We once had a Copilot so efficient, it wouldn’t stop working.

Literally.

It was supposed to respond to updates in a SharePoint list. Instead, it triggered itself… over and over. A classic case of "too helpful for its own good." By morning, we had chewed through 10,000 messages. The logs looked like a drum solo to be honest.

So this was a simple, production ready Copilot, and the mistake? Entirely avoidable — if we had estimated usage beforehand.

Now, we don’t start a single project without the Copilot Studio Estimator.

And this is why,

If you're new to Copilot Studio, here’s the part no one emphasizes enough:

Every Copilot interaction consumes messages. And messages = money.

You don’t pay per user or per flow. You pay per message, and those rack up based on what your Copilot touches.

Things like:

• Every time it talks to a user
• Every time it grounds a response using SharePoint
• Every time it runs a flow or calls a connector
• Every time it thinks

What makes it tricky? These messages don’t feel like costs while you’re building. You’re adding features, making it smarter, improving UX well all good intentions. Until you deploy and start getting usage reports that look like:

“You’ve exceeded your consumption quota by 14,352%.”

So what is the Copilot Studio Estimator?

It's a free, still-in-preview tool from Microsoft that helps you forecast how many messages your agent will consume before you ship it. this thing gives you a gut check.

No guesswork. Just sliders, dropdowns, and message breakdowns.

Let’s break it down.

What Actually Affects Your Message Count?

So. Many. Things.

Here’s a non-exhaustive list of culprits that affect your monthly message volume:

1. User Count & Frequency

• 500 users logging in once a week?
• 50 power users running it 20 times a day?

That’s the base multiplier. Always start here.

2. Knowledge Sources

• Tenant graph grounding = ~10 messages per query
• SharePoint, OneDrive, web = ~2 messages

Yes, just searching your own docs eats messages. That’s not a bug — it’s part of how grounding works.

3. Orchestration Style

• Classic: like Power Automate with structured flows
• Generative: powered by AI models like GPT-4o or GPT-4 Turbo

Guess which one eats more? Yep, generative orchestration is message-hungry, especially if you’re chaining prompts together or handling ambiguity.

4. Triggers (especially Autonomous)

This is where we got burned.

Autonomous triggers like:

• Scheduled flows (e.g., run every 15 mins)
• File added/modified
• Item updated in a list

They’re awesome… until they become loops. And then they’re expensive. 😅

5. Actions and Agent Flows

• Each step — like sending an email or updating a record — = messages
• Compound flows? More steps = more messages

Don’t assume a short flow is a cheap flow.

6. Model Tier

Let’s talk GPT:

Choosing the premium model is like ordering caviar when you just need a sandwich. Great power, but know the cost.

🛠 How the Estimator Helps You (For Real)

This is not a fancy spreadsheet. It’s an actual web tool where you plug in your use case:

1. User base – how many people? how active?
2. Orchestration type – classic or generative?
3. Triggers – are you running on a schedule or event?
4. Knowledge grounding – what’s your data source?
5. Model tier – which GPT flavor?
6. Actions/flows – what’s the agent doing?

On the right, it will show you:

→ Estimated monthly message count, broken down by:

• Knowledge
• Actions
• Triggers
• Modifiers

And yep, it’ll show you how each change affects the total.

🤔Keep in Mind…

This tool is incredibly useful — but let’s set expectations:

• It’s not a billing tool
  You won’t get pricing or license info, just message projections
• Usage can vary
  This is a model, not a guarantee — but it’s damn close
• Still in preview
  Expect improvements. Soon, it might pull telemetry automatically or integrate directly into Studio

How to Use the Estimator (Step-by-Step)

Here’s how we use it at Armely — every single time:

1. Open the estimator: https://microsoft.github.io/copilot-studio-estimator
2. Choose your estimation type
   • If you’re in early stages, use telemetry-based defaults
   • If you have numbers, go manual
3. Fill in agent traffic
   • How many users?
   • How many interactions per user per month?
4. Pick the agent type
   • Internal or customer-facing
5. Set orchestration mode
   • Generative = more messages, better UX
   • Classic = cheaper, more rigid
6. Add knowledge sources
   • SharePoint, web, tenant graph?
7. Include any triggers/actions/prompts
   • Especially autonomous flows — be honest
8. Pick your model tier
   • Start with GPT-4o Mini unless you have a very good reason to go premium
9. Review the result
   • See what’s eating up your messages
   • Adjust until you’re happy (or at least not shocked)
10. Screenshot & share with stakeholders
    It’s a fantastic conversation starter and budget alignment tool.

Final Thoughts

The Copilot Studio is a project-saving, budget-protecting, sanity-preserving tool that every builder should use.

We’ve learned the hard way what happens when you deploy without estimating. Now we design with numbers, not just hopes.

So before you launch your next Copilot:
✅ Check your triggers
✅ Pick your model wisely
✅ Run it through the Estimator

Future you (and your finance team) will thank you.