Fabric Extended setup guide

Fabric + Copilot Studio is an optional extension of the standard Fabric dashboard for tenants running Copilot Studio agents. It is a superset of the base Fabric build — it reads the same core tables plus the agent tables — so it works only once the core ingesters are producing data.

Before you start

• Stand up the Fabric path first. This add-on builds on the standard Fabric build; get it producing data before you add this layer.
• The folder is self-contained. The core ingesters are mirrored into the add-on’s notebooks/_core/ folder, so everything you need runs from inside it — no cross-folder downloads.
• Provision your Fabric Lakehouse and Entra app registration exactly as for the base Fabric path (same parameters, RBAC roles, and Graph permissions).

1. Run the core notebooks

Run the three required backbone notebooks in notebooks/_core/ first: audit logs → licensed users → org data.

Then add only the optional producers you use:

• Product Feedback — optional.
• Cost consumption — optional.
• Agent 365 — optional; use Copilot_Agent365_Registry_Ingester.ipynb by default. Fall back to Copilot_Agent365_Lander.ipynb only when the registry ingester’s app-registration permissions are unavailable.

The two Agent 365 notebooks target the same dbo.agents_365 table — pick one and do not run both.

2. Parse Copilot Studio transcripts

Run notebooks/Copilot_Agent_Transcript_Parser.ipynb to land the Copilot Studio transcript tables (agent_sessions, agent_turns, agent_errors, agent_subagents, agent_catalogue, agent_performance) from Dataverse.

Permissions: Fabric’s notebookutils.getToken cannot mint a Dataverse token, so the notebook must use an app-registration service principal added as an Application User in each Dataverse environment, with a security role granting Read on the Conversation Transcript table (for example System Customizer, or a custom read-only role). Authentication is client-credentials to {env}/.default — no extra Graph permission is needed for this step.

3. Light up agent credit consumption (optional)

Microsoft’s per-agent Copilot Studio message-credit numbers come out of the Power Platform Admin Center as three CSV exports (per environment, per agent, per user). There is no API, so the flow is export → land in the Lakehouse → ingest:

1. In the Power Platform Admin Center → Billing, download the three EntitlementConsumption…MCSMessages… reports.
2. In your Lakehouse Files area, create a folder named credit_consumption (lower-case) and drop the three CSVs in — or automate the landing with the provided Power Automate flow.
3. Import and Run all on notebooks/Copilot_Credit_Consumption_Ingester.ipynb. It builds the credit_consumption_tenant, credit_consumption_agent, and credit_consumption_user tables.

Screenshot planned

4. Connect the extended template

Open ValueLens - Fabric (+ Studio Agent Deepdive).pbit in Power BI Desktop and supply the same Lakehouse parameters as the base template, then set the toggles explicitly:

• Set Enable_Dataverse = Include to light up the Copilot Studio agent pages (required for this build).
• Set Enable_Consumption = Include only if you completed step 3 (PPAC credit consumption); otherwise leave it Exclude.

Screenshot planned

Limitations and troubleshooting

• Cross-environment / cross-tenant identity. The Dataverse agent tables key on the user’s AAD objectId, while org data is Entra from this tenant. They reconcile only when the transcripts and the Entra directory describe the same users in the same tenant. If agents are published in a different environment or tenant (common in demos), those users won’t exist in the org table and the credit-by-organization breakdown won’t attribute cleanly — this is a data-alignment limitation, not a model bug.
• Billing settling. After the ingester runs, Fabric’s SQL endpoint needs a moment to see the new tables — if a refresh is empty immediately after, wait about a minute and refresh again.

More symptoms and fixes — including the extended Dataverse and credit-consumption issues — are in Troubleshooting & FAQ.