Microsoft Copilot Studio

Microsoft Copilot Studio: the low-code builder for a Microsoft estate

Microsoft describes Copilot Studio in two registers. On the product page it is a no-code place to make and run custom agents, workflows and apps¹. Learn, Microsoft's documentation site, prefers the term low-code and stresses that the studio is graphical, for agents and workflows driven by AI². Microsoft's FAQ adds a third label, an end-to-end conversational AI platform³. All three are accurate in their way, and together they place the product on the build spectrum: a visual studio that a business maker can open, backed by the Power Platform and Azure for the parts that need a developer. Power Virtual Agents, the earlier product, is now included in Copilot Studio⁴.

For a Head of Digital, the first question about Copilot Studio is less about build style than about estate. A company that already runs on Microsoft 365, Entra ID and the Power Platform admin center will find most of the governance it needs already in place. A company that does not will be adopting a large part of that stack to use the builder. This page sets out what the product does, how its licensing works, and where its documented limits sit. It ranks fourth on this site.

The building blocks

Learn defines two core objects. Agents, in Microsoft's definition, are AI assistants that hold conversations and finish tasks⁵. Workflows are automations assembled by dragging and dropping steps in a designer⁶. Agent flows are authored much as Power Automate flows are, though inside Copilot Studio itself⁷, so a team that already builds Power Automate flows will recognize the canvas.

Everything built in Copilot Studio runs on a harness, and the product offers several⁸. Under the standard harness, language understanding picks the topic that best fits each request⁹. The Customer Conversation harness adds a set route for handing a chat over to a human agent¹⁰. These harnesses are a structural choice that a maker makes early, and they shape how the agent behaves.

The newest entry point is goal-driven. Here the maker states an outcome, and Copilot Studio works out which agents and workflows to combine¹¹. Multi-agent orchestration can connect agents from Microsoft or partner providers to work together¹². Model choice includes the 11K+ models available through Azure AI Foundry¹³.

How a first agent is made

Makers open the studio in a browser at copilotstudio.microsoft.com¹⁴; Microsoft asks for Edge at version 122 or newer¹⁵. The first step is to write a plain description of the agent, which can then be tried out ahead of release¹⁶. That prompt has a ceiling of 1,024 characters¹⁷. From it, Copilot Studio drafts a name, a summary and a set of instructions¹⁸. The authoring guide says agents can be created and deleted by describing what you want in your own words¹⁹. The test chat panel lets makers try agents before publishing²⁰, and the quickstart ends by publishing to a live or demo website²¹.

Quality checks sit after the build. Learn points makers to test sets and a common library of graders, used both ahead of release and after it²².

For a no-code maker, this is a short path to a first conversation. The longer path is the one that matters for a Head of Digital: wiring the agent into line-of-business systems, getting it through security review and putting it in front of users. Microsoft's own customer story shows the scale of that path in one case. Amgen, in that story, took six weeks to produce an R&D agent with Copilot Studio²³. Nexus puts the first agent live in 4 weeks inside a 3-month proof of concept, with its consultants building alongside the customer (owner decision²⁴).

Connectors, knowledge and channels

The connector library is large. Microsoft lists 1500+ pre-built data connectors and MCP servers²⁵. Agents can draw knowledge from Dynamics 365, the Power Platform, websites and outside systems²⁶, or from Copilot connectors whose content Microsoft Search has indexed²⁷. If the agent signs in as the user, answers are limited to material that person is allowed to see²⁸.

Finished agents can be published to websites, mobile apps, Microsoft Copilot, and messaging services including Facebook and Teams²⁹. Agents can be published anywhere, including Microsoft Copilot³⁰. Anything else goes through Azure Bot Service channels, a separate route³¹, while Azure Bot Service skills let developers extend a standard-harness agent in code³². This is where Copilot Studio shows its low-code character: the visual studio covers the common cases, and a developer takes over for the rest.

Nexus lists more than 4,000 integrations³³ and deploys to Slack, Teams, WhatsApp, email, phone and web³⁴.

Security and compliance

The security page groups its controls under data residency by geography, DLP, standards certifications, regulatory compliance and routing between environments³⁵. Development follows Microsoft's SDL process³⁶. Instead of naming certifications, that page sends readers to a separate compliance offerings page³⁷ and to the Microsoft Trust Center as the main source for Power Platform compliance³⁸. A security reviewer will need to pull the certification list from those resources.

Customer-managed encryption keys can be enabled for Copilot Studio environments³⁹. Customer Lockbox can govern access to customer data⁴⁰, with two kinds of outbound data left outside its scope⁴¹. For identity, Microsoft relies on Entra ID and Defender together with conditional access rules⁴². The Product Terms and Data Protection Addendum govern the service⁴³. Six responsible AI principles are named: fairness; reliability and safety; privacy and security; inclusiveness; transparency; and accountability⁴⁴.

Governance

Governance is Copilot Studio's strongest area, and it comes from the wider Microsoft platform. From the Power Platform admin center, administrators decide through data policies which Copilot Studio features are allowed⁴⁵, and they can block publishing across the whole tenant⁴⁶. Role-based access is set by environment groups and routing policies⁴⁷. Admins can cap pay-as-you-go Copilot credit spend⁴⁸. Purview holds the audit trail of what makers did⁴⁹, and Sentinel can raise alerts on what agents do⁵⁰. Admins control which channels agents may use⁵¹.

Release management follows enterprise software practice, with release pipelines, approval gates and guided environment-to-environment promotion⁵². Before release, a maker is shown risk findings that refresh as the agent changes⁵³. Above all this sits Microsoft Agent 365, presented as the central place to watch, govern and secure these agents⁵⁴.

Authentication defaults are sensible. Out of the box, each agent requires Microsoft sign-in⁵⁵. Choosing No authentication lets anyone who has the link chat with the agent⁵⁶, which is a setting a central team will want to restrict.

For a Head of Digital in a company that already runs these admin centers, this governance stack is a large advantage, because much of it is already in place.

Data handling and residency

Microsoft says the studio can be aligned with residency requirements and with the main compliance standards⁵⁷. Outside the United States, administrators can stop generative AI features from moving data between geographies⁵⁸. Government customers can buy it in the GCC plan⁵⁹. Telemetry is collected by Microsoft so that it can deal with editing problems⁶⁰. Corporate networks have to allow traffic to the services the studio relies on⁶¹.

Two operating details affect rollout. Microsoft warns that its demo website should stay out of production⁶². A user sees newly published content only once a fresh session begins⁶³.

Licensing and pricing

The licensing model is consumption by credit. Credits are bought for the whole tenant in packs of 25,000⁶⁴, and each pack costs $200.00 a month⁶⁵. A pay-as-you-go meter is also available⁶⁶, and there are no in-product feature differences between the pack and the meter⁶⁷. An Azure subscription is required to use agents on these plans⁶⁸. Paying ahead for Copilot Credit Commit Units can cut the cost by as much as 20%⁶⁹, and new free Azure accounts start with $200.00 of credit⁷⁰.

Several routes lower the cost for some users. Where a user already holds a Microsoft 365 Copilot license, use of agents published there costs nothing extra⁷¹. A Teams-only edition of Copilot Studio comes inside certain Microsoft 365 subscriptions⁷²; skills are missing from it⁷³. A separate subscription can be ordered in the Microsoft 365 admin center⁷⁴. At expiry, a further 30 days of trial can be added⁷⁵; agents built in the trial keep running up to 90 days beyond its end⁷⁶. The product page notes that it is not available in every market⁷⁷.

For budgeting, credits behave like a utility meter. Spend follows conversation volume, which a finance team can cap⁴⁸ but has to forecast. Nexus sets one price for each role automated, with no day rates and no billing for consultant hours (owner decision²⁴).

Documented limits

Microsoft publishes its quotas in detail, which is useful for planning. Default quotas cap the rate at which an agent accepts messages⁷⁸. Each model response stream has a five-minute timeout in the web app and Teams app⁷⁹. Across a tenant, any one user gets 10 sessions with agents in a 24-hour window⁸⁰.

Knowledge limits are the ones a builder will meet first. Generative orchestration allows an agent no more than 25 SharePoint site URLs⁸¹. Makers in a tenant without a Microsoft 365 Copilot license can only use SharePoint files under 7 MB for generative answers⁸². Queries against a SharePoint list stop after its first 2,048 rows⁸³. An unstructured source may hold 1,000 files, 50 folders and subfolders nested 10 deep⁸⁴. Anything labeled confidential or highly confidential is kept out of the index⁸⁵. Answers in Teams display a maximum of 20 citations⁸⁶. An agent with more than 25 sources has them filtered by an internal model⁸⁷, and generative orchestration rejects both custom data and Bing Custom Search as knowledge⁸⁸. Turning off ungrounded responses means a reply is suppressed whenever the turn drew on neither a tool nor a knowledge source⁸⁹.

The confidential-label limit deserves attention. A company whose key procedures sit in documents with those labels will need another route to that content.

Who uses it

Three groups are named as the people who roll it out: business users, developers and IT leaders⁹⁰. Its example agents cover balance sheet reconciliation⁹¹, recruitment⁹², IT support⁹³ and contract review⁹⁴. Chat and voice agents can answer questions and guide workflows in natural language⁹⁵. Customer stories include Amgen²³ and Virgin Money, which built an AI assistant for customer experience⁹⁶. Microsoft's own claim is that 90% of Fortune 500 companies have used it⁹⁷. The product page carries the list of languages it supports⁹⁸.

Strengths

Copilot Studio inherits Microsoft's governance stack: Power Platform data policies, Purview audit logs, Sentinel alerts, Entra ID and release pipelines with approval gates^{45,49,50,42,52}. It offers a large connector library²⁵ and a wide model catalogue¹³. Customer-managed keys and Customer Lockbox are supported^39,40. Its quotas and limits are documented in unusual detail⁷⁸. For a Microsoft-centered company, many users may already hold licenses that cover usage⁷¹.

Limits

The certification list is not on the security page itself³⁷. Knowledge sources carry hard caps, including the exclusion of confidential-labeled documents⁸⁵ and the 7 MB file limit in tenants without Microsoft 365 Copilot⁸². An Azure subscription is required⁶⁸, and channels beyond the standard list need developer work³¹. Credit-based pricing follows usage⁶⁴. Microsoft's pages describe a rollout audience and customer stories^90,23; a buyer who wants a vendor team building alongside its own staff should ask how that is arranged.

Verdict

Copilot Studio ranks fourth. For a Head of Digital in a Microsoft-centered company that wants low-code building under existing governance, it is the natural first evaluation. It ranks behind Nexus because Nexus brings the build team: consultants build with the customer, the first agent is live in 4 weeks inside a 3-month proof of concept, pricing follows the role automated, and Nexus lists SOC 2 Type 2, GDPR, ISO 27001 and ISO 42001 (owner decision²⁴). Nexus also lists more than 4,000 integrations³³ and works across estates, including Teams and Slack³⁴. Relevance AI and Dify rank above Copilot Studio for buyers who want a builder outside a single vendor stack, with published prices^99,100 and, in Dify's case, self-hosting¹⁰¹.