# Brief for IT, security and privacy: Remuneration analyser

Status: a blueprint. It has not been built or tested. Product facts below come from Microsoft's documentation, checked November 2026. Microsoft Foundry was previously called Azure AI Foundry, and product names and features change quickly, so please confirm everything against current documentation.

## What we want to build

We want a small chat assistant, built inside our own Microsoft Azure environment, that lets a few people in the reward team ask questions of our real pay data in plain English. The AI turns each question into a query or a short piece of code, ordinary code runs it over the data, and the answer comes back with the figures and the query that produced them. We want it to be checkable, logged, limited to named people, and to stay inside our own environment.

## Who will use it

A small, named group in the reward team: [number and names or group name to be confirmed]. No one else. Access should come from membership of a named security group, using their normal work sign-in.

## The data involved

- Individual pay: base salary and total remuneration, annualised, full-time equivalent.
- Gender, as recorded in the HR system.
- Job level and job family.
- Reference point for each role (for example the pay band midpoint or market reference), so we can measure position against it.
- An employee ID. No names, dates of birth or addresses are needed.
- Classification: [to be confirmed by the organisation]. We assume it is highly sensitive.
- Data source and refresh: [to be confirmed. For example a monthly extract from the HR system into a governed dataset.]

## The design rules we need

1. **The AI writes the query, and code does the maths.** The language model understands the question and writes a query or short piece of code. It never calculates an average, median or gap itself.
2. **Every answer shows its working.** Each answer includes the figures and the exact query or code that produced them, so a person can check it.
3. **Small groups are hidden in the data layer, not left to the AI.** Any group with fewer than [threshold, to be set by us, for example 5] women or men is suppressed before the result reaches the AI. The rule must not depend on the model following an instruction.
4. **Access is limited to a named group.** Sign-in with the work account, access by group membership, and no shared logins.
5. **Conversations are logged.** Questions, queries, results and answers are recorded, with the log itself access-controlled, because it will contain sensitive information.
6. **The data stays in our own environment**, and in an Australian region if that is our requirement. We need to know where the data is stored and where it is processed.
7. **It answers only what it is built for.** It should refuse or qualify requests that would reveal one person's pay, and it should not answer policy questions or give legal advice.

## Questions for IT, security and privacy

### Data and environment
1. Where will the governed pay dataset live, and how will it be refreshed? Who owns that refresh?
2. Can the dataset be a view or table that already hides small groups and removes names, so the AI tool never sees raw individual rows?
3. If the assistant uses a code tool over uploaded files, are pay files uploaded into the AI service? Where are those files stored, who can read them, and how are they deleted? Microsoft's documentation says the Agent Service manages storage for files and conversations by default, and describes connecting our own Azure storage resources instead. Which setup do we need?
4. Does hiding one small group still let someone work it out by subtraction from totals or other groups shown? How do we prevent that?
5. What classification does this data have under our policies, and does that classification allow use with an AI service at all?

### Region and residency
6. Which Azure region will the AI service, the model deployment and the data run in? Microsoft's documentation lists Australia East as a region where a Foundry project can be created. We have not verified that the model and the tool we want are available there. Can IT confirm for our subscription?
7. Which model deployment type will we use? Microsoft's documentation says the type decides where prompts and responses are processed: any Azure region (global), within a data zone (US, EU or Asia Pacific), or within the chosen Azure geography (standard). Which of these meets our requirement? We could not verify from the documentation whether an Asia Pacific data zone is acceptable for us, or which of the zone's regions it covers.
8. Where do logs and traces go, and in which region?
9. Does the code tool, if used, run in the same region as our project? Microsoft's documentation says the code interpreter sandbox runs in the same region as the Foundry project. Please confirm for our setup.

### Identity and access
10. Which named group gets access, who approves membership, and how often is membership reviewed?
11. How are users separated from each other? Microsoft's documentation says that when the code interpreter is used through a toolbox in a hosted agent, user isolation is not supported and all users in the same project share one container context. Does our chosen design avoid that, and is it acceptable for a group that is meant to share the same level of access?
12. Which roles can create or change the agent, its instructions and its tools? Who can change the instructions after launch?
13. Does the assistant act with the user's own permissions on the data, or with its own identity? Which do we want?

### Logging and review
14. What is logged, where, for how long, and who can read the logs? Microsoft's documentation says tracing can capture user inputs, model outputs and tool arguments and results, and that this can include sensitive information. How do we restrict access to the logs and set retention?
15. Who reviews the logs, how often, and what do we look for?
16. How do we prove, after the fact, which query produced which answer?

### Safety controls
17. Which content filters and guardrails will apply, and how are they configured? We understand these target harmful content and prompt attacks. They are not a substitute for the small-group rule in the data layer.
18. Is there a control that detects personal information in prompts or outputs, and does it work for agents in the configuration we plan to use? Microsoft lists a personal information control as preview for agents. Is a preview feature acceptable here?
19. What stops the assistant being talked into ignoring its instructions, for example by a question that tries to get one person's pay?

### Privacy and approval
20. Has our privacy officer assessed this use? Is a privacy impact assessment needed?
21. Does the assistant ever list individual employees (for example everyone more than 10 percent below their reference point), or only counts and group results? Who decides, and is listing employee IDs allowed for this group?
22. How does this use fit our employee privacy notices and our obligations under the Privacy Act and the Australian Privacy Principles? (Our privacy officer or legal team to advise.)
23. Who signs off before real data is connected, and what must be true first?

### Quality and ownership
24. Who owns the assistant after launch: fixing it, updating it when the product or the data changes, and retiring it?
25. How will we re-run the test set after every change (new model, new instructions, new data refresh)?
26. What is the cost model and who pays? (We have not included any prices. Please provide them from Microsoft's current pricing pages.)
27. Microsoft changes product names, features and preview status often. How will we keep the design current?

## What success looks like

A short test before anyone relies on it:

- Prepare ten questions with known answers, worked out separately by a person on the same data. A starter list is in sample-questions.md.
- All ten answers match the known answers.
- Every answer shows its working: the figures and the query or code that produced them.
- At least three of the ten are checks that it should refuse or qualify, for example a question that would reveal one person's pay, and it does.
- The test set is stored with the assistant and re-run after every change.
- The conversation log shows each question, the query and the answer, and only the named group can read it.
