AI super users: Making MS Copilot agents at Amplify Care

This is the first article in a monthly series on Canadian AI Healthcare Super Users: people inside health organizations who build with AI rather than wait for it. Each piece looks at one person’s stack and the agents they have built. We start with Stewart McKendry, director of AI at Amplify Care, a not-for-profit digital health organization. Most health organizations have Microsoft Copilot because they already run on Microsoft 365. I asked Stewart what it takes to make it useful. Our conversation has been edited for length and clarity. (W.F.)

Access is not adoption: Most Canadian healthcare and public-sector organizations I have worked with run on Microsoft. Employee accounts, email, documents and collaboration all sit in Microsoft 365, so IT can approve Copilot using the contracts, access rules and administrative controls already in place. For most employees, “Copilot” means Copilot Chat, Microsoft’s general-purpose chat application. It runs in a browser, as a desktop app, and inside Teams.

Approval is not the same as use. Ontario’s Auditor General found that Copilot Chat accounted for about 6 percent of the Ontario Public Service’s tracked use of popular generative AI websites from April to August 2025, even though it was the only one approved.

Other services accounted for 94 percent. Those figures describe tracked website traffic. The report did not establish how many employees rejected Copilot, or why. But they match what I see: a large gap between what the tools can do and what employees know how to do with them. People need examples from their own work and help applying the tool to a real task. Handing out access leaves that learning to each employee.

Will: Stewart, where do agents come in?

Stewart: OpenAI and Anthropic each put conversation, research and software creation in one desktop application. Microsoft splits comparable work across Copilot Chat, Cowork, Copilot Studio and GitHub Copilot. An employee who wants to build a custom agent must find and learn another Microsoft product. That may be changing.

On September 25, Microsoft announced a new Copilot that brings Chat and Cowork together in one place. It adds Code, which builds apps and workflows from a plain-language description, and Autopilot, an agent that keeps working in the background. It is rolling out first to Microsoft’s early-access Frontier program. Whether it delivers remains to be seen.

Will: Which product do you use, and what do you mean by an agent?

Stewart: At Amplify Care, that product is Copilot Studio. It is now our preferred way to build and deploy agents for our own teams in a secure environment. By “agent,” I mean an AI tool configured for a particular job:

• The model generates responses.
• Instructions define the job.
• Knowledge supplies Amplify Care’s own information.
• Tools let the agent read or update applications such as SharePoint, Outlook, Salesforce or Asana.

An employee comes back to a tool that already knows its job, the information it can use, and what it is allowed to do.

Will: What does building one look like?

Stewart: Copilot Studio gives us a visual workspace. I also use Claude Code and ChatGPT for the actual writing of the agents. We describe the job, choose from the models approved for our environment, and add the knowledge and tools. We test the agent, see where it fails, adjust it and publish a new version, often without writing custom software. It works especially well for internal agents because our employees already have Microsoft accounts and belong to organizational groups.

We share an agent the way we would share any other Microsoft 365 resource. An agent that uses the signed-in employee’s identity applies that person’s existing access. Other enterprise AI platforms support similar controls, with more setup.

Will: Can you give me a real example?

Stewart: Amplify Care’s deployment teams help with the adoption of technology. Each adoption process is a project that moves through several stages before the individual goes live. Our teams manage thousands of these projects across Salesforce, Asana, email, calendars and other records.

Without an agent, a deployment specialist may have to reconstruct a project’s history from a Salesforce record, Asana tasks, meeting notes and email. The information already exists. The employee has to find it and keep every system current, and that repeated coordination is what we are trying to reduce.

Our Deployment Copilot securely gathers the current state of a deployment, helps maintain its records, identifies the next step, and shows who needs a follow-up. We can then measure whether staff use it and whether it cuts the record-keeping around each project.

Will: What happens when Copilot Studio can’t do the job?

Stewart: When an agent needs code, a custom connection or its own interface, we use Microsoft Foundry and Azure. Foundry lets us deploy an approved model in a region and deployment type suited to our data requirements. It can also index documents so an agent can search by meaning, retrieve relevant passages and cite them. Azure hosts the application and provides its storage, identity and APIs.

Will: What have you built that way?

Stewart: For Salesforce, we wrote a custom tool on Azure that defines exactly which records an agent can read or update. It securely exposes the fields and actions deployment work needs, without giving the agent broader access to our data. For our Knowledge Translation and Evaluation team, we indexed published research, so a change management specialist can ask what evidence supports an approach such as central intake and get findings tied to their sources.

Will: Who writes the code?

Stewart: I write much of this software with Codex. I keep each task in its own conversation, tied to a version-controlled repository, and review the changes before they are merged. The code lives in Azure DevOps and runs on Azure. Some of these applications expose their functions to Copilot agents as custom tools.

Will: Who does this work, and what do you call the role?

Stewart: We call this work AI enablement. The person doing it is often called a forward deployed engineer (FDE). At Amplify Care, that means learning how a team works, finding a repeated task, and turning it into a tool the team can test. The engineer defines what the agent may read or change and revises it with the people using it. Employees supply the details that a general AI tool cannot know: a missing Salesforce field, an irrelevant evidence result, a follow-up the agent should have suggested.

Will Falk is a retired management consulting partner who spends his time with start-ups and as a fellow at four institutions. He is a contributing editor at Canadian Healthcare Technology magazine.

Five agents and AI tools at Amplify Care

• Deployment Copilot: Connects Asana, Salesforce, Outlook and public information to keep onboarding records current and identify the next follow-up.
• Healthmap AI Assistant: Turns plain-language questions into eReferral Healthmap searches, so teams can compare eReferral receivers, save a shortlist and export it.
• Knowledge Translation & Evaluation Agent: Finds sourced results in Amplify Care case studies and evidence briefs, for use in conversations with clinicians and partners.
• French Translation Agent: Drafts Canadian French versions of Amplify Care resources and creates approved Word documents after bilingual review.
• Amplify Portfolio Status: Combines Salesforce opportunities and Asana delivery updates, so leaders can find the projects, risks and decisions that need attention.