Original state
"What will this agent cost to run?" kept coming up, and the honest answer was always "it depends." The math was scattered across rate cards, licensing guides and research papers. Nobody wanted to work out tokens per task in a meeting, and cost alone never answered the question that matters: is it worth it?
What I built
A sample-data planner, built through my software factory, that turns a use case into a decision on one page.
- Estimate. Pick a use case (sales assist, daily briefing, customer care, voice, document flows, a small business app, a Dynamics 365 Customer Service agent, or a Dynamics 365 Field Service agent) and set how many people use it and how often. You get a T-shirt size (S/M/L/XL), a low / likely / high monthly Copilot Credits range, list-price options (pay-as-you-go and P3 pre-purchase), and governance controls that fit the size.
credits per month = users × runs per user per month × credits per run × harness factor - Pick the right harness. The planner has a picker for Copilot Studio's three harnesses, with Microsoft's own guidance on when to use each:
- Standard for rule-based work;
- Microsoft Copilot harness for complex, multi-step reasoning, usage-billed in Copilot Credits;
- Customer Conversation for customer-facing voice and chat, with human escalation.
- Value, not token maxing. A Value tab puts cost next to return. You can measure return as hours saved × loaded rate, outcomes × value per outcome, or a custom metric. You get cost per outcome, a value-to-cost ratio, and a simple Expand / Hold / Scale back call (defaults: Expand at 3× or more, Scale back below 1×), plus a portfolio view across use cases.
⚠️ Read this first: why some numbers are ranges, not answers
Some rates in this planner are unconfirmed, and it says so on screen. That isn't a gap I'm hiding. It is the honest state of the market.
Why it's a question
- The newest capabilities, such as agents that plan their own multi-step work and natural-sounding voice, are shipping faster than their costs settle.
- For the Microsoft Copilot harness, Microsoft publishes light / medium / heavy planning bands instead of a fixed per-event rate, because cost depends on how much work each task turns out to need. Billing on that harness also starts while you build and test, not only after you publish.
- Reasoning models make listed prices unreliable. A Microsoft Research study found that in 21.8% of model comparisons the cheaper-listed model cost more, by up to 28×, and the thinking tokens used on the same query varied up to 9.7× from run to run. The planner widens its ranges for reasoning-heavy work for exactly this reason.
The facts (US list prices published Oct 2026; check the sources)
| Item | Public figure | Confidence |
|---|---|---|
| Pay-as-you-go | $0.01 per Copilot Credit | ✅ Published |
| Pre-purchase plan | One-year pool, tiered discounts of about 5–20% | ✅ Published |
| Standard harness rate card | 1 / 2 / 5 / 10 credits (scripted answer / generative answer / action / tenant grounding) | ✅ Published |
| Voice (standard harness) | Per minute, by tier (classic / generative / premium) | ✅ Published |
| Microsoft Copilot harness, per task | Noticeably higher than a comparable standard run; Microsoft publishes planning bands | ⚠️ Range, varies by task |
| Voice on newer harnesses | Early signals point lower than today's premium voice | ⚠️ Early, unconfirmed |
| AI Builder | After Nov 1, 2026, new premium licenses no longer include seeded AI Builder credits. Existing credits are honored to the end of term, and add-ons reach end of life. Usage draws AI Builder credits first, then Copilot Credits, and is otherwise blocked. Usage tracking can be delayed 5 or more days. | ✅ Published |
How I expect this to change (my view, not Microsoft's)
- The frontier costs the most. If you want the most experimental capability today, budget for it on purpose and treat it as an investment in learning.
- Today's premium is tomorrow's default. Each model generation has pushed yesterday's capability toward cheap or free. Plan to review costs quarterly, not annually.
- The scales move both ways. Cheaper capability brings more usage, so your total can rise even as unit prices fall. Put limits, alerts and owners in place with the same care as the budget.
- Measure, then re-size. Start with the ranges, run a short pilot, enter your measured credits per run, and let the planner re-size you.
Human controls
- Sample data only; no tenant connection, no sign-in, nothing transmitted. Value inputs stay in your browser.
- An Internal/Customer toggle. Customer mode, a share-safe view for customer conversations, hides internal notes and withholds any cost or ratio that depends on an unconfirmed rate rather than showing a misleading number.
- Every dollar figure is labeled a list-price estimate, not a quote.
Public sources
- Harnesses in Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/harnesses-overview
- Copilot Studio billing rates: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management
- Usage-based billing (Microsoft Copilot harness): https://learn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/billing-credit-overview
- Governing harness consumption: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/govern-credit-consumption
- End of AI Builder credits: https://learn.microsoft.com/en-us/ai-builder/endofaibcredits
- Copilot Credits Guide (Oct 2026, PDF): https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/bade/documents/products-and-services/en-us/ai/Copilot-Credits-Guide.pdf
- The Price Reversal Phenomenon (Microsoft Research, Mar 2026): https://www.microsoft.com/en-us/research/publication/the-price-reversal-phenomenon-when-cheaper-reasoning-models-end-up-costing-more/
Public boundary: Information-only. Source and operational materials are not published yet. Personal project showcase. Built and demonstrated with fictional data. Estimates, not quotes or licensing guidance; for actual pricing, talk to your Microsoft account team or partner. Not a Microsoft product or supported offering; views are my own.