AI and Power BI.

What Copilot in Power BI does (October 2026)

Copilot in Power BI summarises reports, answers questions about data, drafts report pages and DAX, writes narrative visuals and adds summaries to subscriptions. Per Microsoft’s overview, last updated 25 August 2026, the report Copilot pane is generally available while the standalone Copilot and Copilot in apps are in preview. It needs paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher); a Pro or Premium Per User licence alone is not enough, and trial capacities are excluded. Microsoft also warns that without preparing the semantic model for AI, Copilot can return generic or inaccurate results.

Two New Zealand points matter. The standalone Copilot and Copilot in apps are not yet available in the New Zealand North region. And for capacities outside the US and EU boundary, Copilot is off by default until a Fabric admin allows cross-geo processing; Microsoft’s table shows Australian capacities processed in the US. NZ list prices on 19 October 2026, excluding GST and paid yearly: Pro NZ$22.60 and Premium Per User NZ$38.80 per user per month.

Integration routes and when to use each

Automation 1: a monthly narrative with reconciled metrics

Inputs

About a dozen certified measures pulled with executeQueries for this month, last month and the same month last year.

Reconciliation in code

Revenue and expenses must tie to the ledger trial balance, and debtors to the aged receivables report, within set tolerances. A failed check stops the pack and shows the variance.

Model step

Commentary written only from the numbers table plus notes from budget holders. Each sentence carries the metric ID it relies on.

Validator

Every figure in the text must match the table after rounding; percentage changes are recomputed in code.

Human approval

The finance lead edits and approves; the CFO signs off.

Output

The approved text is stored with its data snapshot and version, then published to the report or the board pack. Nothing writes to the semantic model.

Failure and recovery

Fix the source, rerun, and keep the last approved version until a new one is signed off.

Many NZ organisations run a 1 April to 31 March financial year, so check that year-to-date measures use your fiscal calendar. The workflow is mapped in management reporting commentary, and our AI reporting service builds it end to end.

Automation 2: documenting a model with the Authoring MCP server

1. Connect safely. Point the Authoring MCP server at a copy of the model or at Power BI Project files under source control.
2. Inventory. Ask for every measure with its dependencies, and flag duplicates, unused measures and missing descriptions.
3. Review. A developer accepts proposed descriptions and fixes as a pull request. Nothing is applied to production directly.
4. Benefit twice. Clear descriptions also help Copilot, which Microsoft says depends on a model prepared for AI.

Permissions and approval design

For executeQueries, a tenant admin must enable “Dataset Execute Queries REST API”, the caller needs read and build permission on the dataset, and the token needs Dataset.Read.All. Service principals need their own tenant setting and are not supported on datasets with row-level security or SSO, so a reporting bot on a secured model must run as a named user or use an unsecured certified model built for the pack. The hosted Authoring MCP server needs the admin setting “Users can use the Power BI Model Context Protocol server endpoint”.

Microsoft says REST API request and response data may be processed outside your tenant’s home region. If the model contains payroll or customer data, record that in your privacy impact assessment; see data residency for AI systems.

Costs and limits (as listed 19 October 2026)

• executeQueries: one query per call, one table per query, at most 100,000 rows or 1,000,000 values and 15 MB, and 120 queries a minute per user.
• Copilot: needs capacity on top of user licences; new capacity can take up to 24 hours to be recognised.
• Model cost, worked example: a monthly pack at 20,000 input and 2,000 output tokens on Claude Sonnet 5.5 costs about US$0.06 per draft. The real cost is modelling and reconciliation work.

Where it goes wrong

• Silent truncation. A query that exceeds the row limit returns HTTP 200 with partial data and an error in the body. Always check the error field.
• Invented figures. Any number the model computes itself will eventually be wrong. Validate every figure.
• Calendar mismatch. Year-to-date on a calendar year looks plausible and is wrong for a March balance date.
• Unprepared models. Cryptic measure names give Copilot and custom agents poor answers.

Most packs pull from the ledger too; see our Xero and Microsoft 365 guides.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.