← Back to Copilot Adoption Hub

Author: Lukáš Barda · Updated November 9, 2026

Prompts and templates from the From emails to an AI agent workshop for managers. No more generic presentations: you find recurring tasks in your own work, pick the ones with the biggest impact and turn the best one straight into a Copilot agent prototype – with an owner and a success metric.

Before you start:
  • Run the prompts in M365 Copilot Chat with Work IQ turned on. Only then can Copilot see your emails, files and meetings.
  • Copilot only goes through your own mailbox and data you have access to. Every participant gets their own list of opportunities.
  • The goal is not to pick the most interesting technology, but a problem that really keeps coming back.
  • Always ask Copilot not to make anything up and to mark uncertain conclusions as UNVERIFIED.

Tip: After pasting, replace the text in [square brackets] to fit your situation.

The manager’s role in AI adoption

Before you look for use cases, talk it through briefly as a team. The outcome should be a list of 3 to 5 main barriers, each with a suggestion for how you as a manager could reduce or remove it.

  • What stops people from using AI regularly today?
  • Where can you remove a barrier as a manager?
  • Do you already have agents? Which ones, and who uses them?
  • How do you know you need an agent – and what kind?
1

Finding opportunities in your email

Your inbox is the best map of repetitive work. Let Copilot go through the last two weeks of email and find topics, questions and manual tasks that keep coming back. The main working prompt has two iterations.

Iteration 1 · Find recurring topics

Go through my work emails from the last [14] days and find topics that keep recurring. Focus on questions, problems or manual tasks that could be candidates for an AI use case.

Then continue in the same chat:

Iteration 2 · Candidate table

Put the output in a table: Topic | Frequency | Current process | Estimated time required | Possible AI solution | Data needed | Risks Do not make up missing information. Mark unverified conclusions as UNVERIFIED.

The result is a list of repeated questions, problems and manual work – raw material for use cases, not a finished solution.

2

From problem to use case

From the table, everyone picks 2 to 3 candidates and adds who would use the solution, what the input is, what output they want and how often the situation occurs. The result is a short use case card.

Use case card

For the topics [topic 1], [topic 2] and [topic 3] from the table, prepare short use case cards. For each one, list: – the user (who will use it), – the input (what it starts from), – the desired output, – the frequency (how often the situation occurs). Do not guess anything the emails don’t show – mark it as UNVERIFIED.

A good use case has a clear input and a clearly expected output. If you can’t describe them, it’s not a use case yet.

3

Prioritizing use cases

Score the candidates together. Each criterion gets 1 to 5 points:

  • Impact
    How much time or work it saves
  • Frequency
    How often the situation occurs
  • Reach
    How many people it helps
  • Feasibility
    Is the data needed available?
  • Risk
    Data sensitivity, need for review, chance of errors

Prefer scenarios with high impact and frequency but low to medium complexity.

Help with scoring

Score these AI use case candidates: [list of use cases]. Give each one 1 to 5 points for impact, frequency, reach, feasibility and risk (for risk, 5 = low risk). Put the result in a table sorted by total score and add a one-sentence justification for each score. Recommend 1 to 2 use cases to start with. Where you are not sure, mark it as UNVERIFIED.

Copilot helps with a draft, but the team decides the order. You know what really repeats in your work.

4

Agent design template

Turn the winning use case into an agent design as a group. Go through nine questions – and only then build the agent (in Copilot Chat via Create agent) and try it on sample inputs.

  1. User
    Who will use it?
  2. Problem
    What repetitive work does it solve?
  3. Trigger
    When does the user need it?
  4. Inputs
    Emails, documents, notes, pasted data
  5. Instructions
    What should it do step by step?
  6. Output
    Table, summary, draft reply, action plan
  7. Check
    What must it mark as UNVERIFIED?
  8. Limits
    What must the agent not do?
  9. Measurement
    How much time does it save, how do we judge quality?
Draft agent instructions

Help me design a Copilot agent for this use case: [use case description]. Go through the template with me and ask about anything you don’t know: – User: who will use it? – Problem: what repetitive work does it solve? – Trigger: when does the user need it? – Inputs: emails, documents, notes or manually pasted data? – Instructions: what should the agent do step by step? – Output: a table, summary, draft reply or action plan? – Check: what must it mark as UNVERIFIED? – Limits: what must the agent not do? – Measurement: how much time will it save and how do we recognize a good result? At the end, write finished instructions for the agent, 3 sample inputs and the expected output for each.

Quick prototype test

Here are my agent’s instructions: [paste instructions]. Suggest 5 test situations – 3 typical and 2 edge cases where the agent could make a mistake or overstep its limits. For each, describe the result we expect.

The prototype doesn’t have to be perfect. It has to be testable: clear inputs, a clear output and clear limits.

5

How we measure impact

Without measurement, a prototype never becomes a tool the team actually uses. Set the current time, the target saving, output quality and frequency of use – and agree who will test the agent, by when and how.

Metric and test plan

Suggest a simple plan to verify the impact of the agent [agent name]: – how long the task takes today, – what the target saving is, – how we recognize a good output, – how often the agent should be used, – who will test it, by when and how they will share feedback. Put the output in a short table.

Every agent leaves the workshop with an owner, a deadline and a next step. Otherwise it stays just a nice idea.

Questions that help you find a use case

If your email doesn’t point you anywhere, try answering these questions – on your own or with your team:

  1. Which task do you do repeatedly every week?
  2. What questions do people keep sending you over and over?
  3. Where do you most often look for information?
  4. What output do you create repeatedly in the same structure?
  5. Where do you manually retype or summarize information?
  6. What takes 30 minutes today and could take 5?
  7. Which process has a clear input and a clearly expected output?
  8. Which decisions must always stay with a human?

Frequently asked questions

What is a Copilot agent?

An agent is a customized version of Copilot for one specific repetitive task. It has its own instructions, data sources and limits – for example, it drafts a reply to a frequent question or summarizes materials in a fixed structure. You can create a simple agent right in Microsoft 365 Copilot Chat.

How do I find a good use case for an AI agent?

Start with your own inbox: let Copilot go through the last two weeks of email and find recurring questions, problems and manual tasks. A good candidate repeats often, has a clear input and expected output, and doesn’t work with overly sensitive data.

Which agent should I build first?

Score the candidates 1 to 5 on impact, frequency, reach, feasibility and risk. Pick a use case with high impact and frequency but low to medium complexity – not the most technologically interesting one.

How do I measure the impact of an AI agent?

Before launch, write down how long the task takes today and set a target saving, output quality criteria and expected frequency of use. Name an owner who will test the agent and share feedback by a set deadline.

Lukáš Barda – AI and Microsoft Copilot trainer. He helps companies and their teams bring AI into everyday work in a practical, clear way without unnecessary jargon. AI and Copilot training