product documentation | Public documentation and the support ticket archive |
| Data owner | Head of marketing | Head of marketing for documentation; director of customer support for tickets |
| Sensitivity | Public | Customer names, email addresses, account details |
| Access design | No sign-in needed to read the sources | Connector account reads all tickets; user permissions in the source are not applied |
| Audience | All employees | All employees, narrowed to the support and success groups |
| Output handling | Answers kept 30 days | Answers with customer details redacted; prompts and answers kept 30 days |
| Left out of the index | Not applicable | Attachments and payment fields |
The approval line completes the page: who signed, on what date, and which changes trigger the next review.
The data owner decides whether tickets may be used this way
The support platform has an owner. That person decides whether ticket data may be used for this purpose, and the assistant team does not.
The purpose question comes first. Australia’s privacy regulator, the Office of the Australian Information Commissioner, says privacy obligations apply to any personal information input into an AI system and to output that contains it. Under Australian Privacy Principle 6, personal information may be used only for the purpose it was collected for, unless an exception applies.
Tickets were collected to resolve support requests. Answering an unrelated employee’s question is a different use. The data owner and the privacy lead settle that before the connection is made.
Test the credentials and the output before go-live
- Which account does the connector use? If it reads every ticket, every user of the assistant borrows that reach.
- Does the assistant apply each user’s own permissions in the source, or one shared account’s? Ask the vendor, then test it.
- What can an answer reveal? A summary can surface a customer’s name when nobody asked for it.
- Where are prompts and answers stored, who can read them and for how long?
Then test with the questions a curious employee would ask, such as what a named customer complained about. Check what comes back, and record the test, the result and who signed it off. Write down which ticket fields may be indexed, which employee groups may use the feature and what the assistant should decline to answer.
Agree the triggers for the next review
The OAIC says due diligence for AI products “should not amount to a ‘set and forget’ approach.” Agree the triggers in advance.
- A new source, or a new type of record in an existing source.
- A new audience for the assistant.
- A change to the account the connector uses.
- A change in where prompts and answers are kept.
- A vendor change to how or where the model runs.
IRIS doesn’t inventory AI tools, prompts or outputs, or map findings to AI regulations. For which AI tools staff use, your web proxy or CASB logs show the traffic. IRIS doesn’t connect to Copilot.
IRIS can check the source you plan to connect. For connected sources and detected data types, it records where sensitive and personal data sits, with a scan date. See discovery and classification.
Write the before page for the assistant you already run
Fill the table above for your own assistant this week, starting with the Before column. The first gap you find is the first question to settle with the data owner. The AI assistant oversharing checklist covers what to check in the sources, and the glossary entry on AI governance defines the discipline.
Sources
- NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), action GV-1.6-003 (July 2024; opened November 9, 2026)
- Office of the Australian Information Commissioner: Guidance on privacy and the use of commercially available AI products (published October 21, 2024, updated January 17, 2025; opened November 9, 2026)
Start with a free assessment of one database you choose.
Get a free assessmentRelated resources
Protect your brand. Start with your data.
Book a demo, or start with a free assessment of one database.


