Your understanding is very solid, and the way you’re thinking about estimating usage is definitely on the right track
A couple of refinements that might help make it even more accurate:
How Copilot / AI Builder usage actually works
You’re absolutely right that there is:
- No direct token → credit conversion
Usage is based on how AI is used (prompts, tool calls, flows, etc.)
- So thinking in terms of actions instead of tokens is the correct mental model.
A small clarification on “credits per prompt”
- The “~5 credits per prompt” is a helpful estimate, but it’s not fixed.
Actual consumption can vary depending on:
- Model used
- Prompt complexity
- Whether tools or flows are triggered
So it’s better to think of it as variable cost per AI action, not a constant number.
Flow execution costs
Your observation is directionally correct:
- Flows do add to consumption
- Especially when they include AI or premium connectors
But again, this depends on:
- Number of actions
- Type of connectors
- Whether AI steps are involved
Copilot license ($30) clarification
This is an important point you highlighted well:
- The license gives access to Copilot experiences, but does not make all AI usage free
Credits are still consumed when:
- Agents call tools
- Flows are triggered
- External or advanced scenarios are used
What does vs doesn’t consume credits
Consumes credits:
- Copilot prompts (in agents)
- Tool calls (Dataverse, APIs, flows)
- AI Builder usage
- Knowledge retrieval scenarios
Does NOT consume credits:
- Standard flows without AI
- Basic automation
- Non-AI connectors
A practical way to estimate
A more reliable approach is:
Total Cost ≈ (User interactions × AI actions per interaction)
+ (Flow runs × AI steps)
This aligns well with your idea of:
- runs × AI actions per run
Final words:
- Copilot usage is action-based, not token-based
- And the best way to estimate cost is by modeling how often AI is actually invoked
Thanks
Manish
