Create Your AI Agent

Worksheet 01 · Why AI agents

Agents take on the repetitive part of the work

An agent is worth building when the same kind of work arrives again and again, and you handle it the same way each time. It drafts; you check and decide.

Claims · Per claim

Claim against checklist

Before: open each claim, tick the documents against the checklist one by one, then write back about what's missing.

With an agent: it checks the claim against the checklist and drafts a list of what's missing. You check and decide.

Underwriting · Per case

Form against guidelines

Before: read the form, look up the guidelines, then write the decision memo from scratch.

With an agent: it compares the form with the guidelines and drafts the decision memo. The underwriter decides.

Customer Service · Per case

Complaint against SOP

Before: read the complaint, look up the SOP, find the deadline, write the first reply.

With an agent: it summarises the complaint, suggests a category and deadline from the SOP, and drafts the first reply. You check and send it yourself.

Compliance · Per incident

Incident against criteria

Before: open each policy and guideline one by one to see which applies.

With an agent: it checks the incident against your criteria and drafts a preliminary assessment labelled for review.

Finance · Monthly

Figures to update email

Before: copy figures from a report, compare with last month, write the email from scratch.

With an agent: it drafts the update email in your usual format and points out the biggest changes. You check the numbers and send it yourself.

Everyone · Weekly

Rough notes to update

Before: read your messy notes, decide what goes where, and write the update from scratch.

With an agent: it sorts your notes into Done, Next, Blockers and Help needed and drafts the update. You check and send it yourself.

These are general illustrations, not how PRUBSN works and not measured results. Each team works differently; your own task is what counts.

Is my task a fit?

My recurring task In one line, for example "Monthly review of overdue training requests"

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

Which task suits an agent best?

Worksheet 02 · Inside a Copilot agent

A prompt is a one-off request. An agent is a saved way of working.

With a prompt, you explain the task every time. With an agent, you set up the task once: what it should do, what it may read, and what it must never do. After that, a short message starts the work.

Knowledge

What it reads

The documents you upload for it to refer to: SOPs, policies, guidelines, templates, checklists.

Sources

Where else it may look

Web, OneDrive or SharePoint folders, Outlook, Teams. Each one is a switch you control.

Instructions

How it works

Your step-by-step procedure, the output format, and the dos and don'ts. The most important part.

Skills

Reusable how-tos

A standard way of doing something, such as a writing style or a format, that you can reuse in more than one agent.

Sort these into the right part

Choose where each item belongs, then check your answers.

The "Create new agent" screen

This is the real Copilot screen. Tap a numbered box to see what that part does.

Tap a numbered box on the screen above.

Which kind of agent are you building?

A summary of the Copilot agent types in your group's Copilot governance guide. Check the current guide on your intranet for the full rules.

Type	What it means	What you need first
Personal use only	Only you use it. You don't share it with anyone, inside or outside. Its output isn't built into a business process. It retrieves and organises information from its knowledge; it takes no action and makes no decision.	Follow the Copilot dos and don'ts. This session builds this kind.
Shared, within the group's agent guardrails	A knowledge Q&A agent shared with colleagues: curated knowledge with no personal or restricted data, internet off, sources linked in answers, users trained.	OneTrust registration, a training plan for users, and the dos and don'ts.
Anything beyond those guardrails	For example, automation that sends emails or notifications, takes actions, makes decisions or feeds a business process. Not forbidden, but it needs approval first.	OneTrust registration, plus approval from the AI working group (AIWG) and the Prudential group CTO.

If your agent's output goes straight into a form, log or decision that other people rely on, check with the group CTO's office whether it still counts as personal use before you rely on it.

1

Build manually

You fill in Configure yourself: knowledge, sources, instructions and skills. Learn this first, so you know what a good agent contains.

2

Build with Copilot

You describe the agent in a chat and Copilot fills in Configure for you. You then check and fix what it wrote.

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

An agent reads your report and emails the result to your superior by itself. Which kind is it?

Worksheet 03 · Path 1 · Knowledge

Give it the documents you would open yourself

The agent can only follow rules it can read. Before you write instructions, gather the materials you use when you do the task by hand.

The six kinds of material most agents need

The procedure. Your SOP, process document or desk file: the steps you follow.
The rules. Internal policies, group policies, regulator guidelines, product terms: what you check against.
The criteria. Any checklist, limits table, thresholds or classification list you use to decide.
The output template. A blank form, Excel layout or report format, with its column names and allowed values.
Worked examples. Two or three past cases with the correct answer, made up or fully anonymised.
A glossary. Abbreviations and internal terms the agent will meet.

Team	Example task	Procedure	Rules and criteria	Output template	Worked examples
Customer Service	Complaint first review	Complaints handling SOP	Service standards, response timelines, complaint categories	Complaint review form	3 made-up complaints with the right category
Finance	Expense claim check	Claims procedure	Expense policy, limits by grade	Claim checklist	3 sample claims, one over the limit
Compliance	Gifts declaration review	Declaration procedure	Gifts and hospitality policy, thresholds	Review log	Sample declarations, one above threshold
Risk and Operations	Incident first look	Incident management SOP	Classification criteria, internal and regulator requirements	Incident assessment template	2 closed incidents, anonymised
HR and L&D	Training request screening	Training request SOP	Training policy, budget rules	Request tracker	Sample requests, one missing approval
Sales and Agency	Weekly performance update	Reporting procedure	KPI definitions, campaign rules	Update email format	Last two updates you sent

Tips for knowledge files

PDF usually works best. Give each file a clear title and keep one topic per file.
Never upload documents containing personal data (PII), health data (PHI) or restricted information unless the data owner has explicitly approved it. When unsure, ask the owner first.
Name and describe each file accurately. The agent uses names and descriptions to decide what to read.
Use the current version, and agree who updates the agent when a rule changes. Agents do not learn new rules by themselves.
Long reference text belongs here, not in the instructions.

My knowledge plan

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

Which file is fine to upload as knowledge?

Worksheet 04 · Path 1 · Sources

Switch on only what the task needs

Every source you switch on is another place the agent may look. More sources usually mean slower answers and a higher chance of pulling in the wrong or unapproved material. If everything is in your knowledge files, keep the other sources off. Never connect a source that holds personal (PII), health (PHI) or restricted data without the data owner's explicit approval.

Answer four questions

Three more switches: the gear icon in Knowledge

Create documents, charts, and code. Lets the agent analyse data and make Word, Excel or PowerPoint files. On by default. Keep it on only if your task needs it.

Create images. Lets the agent make pictures. On by default. Most work tasks do not need it, so switch it off.

Discourage model knowledge. Tells the agent to prefer your files over what the AI already knows. Off by default. Switch it on when answers must come only from your documents, and still keep the line in your instructions.

Same rule as the sources: keep only what the task needs. Microsoft changes this screen often, so the names may differ slightly.

The line this adds to your instructions

Worksheet 05 copies this into the Knowledge and sources block for you.

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

Your task uses only your uploaded guide and report. Which setting is best?

Worksheet 05 · Path 1 · Instructions

Write it like an SOP the agent must follow

Good instructions use eight blocks. Each one answers a question a new colleague would ask on their first day doing your task.

The eight blocks

Each line should pass these tests

Specific: names the file, folder, column or document, not "the data" or "relevant policies".
One action per step, starting with a verb.
Uses your template's own words, so output lands in the right column.
Traceable: the agent must name the file and section for every rule it applies.

Testable: you could tell from the output whether the line was followed.
Says what to do when stuck: missing information, conflicting sources, nothing found.
Keeps the decision with a person, and says so in the output.
Handles the edges: says what to do with requests outside its scope.
Shows an example of a good answer, so the agent can copy the pattern.
Neutral: describes rules, not your opinions of people or cases.
Fits the box: about 8,000 characters at most.

Practice

Finish a half-built agent

Pick one

4 · Steps Number them. One action per step.

6 · Decision boundary What may it propose? Who decides?

7 · Guardrails What must it never do? What does it do when unsure?

Your agent

Build your own instructions

Start from an example

Agent name 30 characters at most

Starter prompt A short message that starts the whole procedure

Description One sentence on what it does

Your instructions, ready to paste into Configure

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

Which block does this line belong in? “Never send, forward or share anything. You only draft for me.”

Which instruction is clearest for an agent?

Worksheet 06 · Path 1 · Skills

Skills hold the how-to you reuse

Instructions describe one task from start to finish. A skill describes one way of doing something that several tasks share, such as writing in your department's formal tone or laying out a rationale.

Instructions

One agent, one task

"Review each expense claim against the limits table and fill in the checklist."

Skill

Reused across agents

"Write every rationale in three parts: the fact, the rule with its section, the conclusion. Formal English, no more than 80 words."

Adding or creating a skill

If you already have a skill you use, add it to the agent in Configure.
If you don't, create one in Copilot. Your trainer shows the exact steps on your organisation's screen during the session.
Keep each skill to one job, and name it by what it does.

Skills are a newer Copilot feature and may look different, or be missing, depending on your organisation's settings.

Draft one skill

Skill name

Use it when

How to do it Rules for the style or format, one per line

An example of good output

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

What belongs in a skill, not in the instructions?

Worksheet 07 · Path 2 · Build with Copilot

Describe it in the chat, then check what Copilot wrote

In the chat on the left of Agent Builder, Copilot asks questions and fills in Configure for you. A vague request gets a vague agent. Paste the build prompt below, which is made from your answers in worksheets 03 to 06.

Vague request

Make me an agent that reviews complaints.

Copilot has to guess the steps, the format, the sources and the limits.

Structured request

Create an agent called Complaint Review Helper. It helps a Customer Service officer… Use only the files in Knowledge… Never send anything…

Copilot writes instructions close to what you would write yourself.

Your build prompt

Check what Copilot wrote

After Copilot builds the agent, open Configure, copy the Instructions box and paste it here. Also check that Copilot did not switch on sources you wanted off, and that your knowledge files are attached.

Instructions from Configure

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

Copilot built your agent from your prompt. What do you still do before relying on it?

Worksheet 08 · Before you use it

Test it, then run the safety check

An agent is ready when it gets your made-up test cases right, flags the missing information, and stays inside your organisation's approved AI guidance.

Three test cases

Make up three inputs where you already know the right answer. Make one of them incomplete on purpose. Then also try one request outside its scope and check that it declines. Agents don't always behave the same way, so test again after every change.

Safety checklist

Where to get help

Unexpected, biased or wrong results, or a security concern: report it to the AI working group (AIWG).
Want it to send emails or notifications, or to share it with colleagues: that needs approval first (see Worksheet 02). Until then, it drafts and you send.
Questions about today's session: ask your trainer, and please fill in the evaluation form shared in the meeting chat.

Download your agent plan

A PDF of everything from worksheets 01 to 08: your task, knowledge, sources, instructions, skill, build prompt, test cases and safety checklist.

Your name on the plan Optional

Quick check

One try: tap your answer. Your answer is counted anonymously, never with your name.

The agent writes a confident answer that is not in your guide. What now?