Design and build integrated AI agent solutions in Copilot Studio
£1,199 (ex VAT)
What you will be able to do afterwards
Building a Copilot Studio agent that answers questions from a SharePoint site takes an afternoon. Building one that raises a return, waits for a manager to approve a refund, looks up the customer in Dataverse and hands the delivery question to a specialist agent is a different job. By the end of the three days you have built exactly that, and you know how to test it, publish it and keep it healthy.
Why learn it with us rather than free
The AB-620 learning paths are on Microsoft Learn, free, and an experienced builder with a developer environment can work through them alone.
What a course adds is three days with nothing else competing for them, a Microsoft Certified Trainer to answer the integration questions your own systems raise, and five cumulative hosted labs in which you build one complete agent solution from first instruction to published, monitored agent. A practice exam and an exam voucher are included.
If you have never built an agent in Copilot Studio, this is too steep a start: Microsoft expects hands-on experience first. If your role is designing the overall agentic solution rather than building it, AB-100 is the architect course, and we will say so.
What the three days cover
The course follows Microsoft’s official AB-620T00 courseware, released in October 2026 and built around the GitHub Copilot harness in Copilot Studio: five learning paths, nineteen modules and five hosted labs. Each learning path is taught, then followed by its lab, and the labs build on each other, so by Day 3 you have one working Customer Support Rep Assistant with workflows, tools and a connected specialist agent. Labs, discussion and knowledge checks take roughly half of the class time.
Explore Copilot Studio and agent harnesses
What a harness does, how the GitHub Copilot, standard and Copilot Chat harnesses differ, and which one fits a given business scenario. The lifecycle and authoring surface of a GitHub Copilot harness agent.
Manage usage and cost for agents
How billing differs by harness, what drives consumption at build time and at run time, and the admin controls that keep Copilot Credits under control. Design choices are the biggest cost lever.
Create instruction-driven agents
Defining purpose, scope and boundaries before writing anything, then writing instructions that control behaviour, tone, grounding and escalation. Adding file and website knowledge, then previewing and refining the agent.
Design reusable behaviour with skills
When a skill is the right component, authoring skills from templates or from scratch, and packaging them so several agents can reuse the same behaviour. Lab 1 builds the environment, solution, agent and two skills.
Automate agent tasks with workflows
When to rely on agent reasoning and when a deterministic workflow is safer. Building and publishing a standalone workflow, adding it to the agent as a tool, and testing and troubleshooting its runs.
Add human-in-the-loop steps to workflows
Deciding which steps need a person, collecting missing values with a request-for-information action, and branching on an approval decision. Lab 2 adds return and refund workflows with approval checkpoints.
Design integration strategies for agents
Telling knowledge sources, tools and connected agents apart, recommending the right approach for a requirement, and confirming the identity, connections and policies each one needs. Writing descriptions that help the agent choose.
Ground agents in enterprise knowledge
Comparing knowledge source types, adding a Copilot connector as knowledge, validating grounded answers and planning evaluations for them.
Extend agents with tools
Choosing and configuring connector tools for enterprise actions, and why adding a tool does not guarantee the agent uses it. Refining names and descriptions, then testing tool selection from the run trace.
Connect agents to MCP servers
When an MCP server is a better fit than a connector or workflow, adding one with the right authentication, choosing which of its tools the agent may use, and troubleshooting MCP calls.
Connect agents to Microsoft IQ
Where Foundry IQ, Fabric IQ and Work IQ fit alongside knowledge and tools, and connecting an agent to each. Lab 3 connects Dataverse data, the Work IQ MCP server and an Outlook action.
Design multi-agent solutions
When several agents are genuinely better than one, the main coordination patterns, and the implementation options available on each harness.
Build connected agent solutions
Preparing a specialist agent so others can connect to it, adding it with a distinct routing description, testing delegation, and calling an agent from a workflow. Lab 4 adds a fulfilment specialist and tests the handoff.
Design a governance and security strategy
How controls apply at tenant, environment and agent level, a zoned approach to classifying agents by risk, authentication choices, channel readiness, and monitoring through Agent 365 and the agent inventory.
Design a responsible AI strategy
Turning Microsoft’s responsible AI principles into design decisions, configuring content moderation, choosing models with safety in mind, and designing guardrails and transparency.
Evaluate agents
Building test sets from real scenarios, running evaluations on the Evaluate tab, triaging failures, and checking a model change against agreed acceptance gates before it goes live.
Publish and deploy agents to channels
Configuring end-user authentication, publishing identifiable versions, and deploying to the demo website, a web app, Microsoft Teams and Microsoft Copilot.
Monitor and improve agents
Reading the Monitor tab, diagnosing issues from session evidence and transcripts, extending monitoring with Application Insights, and using telemetry to show the agent’s value.
Implement application lifecycle management
Planning development, test and production environments, packaging an agent and its dependencies in a Power Platform solution, and moving it safely between environments. Lab 5 evaluates, publishes to Teams and monitors the finished agent.
Who this course is for
A good fit if you are
Not the right course if
What is included
Hands-on, in a hosted lab
Five hosted labs, around six and a half hours in total, and they are cumulative: you build one Customer Support Rep Assistant from an empty environment to an evaluated, published and monitored agent, adding workflows, Dataverse data, an MCP server, an Outlook action and a connected specialist agent along the way. There is nothing to install and your own tenant stays untouched.
Three ways to take it
Online
Live over Microsoft Teams with a real trainer and full class interaction. £1,199 per person.
In person
See the In Person course dates for locations and associated costs.
Onsite / closed
At your site, just your team. For this subject a private course changes the conversation, because we can talk through the systems your agents actually need to reach, the governance zones they will live in and the Copilot Credits they will consume. Best value for groups. Ask for a quote.
What it costs to not do it
Most organisations now have Copilot Studio agents. Far fewer have agents that do anything beyond answering questions from a document library, because the integration work is where projects stall.
The usual result is a pile of pilots. Agents that impressed in a demo, could not reach the systems that mattered, and were quietly abandoned, or were connected in a hurry with more access than anyone intended.
Three days gets a builder to the point where they can design an agent properly and connect it safely. Against the cost of another stalled pilot, that is a reasonable trade.
There are no reviews yet. Be the first one to write one.
The course ends. The integration work does not.
The hard questions usually arrive when you try to connect an agent to your own systems and the API does something the lab did not. Both of these are included in the price.
Trainer email support
Email our support team with the questions that come up as you apply it to your own work. Not a ticket queue, and not a chatbot. The people who teach you, answering the thing you are actually stuck on.
The learner community
A private community for people who have trained with us, with a board for each technology. Ask questions, read what others have hit, and get answers from our trainers and other delegates. Your access does not expire when the three months do.
Where this fits
The Copilot Studio agent path
AB-620 is the associate-level build course. Here is what sits either side of it.
PL-7008 Create agents in Copilot Studio
The old one-day Copilot Studio introduction retired in July 2026. Build some agents yourself before AB-620.
AB-620 Integrated AI agents in Copilot Studio
Three days building, integrating, evaluating and publishing Copilot Studio agents, leading to a certification.
AB-100 Architecting agentic AI business solutions
Three days on designing agentic solutions across the Microsoft platform, for architects.
Others in Microsoft 365
Other courses in this area, in case a different level or focus suits you better.
MS-4022: Extend Microsoft 365 Copilot in Copilot Studio
MS-4022 · Microsoft Official Course Extend Microsoft 365 Copilot with agents built in Copilot…
MS-4021: Copilot Immersion Experience
MS-4021 · Microsoft Official Course Copilot Immersion Experience: two hours to see what Microsoft…
MS-4019: Transform your everyday business processes with agents
MS-4019 · Microsoft Official Course Transform your everyday business processes with agents One day…
Ask before you book
Tell us what you have built in Copilot Studio so far. If AB-620 is right we will say so, and if AB-100 or more hands-on time first suits you better, we will say that instead.
Onsite & closed courses
A private course for up to ten people costs roughly the same as five public seats, groups of 11 to 14 are topped up per person, and above 14 we add a second trainer. In person and onsite are priced on request.
Not sure AB-620 is the right course?
Tell us a little about where you are with Copilot Studio and the systems your agents need to reach. A trainer will come back to you, usually the same working day.
If AB-100, AI-103 or a private course suits you better, we will say that instead.
Prefer to talk? Call +44 121 269 8660
What you get that you will not get everywhere
Hands on, not slides
Five cumulative hosted labs. You build one complete agent solution yourself, from first instruction to published and monitored.
Taught by a certified trainer
A Microsoft Certified Trainer, live, with room for questions about your own agents and integrations.
Real prices, published
Every course shows its price and what is included. No “contact us for pricing”, no surprises.
We say when not to book
If you need more Copilot Studio experience first, or AB-100 is the better fit, we will tell you.
There are no reviews yet. Be the first one to write one.
Read before you book
Teams we have trained
Is AB-620 a current course?
Yes. Microsoft released the current version of AB-620T00 in October 2026, rebuilt around the GitHub Copilot harness in Copilot Studio. It leads to the AB-620 exam and the Microsoft Certified: AI Agent Builder Associate certification.
How long is the course, and are there labs?
Three days, with five hosted labs that build on each other. Together they take around six and a half hours, and with discussion and knowledge checks that is roughly half the class time. You finish with one working agent solution you built yourself.
What experience do I need?
Microsoft expects hands-on experience building agents in Copilot Studio, an intermediate understanding of generative AI, proficiency with REST APIs and familiarity with the Microsoft cloud platform. If you have only used Copilot Chat, this is not the place to start.
What is on the AB-620 exam?
According to the official study guide, three areas. Planning and configuring agent solutions is 30 to 35 per cent. Integrating and extending agents in Copilot Studio is 40 to 45 per cent, the largest area, covering enterprise knowledge sources, tools such as MCP and REST APIs, multi-agent collaboration and Azure integration. Testing and managing agents, including evaluation and application lifecycle management, is 20 to 25 per cent. A score of 700 or greater is required to pass.
How much does it cost, and is the exam included?
The scheduled online course is £1,199 per person excluding VAT. That includes a Microsoft Certified Trainer, Microsoft Official Courseware, hosted hands-on labs, an official AB-620 practice exam, a Microsoft AB-620 exam voucher, and three months of trainer email support afterwards.
Do I need to know how to code?
You do not need to be a full-time developer, but you do need to be comfortable with REST APIs and integration patterns, and Power Fx helps. Much of the course is about connecting agents to other systems, which is closer to integration work than to drag-and-drop building.
Does the course still teach topics and child agents?
No. The October 2026 release teaches the GitHub Copilot harness build path: instructions, skills, workflows, knowledge, tools and connected agents. If your organisation still runs classic topic-based agents, tell us and we will talk through how the course maps to what you have.








