What is Copilot Studio for retailers in 2026?
What is Copilot Studio used for in a store? Understand agents, catalogue connections, access and governance before choosing a build-and-run approach.
- Published
- Last updated
Microsoft Copilot Studio is a low-code platform for building and running AI agents that answer questions and use connected tools to complete tasks. For a retailer, it can power an assistant on the store’s website. Your team supplies the catalogue connections and decides which customer actions the agent can perform.
The main commitment is ongoing operation. A store agent needs current product information, permission checks for private data and a working route to staff. Building the conversation is one part of that job.
What Copilot Studio is used for
Copilot Studio lets retailers build their own customer-facing or staff-facing agents using a graphical editor. Microsoft’s Copilot Studio overview describes agents that use connected knowledge and tools across websites and other channels.
These illustrative retail uses have different access requirements:
| Retail task | What the agent needs | What the retailer must control |
|---|---|---|
| Help a shopper choose a rain jacket | Product attributes and a catalogue search connection | Relevant products, variant availability and accurate prices |
| Explain a returns policy | The approved policy and any product exceptions | Which version applies and when staff must decide |
| Show a customer’s order status | An authenticated connection to the order system | Customer identity and access to that specific order |
| Request an order cancellation | A tool that performs the permitted operation | Eligibility, customer confirmation and the result in the order system |
A product recommendation reads information. A cancellation changes a commercial record. Treating those as separate capabilities lets you release useful guidance before enabling consequential actions.
Building an agent in Studio also differs from appearing in the consumer Copilot shopping experience. The guide to Copilot as a shopping channel explains that separate route to product discovery.
Agents, knowledge and workflows
A store agent takes a shopper’s request, retrieves relevant information and either answers or calls a connected tool. The retailer designs the permitted route between those steps.
Consider an illustrative question: “Do you have a waterproof jacket in medium for under $150?” The agent needs product attributes to identify suitable jackets. A catalogue lookup must supply the eligible variants and their current price and stock. The response can then explain the options using that returned information.
A follow-up such as “Where is my order?” needs a different route. Authenticate the customer, establish which orders they can access and call the order-status tool. Return only the permitted result. An order number typed into chat is insufficient proof of ownership.
Microsoft’s knowledge-source documentation distinguishes public websites, uploaded documents and authenticated business sources. Its public-website option relies on Bing-indexed content. That makes it useful for published guidance, but an indexed product page is insufficient evidence of stock at checkout.
Give each dependency an owner. The merchandising team maintains product attributes. The commerce integration owner maintains live lookups and handles failed requests. Customer service owns policy exceptions and escalation. The agent owner maintains instructions and tests the complete conversation.
Use a workflow when a task needs a defined sequence, such as checking cancellation eligibility before requesting approval. Keep the actual permission check in the connected service. Instructions telling the agent to protect orders cannot replace access controls.
Access and a first store-agent build
Access requires an eligible organisational account and permission to create agents in the chosen Power Platform environment. A browser session alone does not grant access to the retailer’s data or permission to publish.
Before starting, arrange the following:
- Use a work or school account in the retailer’s Microsoft tenant. Microsoft’s access guidance excludes tenant guest users from Studio authoring.
- Have the administrator configure the required tenant entitlement and assign the author’s user licence. Microsoft documents these separately in its licence and access guide.
- Select an accessible Power Platform environment with Dataverse and the required author permissions. Microsoft’s environment guidance describes the agent author role and environment access process.
- Obtain approved product and policy sources, plus restricted credentials for any test connection. Confirm that organisational data policies allow the intended tools and website channel.
A Studio trial supports building and testing in the test chat panel, but it does not allow publishing. A Teams-only entitlement also differs from the standalone website-building route. The Copilot Studio pricing guide covers licences and consumption costs separately.
The following first build follows Microsoft’s standard-harness quickstart. A harness is the engine that runs the agent. Microsoft documents different capabilities for its newer agent experiences, so keep the selected experience consistent across setup instructions.
- Create the agent in a development environment. Describe its task as answering product and store-policy questions. Review the generated instructions and remove any suggested capability outside that task.
- Add a small approved knowledge set. Use a product category and the relevant delivery and returns policies. Confirm that each source is ready before asking questions about it.
- Connect a read-only catalogue lookup. Define its inputs and returned fields, including variant identity and availability. Confirm that a known product returns the same values as the test catalogue.
- Write the failure and escalation paths. If the lookup fails, the agent must avoid claiming stock is available. A request for a person must reach the configured staff destination.
- Test privately in the authoring panel. Use the acceptance cases below and inspect tool inputs and outputs. Correct the source or connection when the retrieved data is wrong.
- Publish to a controlled website test page once the licence permits it. Configure the website channel and authentication, then repeat the cases through that page. Release the customer-facing embed only after the published version passes.
Microsoft’s website publishing instructions distinguish the demo website from a production embed. The demo website is for testing. Use the supported custom-website integration for the storefront.
Acceptance cases for the first build
Use synthetic customer records for private-data checks. These are suggested tests with expected outcomes, not reported product results.
| Test | Expected outcome |
|---|---|
| Ask about a known product attribute | The answer agrees with the approved product record |
| Change a test variant from available to sold out | The next live lookup returns sold out and the agent does not promise availability |
| Request another test customer’s order | The connected service denies access and the chat reveals no order details |
| Ask for a person | The configured staff destination receives the request and enough conversation context to respond |
| Interrupt the catalogue connection | The agent explains that availability cannot be retrieved and avoids inventing a result |
If the authoring panel works but the website fails, inspect channel authentication and policy restrictions first. If an answer contains an old product fact, inspect the retrieved source before changing the prompt. After a release, start a new conversation to verify the published version.
MCP, connectors and computer use
Copilot Studio supports several ways to connect an agent to business systems. Each route exposes only the operations you configure and authorise. A connection’s presence does not establish that a specific store platform or order operation is supported.
| Connection route | How it works | Retail decision |
|---|---|---|
| Prebuilt connector | Uses a published Power Platform connector’s API operations | Check the actual operations, authentication and permitted data |
| Model Context Protocol connection | Lets the agent use tools and resources exposed by an MCP server | Review the server’s tool inventory and restrict access to customer data |
| Custom action | Uses a custom connector, REST API tool or agent flow you configure | Your team owns input validation, access checks and failure handling |
| Computer use | Operates a website or Windows application through screen interaction | Requires a configured machine and tightly restricted credentials |
Microsoft’s connector documentation describes both prebuilt and custom connectors. Check who publishes the connector and what it can read or change. A prebuilt connector is not automatic approval to expose every operation to shoppers.
For the standard harness, Microsoft’s MCP documentation supports tools and resources and requires generative orchestration. The separate Microsoft Copilot-harness MCP authoring experience is documented as preview. Keep that distinction when choosing instructions or assessing support status.
An MCP server can expose a catalogue search tool and an order-changing tool through the same connection. Review both. Microsoft documents API-key and OAuth authentication for MCP servers. OAuth supports delegated access, but the order service must still enforce which records and operations each customer can use.
Microsoft’s computer-use documentation describes Windows-based automation for systems without a direct API. Its model availability table includes generally available and experimental options. Generative orchestration is required.
Computer use with maker-provided credentials can let users act with the maker’s access on the configured machine. Use a dedicated restricted account and an allowed-site list. A shopper-facing order lookup must never inherit a staff member’s unrestricted administration session.
Before enabling a cancellation or refund, require the connected service to validate ownership and eligibility. Show the customer the proposed change and obtain confirmation before executing it. Record the returned transaction result and handle retries without repeating the commercial action.
Studio compared with Microsoft’s other agent tools
Copilot Studio is the building and operating environment. Microsoft’s other Copilot experiences primarily serve the people using an assistant, while Foundry and Power Automate change how your team implements it.
| Microsoft product | Main role | What changes for a retailer |
|---|---|---|
| Microsoft Copilot | A general assistant experience | Using Copilot does not itself create an agent embedded in your store |
| Microsoft 365 Copilot | Assistance in Microsoft’s workplace apps and organisational context | Fits employee work, with access governed by the user’s permissions and entitlement |
| Agent Builder | Creates lightweight agents within Microsoft 365 Copilot | Suits individual or team knowledge tasks |
| Copilot Studio | Builds agents with connected tools and broader deployment options | Fits a custom website assistant when the retailer owns its configuration and operation |
| Microsoft Foundry, formerly Azure AI Foundry | Develops AI applications and agents using models, tools and developer services | Fits teams needing control over application code and Azure infrastructure |
| Power Automate | Automates processes across apps and services | Fits defined back-office workflows that can support an agent |
Microsoft’s Agent Builder comparison directs broader audiences and advanced integration requirements towards Studio. For a public storefront assistant, deployment scope matters as much as the ability to answer a question.
Microsoft now calls Azure AI Foundry Microsoft Foundry. It supports portal-based agents as well as developer-led implementations. Choose it when the build requires that development control and your team can operate the resulting application.
Power Automate and Studio can work together. A conversational agent collects the customer’s request, while a defined flow runs approved business steps. Microsoft also provides agent flows within Studio. Choose the workflow owner and failure handling before connecting an existing automation to public chat.
Governance and ongoing operation
The retailer needs a named service owner who can change or withdraw the agent when its answers or actions fail. Microsoft provides administration and analytics, but your team owns the store-specific rules and connections.
Assign continuing responsibilities before release:
| Responsibility | Accountable team | Ongoing check |
|---|---|---|
| Product and policy accuracy | Merchandising and customer service | Changed variants and policy exceptions produce the intended answer |
| Permissions and credentials | IT and security | Customer access stays restricted when accounts or connections change |
| Tool reliability | Commerce integration team | Failed requests and repeated actions have a controlled outcome |
| Conversation quality | Agent service owner | Review failed answers and run the acceptance set after changes |
| Human escalation | Customer service lead | Staff receive requests and can continue the customer’s task |
| Release control | Agent service owner and administrator | Approve the tested configuration and retain a recovery route |
Microsoft’s data-policy controls can restrict connectors, HTTP endpoints and publishing channels. A blocked website channel can prevent deployment even when the agent works in the authoring panel. Review that policy before committing to a storefront launch.
Treat a catalogue schema change as an agent change. Renaming a variant field can break a lookup while the chat still appears healthy. Keep a small repeatable test set and review conversation failures after source or tool changes. Limit access to conversation records containing customer information.
A custom Studio build fits a retailer with unusual business processes, existing Microsoft integrations and a team accountable for operating them. A packaged storefront agent fits a store that wants established shopping capabilities without building those connections and conversation flows itself. Before choosing Studio, assign the service owner and demonstrate the product lookup, denied-order case and staff handover in the intended website channel.
Configure the agent’s knowledge and actions
For this store-agent setup, “training” means choosing approved knowledge and writing instructions that describe the permitted task. You also configure the tools it can call and define acceptance examples. This configuration grounds responses in connected information. It does not retrain a foundation model, and completing a Copilot training course does not configure your store’s agent.
Microsoft’s agent-instruction guidance says instructions can use only capabilities you have configured. Writing “check the customer’s order” therefore needs a connected order tool with enforced access controls. A prompt alone cannot supply that connection or grant permission.
Before allowing a consequential action, use these checkpoints in a development environment.
- Confirm the Copilot Studio release available in your environment. Record the selected harness and model alongside the agent version you intend to publish. Match the instructions to those choices using Microsoft’s release information. Features roll out over several days, so an announcement does not prove availability in your environment.
- Restrict the action’s credentials to the permitted operation. Confirm which customer identity reaches the connected service and which order records it can access. Keep a cancellation tool disabled for shoppers while you validate its configuration.
- Check the actual retrieved source. Compare a known variant’s stock with the commerce system and confirm that the policy source contains the approved wording. Correct stale data in its source or retrieval connection before changing instructions.
- Run a permitted request against a synthetic order. For an illustrative cancellation test, the owning test customer confirms an eligible cancellation. The order system must record exactly the intended change and return its result.
- Run a denied request against that same test order using another test customer. The service must reject it without changing the order or revealing its details. Inspect the tool call and stored order state, because a polite refusal in chat alone does not establish access control.
Microsoft’s agent test panel lets you trace the conversation and inspect its variables. After the private checks pass, publish the tested configuration to the controlled website channel and repeat both requests in new sessions. Enable the action for customers only when the published version produces the permitted result and rejects the unauthorised request.
Give your store its own shopping agent
Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.