Licensing and pricing for Copilot Studio agents

Most agent projects are stopped by a licensing question, not a technical one. It is cheaper to answer it in week one.

The three meters#

• Copilot Studio messages. Prepaid packs consumed per interaction; generative answers and actions cost more messages than a simple topic.
• Power Platform capacity. Dataverse storage, premium connectors and per-user or per-app plans for anything that touches Power Apps.
• Azure consumption. Model tokens in Azure AI Foundry, Azure Functions executions, storage and logging.

Estimating before you build#

1. Take an expected conversation and count the interactions end to end.
2. Multiply by the message weight of each step, not by the number of users.
3. Add token cost for grounding and summarisation separately; retrieval payloads dominate it.
4. Model two volumes: pilot and steady state. Only steady state decides the plan.

A per-conversation cost figure with a stated assumption list survives procurement. A monthly lump sum does not.

Where the cost actually goes#

Deterministic work is cheap in code and expensive in the conversation layer. Moving lookups and validations into Azure Functions usually lowers message consumption while making behaviour more predictable.

Practical guardrails#

Cap conversation length, cache grounding results, and log message consumption per topic so the expensive path is visible before the invoice is.