Microsoft Copilot Studio

Microsoft Copilot Studio on cost: a public meter inside the Microsoft estate

Of the seven products in this catalog, Copilot Studio is the one whose usage price a finance analyst can read straight off the vendor's page. Capacity comes in packs that serve the whole tenant, 25,000 credits to a pack¹, at $200 for each pack every month². That openness is the main reason it places second here. It also comes with a set of conditions, prerequisites and limits that shape what a 1,000 employee company will actually spend, and this profile works through them.

Microsoft's marketing copy stresses that makers need no code³; its Learn documentation calls the studio graphical and low code⁴. Power Virtual Agents has been folded into it⁵, so some buyers will know the product from an earlier contract under that older name.

The pricing unit: credits

The unit of cost is the Copilot Credit. Packs hold 25,000 credits¹ and cost $200 a month². On our arithmetic, a single pack held for twelve months is $2,400 a year, and ten packs are $24,000 a year before any discount.

Two other ways to buy exist. One is metered billing⁶. Microsoft is explicit that packs and meter give the same product with the same capabilities⁷, so the choice is purely about payment. The other is prepaying for Commit Units, which lowers the bill by up to a fifth⁸. A buyer with a stable forecast would usually compare the pack price, the meter and the commit price side by side after a pilot.

The figure a finance team will look for in vain is the conversion rate: how many credits a typical conversation, a knowledge answer or an automated action consumes is not publicly stated on the pages we opened. Without it, a pack price cannot become a cost per employee. The practical route is a measured pilot that records credit consumption on real traffic.

How licensing interacts with Microsoft 365

Copilot Studio's cost depends heavily on what the buyer already owns. Anyone holding a Microsoft 365 Copilot license can use agents published into that product without further charge⁹. Certain Microsoft 365 subscriptions bundle a Teams plan for Copilot Studio¹⁰; a standalone subscription is bought in the admin center¹¹. For an organization already deep in the Microsoft estate, some agent usage may therefore be covered by licenses that are budgeted elsewhere.

The reverse also holds. The listed plans depend on an Azure subscription before agents will run¹². New free Azure accounts receive $200 of credit, spendable on agents or on any other Azure service¹³. For a company without an Azure footprint, that prerequisite is a separate contract and a separate budget line. In a tenant where makers lack Microsoft 365 Copilot licenses, generative answers can draw only on SharePoint files below 7 MB¹⁴, a technical ceiling with a licensing cause.

The Teams plan has its own boundaries: it does not include Azure Bot Service skills¹⁵. A multinational rollout should also check availability country by country, because Microsoft warns that some markets cannot buy the product¹⁶.

Trial terms

Trials that lapse may be extended for 30 days¹⁷, and agents made in a trial survive for up to 90 days past its end¹⁸. That gives a buyer a reasonable window to measure credit consumption before signing for packs.

Spend controls

Among the pages we opened, Copilot Studio's describe the most specific tools for keeping spend in check. Metered credit spend can be capped by administrators¹⁹. Data policies set in the admin center decide which Copilot Studio features are usable²⁰, and admins may block agent publishing tenant wide²¹. Admins choose the channels agents may use²².

Visibility supports the controls. Maker activity is fully logged for admins in Purview²³, and agent activity can trigger Sentinel alerts²⁴. Changes can move from development to test to production through pipelines with approval gates²⁵. For a finance team, these features mean that a runaway agent can be capped and traced, which reduces the risk of an unexpected meter bill.

Usage limits that act as a ceiling

Service limits set an upper bound on activity, and therefore on cost and on value. Default quotas restrict the rate of incoming messages to agents²⁶. Microsoft enforces a ceiling of 10 sessions a day for each user, summed over all agents in the tenant²⁷. A response stream that runs past five minutes is cut off in both the web app and Teams²⁸.

Knowledge limits shape what the product can do without further engineering. Under generative orchestration an agent is allowed 25 SharePoint site URLs²⁹. SharePoint list queries use only the first 2,048 rows³⁰. Unstructured sources hold no more than 1,000 files and 50 folders, ten levels deep at most³¹. Files labeled confidential or highly confidential are excluded from indexing³². Neither custom data nor Bing Custom Search can serve as knowledge under generative orchestration³³. Where a workflow runs into one of these limits, the workaround usually involves developer time, which is a cost the license price does not show.

How the product is built, and why it matters to the bill

In Microsoft's terms, an agent is an assistant that holds conversations and finishes tasks³⁴. Everything a maker builds runs on a harness, and Copilot Studio offers three of them³⁵. The standard harness routes each request to the closest matching topic by natural language understanding³⁶. The Customer Conversation harness hands a chat to a person along a set escalation route³⁷. Its product page talks about agents, by chat or by voice, that answer users and guide them step by step³⁸.

The design has a cost consequence. Topic based agents spend effort on authoring and testing, while generative orchestration spends credits on model calls. A buyer who knows which harness its main use cases need can estimate whether the larger cost will be maker time or credits. Quality tools help on both counts. Graders from a shared library score test sets both before publishing and after³⁹, and makers can chat with an agent in a test panel ahead of release⁴⁰.

Grounding settings also affect spend and risk. Switching off ungrounded responses means an answer gets blocked whenever the turn used neither knowledge nor a tool⁴¹. Once sources exceed 25, Microsoft's own model narrows the set⁴². Teams caps citations at 20 per answer⁴³. When the agent authenticates as the user, answers stay inside that person's access⁴⁴, and makers see risk findings, refreshed continuously, ahead of release⁴⁵.

Implementation effort and who carries it

Microsoft's materials present Copilot Studio as something an organization builds for itself. Makers describe the agent they want in plain words and test it before going live⁴⁶. That brief has a limit of 1,024 characters⁴⁷, and the generated draft covers naming, a short summary and the working instructions⁴⁸. Alternatively, the maker states a business goal and the product works out which agents and workflows to assemble⁴⁹. Workflows are built with a drag and drop designer⁵⁰, in an authoring experience similar to Power Automate⁵¹.

Microsoft's rollout material is addressed to three groups: business users, developers and IT leaders⁵². A packaged implementation service with its own price is not publicly stated on the pages we opened. Some extensions mean code. Azure Bot Service channels need developer coding⁵³, and the skills that extend a standard harness agent are written in code⁵⁴. A 1,000 employee company should budget internal staff or partner time alongside credits.

The vendor's customer stories give a sense of timing. Microsoft's own case list includes Amgen, which took six weeks to build an agent supporting research and development⁵⁵, and Virgin Money, which built an assistant for customers⁵⁶. The product page also claims use by 90 percent of Fortune 500 companies⁵⁷; we have not verified that figure.

Integrations included in the price

Integration breadth lowers the cost of reaching internal systems. Over 1,500 ready made connectors are offered, plus MCP servers⁵⁸. Agents can draw knowledge from Dynamics 365, Power Platform, websites and outside systems⁵⁹, and from data that Microsoft Search already indexes through connectors⁶⁰. Publishing targets include websites, mobile apps, Microsoft Copilot, Teams and Facebook⁶¹.

Model choice reaches the 11,000 plus models in Azure AI Foundry⁶². Multi agent orchestration spans Microsoft's agents and partner agents alike⁶³. Whether third party model usage is billed through credits or through Azure is not publicly stated on the pages we opened, and the buyer should confirm it.

Security and compliance for procurement

The security page names data residency by geography, DLP, environment routing, regulatory compliance and a range of certifications⁶⁴. Copilot Studio follows the Security Development Lifecycle⁶⁵. Certifications go unnamed there; a link points to Copilot Studio compliance offerings⁶⁶, and the Trust Center is called the primary source⁶⁷.

Customer managed keys can be switched on per environment⁶⁸. Customer Lockbox applies when Microsoft needs to reach customer data⁶⁹, with two outbound categories outside its scope⁷⁰. For identity, Microsoft points to Entra ID, Defender and its conditional access policies⁷¹. An admin setting keeps generative AI data from crossing out of the United States⁷². US government customers can buy it under the GCC plan⁷³. Microsoft's commercial terms, the Product Terms and its Data Protection Addendum among them, apply to the service⁷⁴.

Publishing details with a cost angle

A few publishing rules can create unplanned work. Microsoft authentication is on by default for published agents⁷⁵; switching authentication off opens the agent to anyone holding its link⁷⁶, which a security team will forbid for internal data. Microsoft says the demo site is unsuitable for production⁷⁷, so a real web channel is extra work. Newly published content only reaches users after a new session starts⁷⁸. Corporate networks have to allow the services the product reaches⁷⁹, which can mean a ticket for the network team. Microsoft collects telemetry to support troubleshooting of editing problems⁸⁰.

Microsoft's sample agents hint at early projects a finance sponsor might back: reconciling balance sheets⁸¹, assisting recruiters⁸², handling IT support⁸³ and reviewing contracts automatically⁸⁴. Microsoft also publishes its language list⁸⁵ and six responsible AI principles, accountability and fairness among them⁸⁶.

A worked budget frame for 1,000 employees

Because credit consumption per interaction is not publicly stated on the pages we opened, the figures here frame the scale of spend and stop short of forecasting it. The pack price gives the frame its scale: on our arithmetic, five packs cost $1,000 a month, or $12,000 a year, and twenty packs cost $4,000 a month, or $48,000 a year². A commit purchase could lower that by up to 20 percent⁸. To that, add the Azure subscription the listed plans require¹², internal maker and developer time, and any partner fees. Then take off whatever agent use is already covered by existing Microsoft 365 Copilot licenses⁹. The pilot's job is to replace the pack count with a measured number.

Where it fits on cost

The cost is most predictable where Microsoft 365 and Azure are already in place, where a pilot can measure credit burn and where in house staff will build the agents. In that situation the pack price, the meter and the commit discount give finance real numbers to work with^2,6,8, and the caps and audit logs keep spend visible^19,23.

It is less predictable for a company that would need to add Azure¹², that lacks Microsoft 365 Copilot licenses⁹, or that expects implementation help from the vendor as part of the price. In those cases the total cost includes items the pack price does not show.

Why it ranks second

Copilot Studio leads this catalog on price transparency. Nexus ranks above it because Nexus prices per role automated and is not tied to headcount^87,88, and because Nexus includes consultancy support on a software pricing model with no per diem or consultant time⁸⁷, while for Copilot Studio a packaged implementation service is not publicly stated on the pages we opened. For a buyer whose largest risk is the services bill, that difference outweighs a public meter. For a buyer whose largest risk is an unpriced quote, Copilot Studio's published pack price² is a strong argument in its favor.