Copilot Studio agent cost calculator

1. Load your exports (optional)

No files loaded. Drop an export to auto-fill the numbers, or just type them in below.

Where the cost comes from

Every dollar of the monthly total, by component - so nothing is hidden.

• Generative answers$64 66.7%
• Classic answers$32 33.3%

Cost optimizations

Concrete ways to spend less, with estimated monthly savings. Levers overlap - don't add them up.

• Right-size the plan instead of buying full prepaid packs$104 / month

  1 full prepaid pack(s) cost $200/mo; a hybrid of 0 pack(s) + pay-as-you-go is $96/mo, so you stop paying for unused capacity.

• Use classic answers for deterministic, high-volume responses$32 / month

  A classic answer is 1 credit vs 2 for a generative one. For each generative answer you can make a deterministic authored response, you save 1 credit. Upper bound shown (if every generative answer were deterministic).

• Pre-purchase credits (P3) for a discount - and they don't expire$12 / month

  A Copilot Credit Pre-Purchase Plan (P3) buys credits up front at ~6% off pay-as-you-go and, unlike monthly capacity packs, pre-purchased credits do NOT reset monthly - they're drawn down as you consume. Worth it for steady volume (1 pack-equivalent/mo). Discount is an editable estimate; confirm with your contract.

• Set a monthly credit cap per agent (guardrail)guardrail

  In the Power Platform admin center (Licensing > Copilot Studio > Manage Agents), set a per-agent monthly cap around 14,400 credits (~$144, about 1.5x expected usage) so a misconfiguration or spike can't run up unbounded overage.

What people forget

Cost levers ranked by impact. Highlighted ones apply to your current setup.

• $$$User license mix (dominant lever)applies

  0% of users are treated as zero-rated (M365 Copilot-licensed on M365 surfaces). Every metered user instead costs ~$0.48/month in credits. This single input changes the bill by orders of magnitude.

• $$$Bring-your-own (Foundry) model is per-token

  A bring-your-own / Foundry model bills per token in Azure - NOT in Copilot Credits and NOT zero-rated by M365 licenses. It frequently dominates the bill, so model choice and token volume matter more than any credit lever.

• $$$Tenant graph grounding is 10x

  Microsoft Graph grounding costs 10 credits per message vs 2 for a plain generative answer - a 5x premium. Use it only where M365 data is genuinely needed.

• $$$Model tier adds per-token billing (100x spread)

  The selected model (GPT-4.1 mini (default)) is on the basic tier. Token billing rates: basic 0.1, standard 1.5, premium 10 credits per 1K tokens - a 100x spread from cheapest to most expensive. Premium/reasoning models add this ON TOP of the feature rate; consider a lower tier for routine turns.

• $$Autonomous triggers bill per run, never zero-rated

  Event-triggered / autonomous runs bill per run under a service identity, so they are always metered (M365 licenses don't zero-rate them). Unbounded triggers are a classic cost surprise - add guard conditions and batch.

• $$Classic flows break - convert to agent flows

  Classic Power Automate cloud flows called by an agent bill via Power Platform request limits and need a Power Automate Premium ($15/user) or Process (~$150/flow) license - a seeded M365 license is not enough, so flows using premium connectors fail. Fix: convert them to agent flows (Copilot Credits, premium connectors allowed, no PA license).

• $$AI-tools tier is per-token (10x across tiers)

  AI tools / prompts bill per 1K tokens at 0.1 (basic) / 1.5 (standard) / 10 (premium) credits - a 100x spread. Premium is for advanced reasoning only; default to the lowest tier that meets quality.

• $$Azure / Dataverse TCO beyond credits

  Custom Azure AI Search SKU, Application Insights ingestion above the 5 GB/mo free tier (~$2.53/GB), and Dataverse storage overflow (5 GB DB / 20 GB file / 2 GB log per pack) are billed separately from Copilot Credits.

• $Generative answers cost 2x classicapplies

  A generative answer is 2 credits; a classic (authored) answer is 1. For deterministic, high-volume responses, classic answers halve the per-answer cost.

• $Quota / throttle ceilingsapplies

  Concurrency is capped before cost is: 1-10 prepaid packs = 50 RPM / 1,000 RPH for generative AI messages; PAYG and M365 Copilot users = 100 RPM / 2,000 RPH. High-burst agents can hit the wall regardless of budget.

• $Content processing is 8 credits/page

  Standard SharePoint/Dataverse RAG indexing is free. Running documents through content-processing AI tools is 8 credits/page - don't pre-process the whole corpus; only process documents that need it.

• $SharePoint RAG feasibility limits

  1,000-file cap and 512 MB/file limit per source; files 7-200 MB need an M365 Copilot license + Work IQ to be answered; Confidential / password-protected docs are silently skipped. Indexing itself is free - cost is only at query time.

• $Multilingual per-conversation uplift

  The agent supports 1 languages; instruction/context tokens are effectively re-sent per language, applying a modelled uplift to the model-token-driven credits.

• $$License vs. metering crossoverapplies

  A user crossing ~3000 credits/month is cheaper to license ($30/mo) than to meter. This profile: ~48 credits/metered-user/month -> recommendation: meter-users.

• $Fair-use caveat on zero-rating

  Zero-rated M365 Copilot usage is subject to Microsoft fair-use limits; it is not an unlimited allowance for high-volume scenarios.