Before you start
You will need:
- A license for Microsoft Copilot Studio
- The Filtered MCP server added as a tool to your agent (see our article on adding the Filtered MCP server)
- Your LMS API or MCP server added as another tool for the agent to use (we will be writing an article about this too)
- An administrator who can approve agents in the Microsoft 365 admin centre
The end state: what the agent can do
Once set up, users can open the agent in Microsoft 365 Copilot and ask a question or pick a starter prompt:

The agent in Microsoft 365 Copilot with starter prompts
For a question such as "A first-time manager is joining my team. What should they learn?", the agent:
- Answers from your internal content first, quoting the module it draws on
- Recommends relevant assets with links, adding a small number from external providers to supplement the response as needed
- Records what the user used, understood and demonstrated in their learning transcript
- Asks a comprehension question and explains the correct answer
- Offers a scenario-based skills assessment, scores the answers on a 1 to 5 proficiency scale and updates the user's skills profile in the LMS

An answer built from internal content, with the source module quoted

The follow-up skills assessment
Set up the agent
- In Copilot Studio, open your agent and go to the Overview
- Under Details, click Edit and add a name and description, for example: "This agent looks out for context in workplace apps and suggests relevant learning resources from the Filtered MCP server to help complete the task."
- Under Select your agent's model, choose the model the agent will use for reasoning and responding.

Agent details and model selection on the Overview tab
Write the instructions
Under Instructions, click Edit, write the agent's instructions (up to 8,000 characters) and click Save. The instructions decide how the agent uses Filtered, so most of the setup effort goes here.

The Instructions field
Write the instructions, using the example below as a starting point and replace the placeholders in square brackets with your own details.
Example prompt
You are an AI tutor for [Organisation], powered by Filtered Intelligence's MCP server. You coach users using verified content and assess their skills, and you help L&D users curate pathways and analyse the library.
If asked a general question or for help with a topic:
- Contextualise the response for the user's role, skills and work context.
- Answer using primarily content with the provider '[Internal provider name]' or 'Internal' in results. To make sure you find it using the MCP, use a low relevance threshold of 0.5 and a high asset limit of 100. Focus on retrieving module transcript data to build the answer, and supplement each response with 1 or 2 assets from external providers. Adopt a friendly, conversational tone.
- Use [the API or MCP server to connect to your LMS] to check the user has not already recently completed the content youβre recommending (if they have done so in the past, a reminder may still be helpful)
- Cite the precise source, module or extracted SCORM segment. Include a hyperlink to the asset itself, formatted as hyperlinked text, not the URL.
- In your first response, deliver a short learning nugget (don't name it as such) rather than recommending a one-hour course, then ask one follow-up question to test comprehension. For compliance-related queries, make it a multiple choice question. Record what was used, understood and demonstrated in the user's learning transcript in the LMS using [the API or MCP server to connect to your LMS] and tell the user you have done so.
- Let the user know whether they got it right, and offer a scenario-based skills assessment.
For the skills assessment, ask 4 scenario-based, free-text questions that let you score the user using [the skill profiling methodology in your business], then update the user's skills profile in the LMS.
For curation or L&D admin queries, return anonymised insight about gaps, weak content and unmet demand.
If asked an off-topic question, explain that you are an AI tutor for [Organisation] that coaches users using verified content, assesses skills, and helps L&D users curate pathways and analyse the library.
What the prompt covers
- Sources: answers come mainly from your internal content. The low relevance threshold and high asset limit make sure internal assets are found, with one or two external assets added to each answer.
- Integrations: specify the APIs or MCP servers the agent will use to integrate with other systems
- Citations: each answer names the module or SCORM segment it draws on and links to the asset.
- Coaching: the agent gives a short answer rather than a full course, checks understanding with one question (multiple choice for compliance topics) and records the outcome in the learning transcript.
- Skills assessment: four scenario-based questions, scored from 1 to 5 using the Dreyfus model, with the result written to the user's skills profile.
- L&D queries: curation and admin questions return anonymised insight on gaps, weak content and unmet demand.
- Off-topic questions: the agent explains what it is for.
Configure the Filtered tool
- On the Overview tab, scroll to Tools.
- Next to Filtered, click β― and then Edit.
- Under Details, check that Server is set to Filtered and that Connection shows a green tick.
- Under Tools, use the toggles to choose which MCP tools the agent can call, then click Save.

The Filtered tool alongside Work IQ in the Tools section

Server and connection details for the Filtered tool

Turning individual MCP tools on or off
The tool list updates automatically whenever the Filtered API changes, so you don't need to reconfigure it.
Tool | What it does |
search | Searches the Filtered Connective Intelligence catalogue and returns matching assets |
get_assets | Returns a paginated list of assets, filterable by language |
get_asset | Returns a single asset by its ID |
get_skills | Returns all available skills |
get_skill_groups | Returns skill groups with their associated skills |
get_skill_labels | Returns skill labels with their associated skills |
get_job_roles | Returns job roles with their associated skills |
get_asset_languages | Returns the languages of assets in the catalogue |
get_asset_formats | Returns the asset formats in the catalogue |
get_asset_providers | Returns the content providers in the catalogue |
get_tagged_assets | Returns assets tagged with a specific skill |
get_tagged_assets_by_label | Returns assets tagged with the skills in a specific skill label |
get_tagged_assets_by_job_role | Returns assets tagged with the skills for a specific job role |
get_asset_quality | Returns quality evaluation data for a specific asset |
Tools is also where you add your LMS integration, for example the Workday API or MCP server, so the agent can write to learning transcripts and skills profiles.
Review the other settings
The tabs along the top of the agent control how it behaves and how you track it.
- Agents: connect other agents your agent can hand tasks to, such as a Microsoft agent that books meetings in people's calendars.
- Topics: add scripted parts of the conversation. Use these for framing that must always appear, such as a reminder that AI can make mistakes and users must take responsibility for what they do with AI outputs, or any other safety policy information.
- Activity: see reporting data on how people use the agent.
- Evaluation: set up test questions with expected answers to check that results are relevant, useful and good quality.
- Monitor: track the agent's performance against your evaluation criteria.
Publish the agent
Clicking Publish alone does not make the agent available to users. You also need to publish it to a channel and submit it for approval.
- Click Publish.
- Go to the Channels tab and select Microsoft 365 and Microsoft Teams.
- Click Availability options, choose Show in Teams app store for org and click Submit to org catalog.

Available channels on the Channels tab

Submitting the agent to the org catalogue
The submission status changes to Waiting for approval. If the agent uses Microsoft authentication, only the Teams, Microsoft 365 and SharePoint channels are available.
Approve the agent
An administrator needs to approve the agent before anyone can use it.
- Go to the Microsoft 365 admin centre (admin.cloud.microsoft.com).
- Select Agents, then All agents, then the Requests
- Find the agent, which shows as Pending review with the request type Publish, and approve it.

A publish request waiting for review in the Microsoft 365 admin centre
Note: every time you update the agent, you need to publish it and get it approved
