Ask questions of your real pay data in plain English and get answers that show their working. A blueprint for building it securely with your IT team in Microsoft Foundry (previously Azure AI Foundry).
This is a remuneration health check assistant, built inside your organisation's own Microsoft Azure environment with your IT team. People in the reward team ask questions of the real pay data in plain English and get an answer with the working shown. Microsoft Foundry is the product it would be built in.
One design rule matters most: the AI does not do the maths. The AI turns the question into a query, ordinary code runs it over the data, and the answer comes back with the figures and the query, so a person can check it. Language models are good at understanding a question and unreliable at arithmetic, so the design keeps them apart.
Use this route when you need real, identifiable pay data, answers must be checkable and logged, access must be limited to named people, and it must stay inside your own environment. Do not use it for exploring ideas with invented data (use the pay gap analyser, use case 01) or for policy questions (use the team assistant, use case 05).
Press Play to watch a short conversation with the analyser, or Show all to read it. This is a script, not a live system.
Scripted illustration using the invented workforce from use case 01. Not output from a real build. You can check the same figures in the live analyser in use case 01.
Five steps from a question to an answer you can check. Press Play, or step through at your own pace.
Signed in with your work account. Nothing has been calculated yet.
You ask in plain English. No query language, no spreadsheet formulas. Just the question, as you would ask a colleague.
SELECT gender,
AVG(total_remuneration) AS average_pay,
MEDIAN(total_remuneration) AS median_pay
FROM governed_pay_view
GROUP BY gender
The AI writes a query. It has not touched the numbers. It has only turned your words into instructions that a database can follow.
| employee_id | gender | level | total_remuneration |
|---|---|---|---|
| E1042 | W | Level 3 | $112,300 |
| E1043 | M | Level 3 | $118,900 |
| E1107 | W | Level 4 | $149,500 |
| E1118 | M | Level 2 | $93,800 |
| E1190 | M | Level 5 | $201,400 |
Code runs it on your data. Ordinary code does the arithmetic, the same way every time. The AI is not involved in this step.
Guardrails apply. Small groups are replaced before the result reaches the AI or you. Only the named reward team group can ask at all.
Men $124,423, women $107,860. Median gap 13.0%.
SELECT gender,
AVG(total_remuneration) AS average_pay,
MEDIAN(total_remuneration) AS median_pay
FROM governed_pay_view
GROUP BY gender
(124,423 − 107,860) ÷ 124,423 = 13.3%
An answer you can check. The figures and the query arrive together. A person can run the same query, or redo the sum, and confirm it.
All figures in this walkthrough are invented, for illustration only. Row values are samples.
Unlike the other use cases, you do not build this one yourself. You build the plan, and your IT team builds the system. This section gives you the picture to show them, plain-English descriptions of each part, the decisions you need to make together, a technical panel for them, and a note to send with the brief.
Microsoft's product names and features change quickly. Every Microsoft statement on this page links to the page it came from and is marked checked November 2026. Where we could not verify something, it is a question for your IT team, not a claim.
Read it left to right. Everything inside the dashed line stays inside your organisation's Azure environment.
The same diagram as a list, for screen readers and for printing:
In plain English, with what Microsoft's documentation says about each one.
People sign in with their normal work account, and access is limited to a named group. Microsoft Foundry uses Microsoft Entra (Microsoft's sign-in and identity service) and role-based access control, which lets you control who can create, run and manage agents.
Source: What is Foundry Agent Service, checked November 2026.This is the screen people type into. Your IT team chooses it. Microsoft's documentation describes publishing an agent to a stable endpoint (a fixed address other software can call) and sharing it through Microsoft Teams and Microsoft Copilot.
Source: What is Foundry Agent Service, checked November 2026.An agent is an AI model paired with instructions and tools. It reads the question, decides what to ask for, and calls a tool. Your instructions tell it to write queries, show its working, and refuse certain requests. Microsoft calls the simplest kind a prompt agent: you set instructions, a model and tools, and Foundry runs it.
Sources: What is Microsoft Foundry and Agent Service overview, checked November 2026.The tool runs what the AI wrote. Microsoft describes two options that fit. A code interpreter tool runs Python code in a sandbox (a sealed-off space) over files you supply. Custom tools let the agent call functions or services your team writes. Which is better depends on whether the tool should ever see individual rows. See the IT panel below.
Sources: Code Interpreter and Agent Service overview (tools and toolboxes), checked November 2026.This is your data, prepared for this use. It holds employee IDs instead of names, and it hides small groups before anything reaches the AI. This part is your organisation's design, not a Microsoft feature. Microsoft's documentation says Agent Service can store files and conversations in your own Azure resources instead of Microsoft-managed ones, which matters if you need full data ownership.
Source: Use your own resources in Agent Service, checked November 2026.Microsoft's agent tracing captures user inputs, agent outputs, tool use and results, token use and timing. Traces are stored in Azure Monitor Application Insights. Microsoft warns that traces can contain sensitive information, so the logs need the same access controls as the data.
Source: Agent tracing overview, checked November 2026.Foundry has guardrails that scan prompts and answers for harmful content and prompt attacks (attempts to trick the AI). These are useful, but they are not the small-group rule. Hiding small groups has to happen in the data. Microsoft marks guardrails for agents as preview.
Source: Guardrails and controls overview, checked November 2026.These are business decisions as much as technical ones. Settle them in a meeting before anyone builds anything.
Which system holds the governed pay dataset, who prepares it, how often it is refreshed, and who can see it. Decide which fields go in and which stay out.
Which named group, who approves membership, and how often it is reviewed. Decide whether everyone in the group sees the same things.
Your small-group threshold, whether it may ever list individual employee IDs, and what it must refuse (for example legal advice or policy questions). Write these down so they can be tested.
A test set of questions with known answers, worked out separately by a person. The assistant must get them all right, with its working shown, before it is used. The file sample-questions.md is a starting point.
Who fixes it, who updates it when the data or the product changes, who reviews the logs, and who retires it. Without a named owner, it will drift.
Verified capabilities, each with its source and marked checked November 2026. Names changed during this research: Microsoft's pages say the brand moved from Azure AI Studio and Azure AI Foundry to Microsoft Foundry, and the old Assistants-style agent API to the Responses API (Agents v2) with conversations, items and responses replacing threads, messages and runs. Confirm current names before you build.
Small-group suppression and the removal of names must happen in the data layer. If the code tool ran over a raw individual-level file, the AI could in principle write code that exposes one person's pay, and an instruction in the prompt would be the only barrier. A safer pattern is a custom function or service, written by your team, that queries a governed view and returns only aggregated, suppressed results, with the query text returned alongside. Whether that is feasible is for your team to judge.
The full question list is in it-brief.md.
Download it-brief.md from the last section of this page and send it with this note. Change the words in square brackets.
This is the one use case on the site designed for real, individual pay data, so it deserves the most care. It involves individual pay, gender, level and job family, which people rightly treat as highly sensitive.
Privacy rules for employee records differ between private employers and government agencies, so check with your privacy officer. Start with the OAIC's guidance on the Australian Privacy Principles and on employment and privacy. General information, not legal advice.
Use the brief to start the conversation with your IT, security and privacy teams, and use the questions as the first draft of your test set. Nothing here has been built or tested yet.
What next: Back to all use cases