Business Process Automation with Power Automate and Copilot Studio Agents: A Worked Example

Short version: Most business processes fail in the gap between a person asking for something and the system that records it. Power Automate closes the system side — triggers, approvals, connectors, logging — and a Copilot Studio agent closes the human side, answering the policy questions and collecting the details in plain language, grounded on the documents you already keep on SharePoint. Put them together and a purchase request that used to be a six-message Teams thread becomes a two-minute conversation that ends with an approved, logged record. This article walks through that exact build step by step, lists the process patterns worth automating first, and covers the governance that keeps an enterprise rollout safe. The hands-on version is the WSQ Business Process Automation with Power Automate and Copilot Studio Agents course (TGS-2022017524), funded up to 70%.

Automation and agents are not the same thing

Power Automate has been the Microsoft 365 automation tool for years. A cloud flow waits for a trigger (a form submitted, a file added to SharePoint, a row updated), runs a sequence of actions through hundreds of connectors, and can pause for a human approval. It is deterministic: the same input produces the same path every time. That is its strength and its limit. A flow cannot tell a colleague which form to use, cannot explain the policy, and cannot cope with “I need a laptop for the new hire, the fancy one, by Monday”.

A Copilot Studio agent is the opposite. It understands language, it can be grounded on your policy documents, and it can decide what to ask next. What it should not do is own the record of what was approved. The design pattern the course teaches is therefore simple:

If you have read our piece on business transformation with agentic AI, this is the same principle: the agent handles the messy front, the workflow handles the auditable back, and a human keeps authority over consequences.

The worked example: a purchase request agent

Here is the build we do in class, end to end. The scenario is a purchase request for equipment under a spending policy, but the same skeleton fits leave requests, IT access, travel, vendor onboarding and customer support escalations.

1. Put the knowledge on SharePoint. One document library holds the procurement policy, the approved-vendor list and the equipment catalogue. Make sure every requester already has read access; the agent will answer with the same permissions the user holds.
2. Create the agent in Copilot Studio. Name it, describe its job in the instructions (“You help staff raise equipment purchase requests under the procurement policy. Ask for item, quantity, cost centre, business reason and needed-by date. Do not approve requests.”), and add the SharePoint library as a knowledge source.
3. Test generative answers. Ask it “can I buy a second monitor without approval?” and check that it answers from the policy document and cites it. Generative answers are the feature that turns a scripted bot into something people will actually use; if the answer is wrong, fix the document, not the agent.
4. Add a topic that collects the request. Trigger phrases like “raise a purchase request”. The topic asks the five questions in turn, with the agent filling any it can infer from the conversation, and confirms the summary back to the user.
5. Build the Power Automate flow. From the agent, create a flow with inputs for the five fields. The flow creates an item in a SharePoint list called Purchase Requests, looks up the requester’s manager from Microsoft 365, and starts a Start and wait for an approval action addressed to that manager, with the policy clause the agent found pasted into the approval details.
6. Return the result to the agent. The flow’s outputs (request ID, approval status, approver comments) go back to the topic, which tells the user “Request PR-0142 sent to Mei Lin for approval; you will get a Teams notification when it is decided.” Microsoft’s documentation on calling flows from an agent covers the input and output mapping.
7. Close the loop. A second flow, triggered when the list item’s status changes, posts the decision to the requester in Teams and, if approved, emails the vendor request to procurement. Every state change is a list row with a timestamp, which is your audit trail.
8. Publish to Teams. Add the agent to Microsoft Teams so it lives where people already are. Pin it in the channel where purchase questions used to be asked.

The whole thing takes an afternoon the first time and an hour the second. The part people underestimate is step one: an agent grounded on a clean, current policy library is dramatically better than one grounded on three contradictory PDFs from different years. SharePoint hygiene is now a business-process skill.

Which processes to automate first

The course covers six families of process because they account for most of the manual effort in a typical Singapore office. The table below is the shortlist we use to help learners pick their first project.

Start with an approval. It has a clear owner, a policy to ground on, an obvious before-and-after measure (days to decision), and every organisation has one that annoys everybody. If the process involves an old on-premise application with no connector, the Power Automate Desktop RPA course covers the robotic side.

Human-in-the-loop is a design choice, not a fallback

There is a temptation, once the agent works, to let it approve small requests itself. The course argues against it, and the argument is worth repeating. An approval flow with a human in it costs the approver thirty seconds and gives you a defensible record. An agent that approves on its own saves those thirty seconds and gives you an unexplainable decision if anything goes wrong. Use the agent to make the human’s decision faster and better informed — policy clause attached, budget remaining shown, duplicate requests flagged — and keep the click with a person. Where an action is genuinely low-risk and reversible (booking a meeting room, resetting a notification preference), let the flow act directly; where it spends money or grants access, do not.

Governance for an enterprise rollout

Everything above can be built by a business user, which is exactly why the last day of the course is about deployment, governance and monitoring. The essentials:

• Environments. Build in a development environment, publish from a production one. Do not let a half-finished agent answer real staff.
• Data loss prevention policies. The Power Platform DLP policies decide which connectors can be combined; a flow that reads SharePoint should not also be able to post to a personal Gmail connector. Set this before the first flow, not after the first incident.
• Agent security. Copilot Studio’s security and governance guidance covers authentication, who can edit the agent, and how knowledge-source permissions are enforced. Authenticate users so the agent knows who is asking.
• Monitoring. The built-in analytics show which topics fire, where conversations abandon and which questions the agent could not answer. Review the unanswered list weekly; it is a to-do list for your SharePoint library.
• Change control. Version the policy documents. When the spending threshold changes, the agent’s answers change the moment the document does, so someone must own that document.

Governance is what turns a clever demo into something the finance director will let run. It is also, in our experience, the section that gets a learner’s automation project approved back at the office.

Learn it hands-on, funded up to 70%

The WSQ – Business Process Automation with Power Automate and Copilot Studio Agents course (TGS-2022017524) is a two-day, lab-driven programme that builds the pattern above from a blank tenant. The six topics:

1. Introduction to business process automation and the Microsoft Power Platform
2. Workflow automation with Power Automate — triggers, actions, approvals and connectors
3. Building AI agents with Copilot Studio — instructions, topics, SharePoint knowledge and generative answers
4. Integrating Power Automate with Copilot Studio agents — calling flows, passing inputs and outputs, human-in-the-loop
5. Enterprise business process automation — approvals, notifications, data collection, support and operational workflows
6. Deployment, governance and best practices — environments, DLP, security, monitoring

Assessment is a written and a practical exam. Learners leave with a working agent-plus-flow they built themselves, which most take straight back to their own tenant.

The nett fee can be further offset with SkillsFuture Credit for eligible Singaporeans, PSEA for eligible younger learners, and SkillsFuture Enterprise Credit (SFEC) for eligible Singapore-registered companies; employer-sponsored SMEs can also claim absentee payroll. The course runs over two full days (9:30am to 6:30pm), with physical classroom and synchronous online (Zoom) options. Figures are taken from the course page at the time of writing — the live page always carries the current fee and the next available dates.

Related paths in the Microsoft Copilot series: Copilot for Power Automate if you want Copilot to write the flows for you, Creating Intelligent Chatbots with Copilot Studio for a shorter non-WSQ introduction, and WSQ AI Agents for Business for the platform-neutral view.

Frequently asked questions

The bottom line

Power Automate has always been able to run a process. What it could not do was talk to the person raising the request, explain the policy, or forgive an untidy ask. A Copilot Studio agent grounded on your SharePoint documents does exactly that, and hands the clean, complete request to the flow. Keep the approval with a human, keep the record in a list, set governance before the first publish, and you have an automation that people use because it is easier than the old way, not because they were told to.

Ready to build your first agent-plus-flow? Register for WSQ – Business Process Automation with Power Automate and Copilot Studio Agents — 2 days, funded up to 70%, SkillsFuture Credit claimable.