This Copilot Studio Agent Lives in my Inbox
- vikashsingh01
- Jul 17
- 6 min read
Table of Contents
Introduction
Like many of you, Copilot has transformed the way I work - whether it’s recapping meetings, speeding up PowerPoint creation, or handling repetitive tasks. It’s brilliant. But there’s one thing I’ve always wished it could do better: write emails like I do (some might say that’s not a bad thing)!
Sure, I can give Copilot an example email as a reference, but that takes time - and it doesn’t remember my tone, phrasing, or intent the next time I write an email. So, I decided to fix that by building my own autonomous auto-drafting email agent in Copilot Studio.
Now, as soon as an email lands in my inbox, my agent kicks in and drafts a reply. It’s trained on:
20+ examples of my writing style
What I define as “good”, “better”, and “best” emails
Tactful language for navigating tricky conversations
A cheat sheet for hyper-efficient writing (focus on value)
It mirrors my tone; straight to the point, genuinely helpful, and just charming enough to get a reply… most of the time.
The best part? I actually use it and there is a clear ROI benefit. This isn’t just a proof of concept, it’s now part of how I work.
If I had to categorise this agent, I’d say it sits somewhere between a Task Agent and an Autonomous Agent. It automates a repetitive workflow with full autonomy, analysing inbound emails and drafting replies without needing input, but it doesn’t yet orchestrate or coordinate with other agents.
Configuration of the Agent
Like most agents built in Copilot Studio, the Instructions are the soul of how it works. This is where the magic happens, and where most of the success (or failure) lies. If your outputs aren’t quite right, don’t abandon the build too soon as it’s usually down to how well the instructions are guiding the agent.
Before I dive into that in more detail, here’s a quick overview of the rest of the configuration:
Orchestration = Enabled
From testing, disabling orchestration sends the agent back to the “classic” Power Virtual Agents experience - rigid, predefined responses.
Enabling it is like saying: “Here are your tools, here are your instructions - now figure out the best way to respond.” It gives the agent the freedom to act more intelligently.
Topics
While this agent can technically respond to chats, that’s not what it’s designed for. So I didn’t bother configuring any Topics (i.e. conversation paths).
Triggers & Actions
This part came together quickly. All I did was add the appropriate ones I needed (shown below), went through the basic setup and it was done. No heavy configuration needed beyond that.
Knowledge Sources
I used four lightweight Word documents based on the themes I mentioned in the introduction. These are grouped using the new Collection of Documents feature (currently in preview). It’s hard to quantify exactly how much this feature improves the replies, but they certainly don’t get in the way.
It's worth noting that I didn’t create these documents upfront. I built the core experience first, and added the knowledge sources later to refine tone, structure, and response accuracy. This is like “fine-tuning” in Azure AI Foundry, even though Copilot Studio doesn’t support this (yet).
Instructions of the Agent
The Instructions are the backbone of this agent. They define its tone, its decision-making, and ultimately whether the outputs get you what you were after. Copilot Studio gives you a lot of flexibility here - and that’s both a gift and a trap. You can write a vague prompt and get vague results. Or you can go deep, specific, and intentional.
I kept the Description basic. It gives the agent just enough purpose to stay on track:
The Instructions is where I went all in. I wrote them like I was onboarding a real teammate: clear, specific, and no fluff. Step-by-step guidance, tone cues, fallback logic, and what to filter out i.e.what not to reply to - like meeting invites or marketing emails. This took the most effort, but made the biggest impact.
Here’s what it looks like 👇
Example Replies
I tested it on three types of emails:
Type | Incoming Email | Auto Drafted Reply | Agent Response Highlights |
Casual | “Hey, can you send over the deck?” | “Sure, I can send you the deck. Please find it attached.” | Quick, friendly, and direct. Matched tone perfectly. |
Technical | “What’s the impact of enabling environment routing in Dataverse?” | The reply was pretty good for the technical information provided. The formatting admittedly wasn’t the greatest, but quick enough to fix manually. There’s a screenshot example of this later in this post. | Mirrored my voice, added a polite caveat, and used placeholders like “[confirm if this applies]”. |
Functional | “We’d like to explore training options.” | “I’d be happy to help you explore training options. Could you please provide more…” | Positioned a helpful reply, added clear CTAs, avoided over-explaining. |
ROI (Time Savings)
Let’s be blunt: this will save me hours per week.
Another bonus: the mental load of having to start a reply from scratch is now gone too.
Type of Email | Manual Effort (Before) | With the Agent | Time Saved |
Simple Acknowledgements | "Thanks, received" / "Noted, will do" | Drafted instantly with correct tone & context | ~30 seconds per email |
Average Workflows | Scheduling, status updates, quick answers | Structured, relevant, on-brand and draft ready | ~3–5 mins per email |
Heavy Hitters | Sensitive, tricky, or technical conversations | Tailored tone + tactful phrasing + fallbacks | ~10–15 mins per email |
Total Weekly Time | ~2-4 hours (depending on volume of emails & complexity) | Review & send, instead of write from scratch | Easily 2+ hours/week |
And because the agent only triggers on eligible emails, I’m not burning messages unnecessarily. FYI: Triggers are a billable feature.
New Learnings
If you test a Trigger without publishing your agent it will fail. You will get the error message ‘The response is not in a JSON format’. Don’t make this same mistake.
It seems like the agent didn’t work instantly when I had the trigger and action configured, I seemingly had to ‘jump start’ it by asking in the chat ‘what triggers and actions are you configured with. There is also a chance this is just confirmation bias to me and I just needed to wait a few minutes.
Whilst configuring Triggers and Actions in Copilot Studio does create Power Automate flows behind the scenes, I also toyed with trying to create an Agent Flow with natural language to see if I could nail both aspects with the one Tool. It likely is possible, but I found that it was slightly fiddly and didn’t quite work as nicely as Triggers and Actions:
Fun quirk: Even though the agent is not configured to properly use the chat feature, I can still ask it directly in chat about topics, and it’ll cook up a draft in Outlook too. Magic.
What it Doesn't Do
It won’t draft the email as a reply to the existing email chain it came from, so there is an element of copy and pasting the email content over to a reply. However, overall it is still faster than not having an agent.
If someone asks for something the agent doesn’t have access to - like my calendar availability - it won’t guess. Instead of writing something vague, it usually drafts something like: “Happy to help, could you confirm what dates you had in mind?”
As seen above, the formatting isn’t the cleanest but I’m sure there is a way to improve on this by adapting the Knowledge Sources and Instructions.
What's Next?
Now that I’ve got my first agent up and running, I’m already thinking bigger. A team of multi-orchestration agents, each with a specialised role - all coordinated by a single Master Agent.
Imagine asking one agent for Power Platform governance advice - and behind the scenes, it taps into a licensing expert, a documentation formatter, or even a tone-of-voice coach. I want to create a system where agents collaborate with each other the same way humans do: passing context, specialising, and working efficiently as a team.
This auto-drafting email agent is just the beginning - the first hire in what I hope becomes a full digital squad.
TL;DR: Most of my replies to emails aren't from scratch anymore. My agent does it, and honestly, it’s probably doing a better job.
Comments