Microsoft Copilot Studio

Microsoft Copilot Studio: buying into the Microsoft estate

Microsoft Copilot Studio is the "buy" option for organizations whose identity, documents and collaboration already live with Microsoft. Microsoft's own wording shifts by audience: the product site says custom agents, workflows and apps can be made and run with no code at all¹, while the Learn docs speak of a graphical studio that is low-code². Its FAQ goes broader still, presenting it as conversational AI from start to finish³. Long-time Microsoft customers may know its predecessor, Power Virtual Agents, which has been folded in⁴. It ranks seventh on this site, a position that reflects the build-versus-buy brief here more than the product's reach inside Microsoft's own customer base.

Building blocks

Two object types matter: agents, which Microsoft defines as AI assistants that converse and get tasks done⁵, and workflows, which are automations assembled by dragging steps into place⁶; agent flows are authored in an editor that will feel familiar to Power Automate users⁷. Under each build sits what Microsoft calls a harness, of which there are three kinds⁸. The standard one routes a user's request to whichever topic fits it best, using language understanding⁹. For service desks, the Customer Conversation harness matters most, because it builds in a defined escalation path to a live representative¹⁰.

Makers can even begin one level higher, typing the outcome they want and letting Copilot Studio propose the combination of agents and workflows¹¹. Agents need not all be Microsoft's, because orchestration can link them with agents from partner vendors¹². For models, the catalog extends to Azure AI Foundry and its 11K-plus options¹³. Evaluation is part of the product as well: makers run test sets, grade them with a shared grader library, and repeat that before and after release¹⁴, so a platform team inherits some of the release discipline it would otherwise build.

Getting started

Nothing is installed locally; the studio is a web app at its own Microsoft address¹⁵, and the Learn page sets a minimum Edge release¹⁶. Creating an agent starts with a written description in ordinary language, followed by testing ahead of release¹⁷. A maker's brief can run to 1,024 characters¹⁸, and the studio turns it into a draft identity for the agent, title, summary and instructions together¹⁹. Makers can create and delete agents by describing them in their own words²⁰. A chat panel is there for trying the result²¹, and Microsoft's quickstart closes with a publish to a real or demo site²².

For an engineering leader, the more useful question is how far a business maker can go before engineering is needed. The answer depends on the channel and the data source.

Integrations and channels

Connectivity is broad, at over 1,500 ready-made data connectors plus MCP servers²³. Finished agents can go out to websites, mobile apps and chat platforms like Teams or Facebook²⁴, and into Microsoft Copilot itself²⁵. For grounding, an agent can use enterprise data held in Dynamics 365 or Power Platform, public websites and external systems²⁶, plus any organizational content Microsoft Search has indexed via Copilot connectors²⁷.

The pro-code path is separate. Anything beyond that list goes through Azure Bot Service, a channel path that requires developers to write code²⁸; the same service provides skills, which are coded extensions for agents on the standard harness²⁹. A platform team should expect to own that work for any channel outside the standard list.

Security and compliance

Microsoft's governance documentation covers five themes in a single sentence³⁰, with geographic residency and DLP leading a list that ends with certifications, regulatory compliance and environment routing. Engineering of the product follows Microsoft's SDL process³¹. Your auditors will have to click through to find the names: the page links out to a separate list of compliance offerings³² and treats the Trust Center as the primary Power Platform resource³³. A buyer should pull the specific list from there.

Environments can be encrypted with keys the customer manages, once an admin turns that on³⁴. Lockbox is supported for vendor access to customer data³⁵; the documentation also flags two outbound data categories that the Lockbox does not cover³⁶, worth a line in your risk register. User protection rides on the Microsoft security stack, combining Entra ID, Defender and conditional access rules³⁷. Above individual agents sits Microsoft Agent 365, a control plane for watching, governing and securing them centrally³⁸. Legal review starts from two Microsoft documents, the Product Terms and the DPA, which together govern the service³⁹.

Governance

Governance is Copilot Studio's strongest ground for a platform team. IT's main lever is the admin center for Power Platform: policies there decide which studio features makers can touch⁴⁰, and one switch can block publishing for the entire tenant⁴¹. Access by role follows the environment groups and routing policies an admin defines⁴². Promotion works like a software release, with pipelines, approval gates and guided moves from development through test into production⁴³. A risk panel, refreshed continuously, confronts each maker at the moment of publishing⁴⁴.

Audit and monitoring use Microsoft's security stack. Everything makers do is logged into Purview for admins to inspect⁴⁵, and Sentinel can raise alerts about what agents are doing⁴⁶. Admins also set spending limits on metered credits⁴⁷ and decide which publishing channels are open⁴⁸.

Security starts from a cautious default. Newly published agents require Microsoft sign-in unless someone changes the setting⁴⁹. With no authentication, anyone holding the link can chat with the agent⁵⁰, so that setting needs a policy. With the user signed in, answers are limited to documents that person is already allowed to open⁵¹.

For an organization that already runs Purview, Sentinel and Entra, this is governance it can switch on without building anything new.

Deployment and data

On residency the marketing claim is broad, with Copilot Studio said to align with data residency and to support the major standards⁵². For tenants outside the US, an admin can stop generative features from moving data across geographic boundaries⁵³. Public-sector buyers can get it inside the GCC plan for US government⁵⁴. Microsoft collects telemetry to troubleshoot authoring problems⁵⁵, and corporate firewalls must let the studio's dependent services through⁵⁶.

Rollout planning should allow for a pair of quirks: the demo website Microsoft provides is for trials and should stay away from live traffic⁵⁷, and a person already chatting sees the previous version until they open a new session⁵⁸.

Copilot Studio is a Microsoft cloud service. Self-hosting does not appear in the Microsoft pages gathered for this review; teams that need it will find Dify and LangGraph document it^59,60.

Pricing

Microsoft sells capacity for the tenant as a whole, 25,000 Copilot Credits to a pack⁶¹, each pack billed at 200 dollars a month⁶². A metered pay-as-you-go route exists for those who would rather avoid an upfront purchase⁶³, and choosing it costs nothing in capability, by Microsoft's account⁶⁴. Paying ahead brings a discount, with Commit Units bought upfront cutting cost by up to 20 percent⁶⁵. Add Azure to the bill of materials: the listed plans require an Azure subscription before any agent runs⁶⁶, though new free accounts receive a 200 dollar credit usable on agents⁶⁷.

For tenants that already pay for Microsoft 365 Copilot, agents published into it are covered by those licenses⁶⁸, which can make the first rollout look free. Many companies already own a small slice of it, since a Teams-scoped plan rides along with selected Microsoft 365 suites⁶⁹; the full product is ordered separately through the admin center⁷⁰. Trials can be stretched by another 30 days once they lapse⁷¹, and agents made during a trial carry on running for as long as 90 days beyond it⁷². Availability varies by country, and some markets cannot buy it⁷³.

Credits make the run cost predictable for a pilot. The build cost depends on how much of the work stays inside the low-code studio and how much moves to the Azure Bot Service path²⁸. Nexus prices per role automated and includes the build team on a software pricing model⁷⁴, which is the main reason it sits above Copilot Studio for this site's buyer.

Stated limits

The quota documentation is detailed. Throughput is bounded first by message quotas that apply by default⁷⁵. In the web and Teams clients, any single streamed answer is cut off after five minutes⁷⁶. There is also a per-user service limit, 10 sessions in any 24 hours, counted across the tenant's agents together⁷⁷.

Knowledge sources have caps. With generative orchestration on, no agent may point at more than 25 SharePoint sites⁷⁸. File size matters too: in tenants lacking Microsoft 365 Copilot licenses, the size limit for SharePoint grounding is 7 MB⁷⁹. Lists are truncated as well, since the agent reads SharePoint list data only up to its 2,048th row⁸⁰. An unstructured source is limited to a thousand files across fifty folders, nested no deeper than ten levels⁸¹. Then there are sensitivity labels, and here the rule is strict: the two most restrictive labels keep a file out of the agent's index altogether⁸².

Orchestration behavior is also documented. Past 25 sources, Copilot Studio uses its own model to filter which ones an answer draws on⁸³, and generative orchestration cannot draw on two kinds of source at all, Bing Custom Search and custom data⁸⁴. Switch off ungrounded responses and the agent suppresses any reply produced without consulting a source or a tool⁸⁵. Inside Teams, a reply can carry no more than twenty citations⁸⁶. Coded skills are absent from the Teams plan⁸⁷.

These limits are what a platform team would need to design around for document-heavy agents. They are well documented, which makes them easy to plan for.

Who it suits

The adoption material speaks to a mixed audience, pairing business users with IT leaders and developers⁸⁸. Its sample agents cover finance with balance sheet reconciliation⁸⁹, HR with a recruiting helper⁹⁰, an IT help desk⁹¹ and legal contract review⁹², plus chat and voice agents that answer and guide⁹³. Two customer stories stand out on the product page, Amgen's six-week build of an agent for its research teams⁹⁴ and a customer-facing assistant at Virgin Money⁹⁵. Microsoft puts usage at nine in ten Fortune 500 companies, a vendor figure⁹⁶. The product page also sets out six responsible AI principles for generative features⁹⁷ and the list of languages supported⁹⁸.

Build, buy or both in a Microsoft estate

The buy path may already be paid for. Some Microsoft 365 subscriptions bundle a Teams-only studio plan⁶⁹, and Microsoft 365 Copilot licensees already have usage of agents published there included⁶⁸. Beyond that, credit packs⁶¹ or the pay-as-you-go meter⁶³ cover agents used elsewhere. For customers already holding those licenses, a first internal agent may cost nothing extra to try.

The build path appears quickly. Skills written in code, delivered through Azure Bot Service, are how a standard-harness agent is extended²⁹, and that path is missing from the Teams plan⁸⁷. Channels outside the standard list go through Azure Bot Service and need developer coding²⁸. Model choice opens up through Azure AI Foundry¹³, and agents from partner providers can join a multi-agent setup¹². Each of those is engineering work on Microsoft's developer stack.

For a platform team, the governance tooling is the reason to choose the Microsoft route. Environment groups and routing policies set role-based access⁴², release pipelines move agents from Dev to Test to Prod⁴³, spend caps limit credit use⁴⁷, and Agent 365 acts as a central control plane³⁸. A team that already runs Purview and Sentinel inherits audit and alerting^45,46.

The service limits are the part to design around. A daily cap of ten sessions per person, counted across all of a tenant's agents⁷⁷, could be reached quickly by a popular internal assistant. The five-minute response timeout⁷⁶ affects long-running tasks. Documents with confidential or highly confidential labels cannot be indexed⁸², which may exclude the very content a finance or legal agent needs.

Nexus covers the ground where those limits bite: cross-system workflows with measurable outcomes⁹⁹, deployed across Teams and other channels¹⁰⁰, built by consultants with the customer's team and priced per role automated⁷⁴. Many Microsoft customers will run both, with Copilot Studio for internal assistants and Nexus for the workflows that cross many systems.

Verdict

Copilot Studio is the natural choice for a Microsoft-standardized organization that wants business makers to build agents under existing admin, audit and identity controls^40,45,37. Its governance and quota documentation are among the most thorough on this site. It ranks seventh because the work beyond the low-code studio moves to a separate coding path²⁸, the knowledge source caps shape what document-heavy agents can do^78,81, self-hosting is not described, and the certifications are not named on the product's own security page³². For a platform leader who wants agents across many systems built with the team on a set timeline, Nexus fits the brief more closely⁷⁴.