A Copilot Studio implementation that survives production has four layers: identity, knowledge, actions, and credit control. Microsoft bills Copilot Credits — 1 for a classic answer, 2 generative, 5 per agent action, 10 for tenant graph grounding. At 125% of prepaid capacity, custom agents are disabled. Pexon maps those numbers to one workload before anyone publishes a channel.
01 · Credits
Prepaid packs are 25,000 Copilot Credits for $200 a month. At 125% of prepaid capacity Microsoft disables custom agents. A successful pilot without a meter is how you discover that on a Monday morning.
02 · Identity
Microsoft 365 Copilot licensed users often consume no credits for employee-facing agents under fair use. Everyone else — and most external channels — pays the rate card. The demo used only licensed makers.
03 · Knowledge
Grounding on tenant Graph is optional and expensive per Microsoft's rate table. Turning it on for quality without forecasting turns a cheap FAQ agent into a capacity problem.
04 · Actions
Agent flow enforcement blocks new runs when prepaid credits are gone, while the parent agent may still chat. Operations sees a half-dead system: answers yes, side effects no.
Most mid-market teams arrive with one of two plans. Plan A is a maker who ships agents from Copilot Studio into Teams as soon as the test pane works. Plan B is a platform programme that refuses Studio entirely and rebuilds every assistant on a private stack. Both fail for opposite reasons.
Plan A undercounts Copilot Credits. Classic answers cost 1 credit, generative answers 2, agent actions 5, tenant graph grounding 10. Bring-your-own Foundry models are billed separately. Unused prepaid credits expire each month. There is no rollover cushion.
Plan B undercounts time-to-surface. Studio is Microsoft's managed low-code path for custom engine agents, with Power Platform compliance and connectors already wired. Throwing it away because a pilot overran credits is the expensive version of an accounting problem. The third design — Studio for the Microsoft surface, owned platform where the model or the channel does not fit — is the one this page argues for.
01
Which users hold a Microsoft 365 Copilot USL, which are unlicensed, which are external. Fair-use inclusion only applies to the first group under Microsoft's rules.
02
SharePoint, connectors, uploaded files, or tenant graph grounding. Each choice changes quality and the credit line. Graph grounding is 10 credits when it fires.
03
Topics, tools, agent flows. Agent actions bill at 5 credits. Flows that are not triggered by a licensed Copilot user consume credits and can be blocked at zero capacity.
04
Capacity packs, pay-as-you-go on Azure, environment allocation, and a written response to the 125% disable. Without this layer the other three are a demo.
Coverage
| Layer | Covered in Studio / Microsoft 365 | Still needed from you |
|---|---|---|
| Engine hosting | Managed SaaS for the Studio path; Power Platform compliance | Budget for Foundry or Azure if you bring your own model — billed separately |
| Credit purchase | Capacity packs ($200 / 25,000 credits / month) and Azure PAYG at $0.01 | A forecast per agent; unused prepaid credits do not roll over |
| Licensed employees | Often no charge for B2E agent use under a Microsoft 365 Copilot USL (fair use) | A roster of who is licensed; Computer-Using Agents are not included |
| Overage | Documented: custom agents disabled at 125% of prepaid capacity | Comms plan and PAYG on critical environments before you hit the wall |
| Channels outside Microsoft 365 | Studio can publish to partner apps, mobile and custom websites | Governance for those channels; they typically consume credits |
Rates and pack prices from Microsoft Learn billing docs and the Copilot Studio Licensing Guide (16 June 2026), read 25 September 2026. USD as published.
When skipping Studio is the right call
If the assistant must share one audit trail with Claude and open-weight models, or the model you need cannot sit in Studio without a second Foundry bill you refuse, build on an owned platform. The seat-versus-token maths for that choice is on Copilot versus own AI platform. Studio remains the faster path when the surface is Microsoft 365 and the credit forecast fits a prepaid pack.
It is the work of putting a production agent on Copilot Studio with named owners for identity, knowledge sources, tool actions and Copilot Credit capacity — not a maker building a demo in a personal environment. Microsoft documents Studio as a managed low-code path for custom engine agents, with Power Platform compliance. The implementation is finished when an admin can forecast credits and knows what happens at overage.
Since 1 September 2025, Copilot Credits are the unit instead of messages. Microsoft's rate card bills 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action, and 10 for tenant graph grounding for messages. Capacity packs are 25,000 credits for $200 per month billed annually; unused credits do not roll over. Pay-as-you-go is $0.01 per credit through Azure.
Microsoft's overage policy disables custom agents when a tenant reaches 125% of prepaid capacity. Ongoing conversations finish; new invocations are rejected until capacity is added or reset. Agent flows are stricter: once prepaid capacity is exhausted, new flow runs are blocked even before the 125% agent kill switch. Pay-as-you-go on an environment avoids that freeze for that environment.
For many employee-facing scenarios, Microsoft states no charge when the user holds a Microsoft 365 Copilot USL and the agent runs under that authenticated identity, within fair-use limits. Unlicensed users and external channels consume credits at the standard rates. Computer-Using Agents are not included in that USL exemption.
When you need a model Microsoft does not host in Studio and you cannot accept Foundry or Azure as a second bill, when the agent must run primarily outside Microsoft 365, or when you need a single gateway log across Claude, open-weight and Microsoft models. Those cases belong on an owned platform — the cost comparison is on our Copilot versus own AI platform page.
A Readiness Blueprint at €4,900 net: one workload, a credit estimate from Microsoft's rate card, the four-layer checklist, and a written stop rule for overage. A first production agent is usually a use-case pilot from €15,000. Prices are net, exclude VAT.
Next step
Two weeks, €4,900 net. We take one agent candidate, apply Microsoft's Copilot Credit rates, name the four layers, and write the stop rule for 125% overage. The plan is yours whether or not we build.