An assistant that works only from your own policy. It answers with the clause number, drafts replies to managers, and checks proposals against the rules, so the whole team works to the same standard.
Status: the build steps on this page are written from Microsoft's documentation and have not yet been tested by us inside Microsoft 365 Copilot. The replays below were produced by Claude following the same instructions and policy.
Most reward teams have a policy that few people can recall clause by clause. Newer team members search for answers, and experienced colleagues get interrupted to confirm them.
This is an assistant for the remuneration team only. It is an agent: an AI assistant set up in advance with standing instructions and its own documents. It is grounded, which means it answers from the documents you gave it, not from general knowledge.
It does three jobs. It answers “what does our policy say about X?” with the clause number, and says plainly when the policy is silent. It drafts a reply to a manager's pay query in the team's tone. It checks a pay proposal against the policy: what complies, what breaches and what is missing.
It never decides. It prepares the work, which lifts quality and consistency, especially for newer people, and frees experienced people for the judgement calls.
Three conversations with the assistant, using an invented policy for Banksia Example Co (an invented organisation). Press Play to watch each one, or Show all to read it. Every clause number can be checked against the policy.
All names, roles and figures are invented. Read the invented policy to check any clause number yourself.
Five steps from a policy document to a team that works to one standard. Press Play, or step through at your own pace.
3.2Eligible if started on or before 1 March.
4.3No out-of-cycle adjustment within six months of the last review.
5.1Reward team reviews an offer before any figure is discussed.
Agent for the reward team
Give it your policy. Add the final version of your policy, with numbered clauses, as the agent's knowledge. Numbered clauses are what let it answer with a clause number.
Answers from the policy, with the clause number. Says “the policy is silent” when it is.
Drafts a reply to a manager. Outcome first, plain English, clause in brackets.
Checks a proposal in a fixed table, then lists missing information and questions.
Give it standing instructions. You write them once. Microsoft allows up to 8,000 characters of instructions (Microsoft Learn, checked November 2026).
No. They are first reviewed in the next annual review. Clause 3.2
“Employees who start after 1 March are first reviewed in the next annual review.”
Ask, and get the clause. The answer comes with the clause number, so the team member can check it in seconds. If the policy does not cover the question, it says so.
Missing information: the highest position held by an existing employee in the same role and level.
Check a proposal. Every requirement is marked in words: Complies, Breaches or Cannot tell. Missing information and questions for the manager follow the table.
A person on the remuneration team makes the decision.
Reviewed by Date
A person decides. The assistant prepares the work. A member of the team reads it, checks the clauses, makes the decision and records it.
All policy text, roles and figures in this walkthrough are invented, for illustration only.
You can build this yourself if you have Microsoft 365 Copilot and are allowed to create agents. Agents that use your own documents as knowledge need a Microsoft 365 Copilot licence, or a tenant with pay-as-you-go billing turned on (Microsoft Learn, checked November 2026). Your organisation may also limit who can create or share agents. Check with your IT team before you start.
The steps below use Microsoft's Agent Builder, the simple agent maker inside Microsoft 365 Copilot. They are written from Microsoft's documentation. Screens and button names change often, so follow what you are doing rather than the exact labels.
Collect the policy and guidelines the team already works from. Use final, approved versions only. Number the clauses (1.1, 1.2 and so on) so the assistant can quote them. Remove drafts, old versions and anything the team should not rely on.
Keep the documents short and clear. Microsoft recommends concise files for Copilot (Microsoft Learn, checked November 2026).
Paste this into the agent's instructions field. It is 5,276 characters. Microsoft allows up to 8,000 characters of instructions (Microsoft Learn, checked November 2026). Change the organisation name, and the clause numbers if your policy numbers things differently.
Microsoft advises against putting instructions in a SharePoint document to get around the 8,000-character limit. Documents are meant to supply facts, and directive text in them can be blocked or changed at run time. Microsoft Learn, checked November 2026. Microsoft also notes that Copilot moves to newer models automatically, so behaviour can shift over time. Re-run your tests now and then.
Agent Builder has a “Try it” tab where you can chat with the agent while you build it. It is switched on once the agent has a name, description and instructions (Microsoft Learn, checked November 2026).
Use it with ten questions you already know the answers to. Write down the answer you expect before you ask. Two of the ten must be topics your policy does not cover. Here are the types to cover.
The same instructions also work as a project or custom assistant in other AI tools your employer allows.
Your policy and guidelines, which may be confidential. And whatever the team types in, which can include individual pay.
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 instructions with your own policy, not the invented one. Save your policy as a Word or PDF file before you upload it, and test before the team uses it.
What next: Back to all use cases