Copilot for Power BI Consulting — Setup, Governance, and a Private-AI Alternative for Regulated Data
Copilot in Power BI brings natural-language analytics, narrative summaries, DAX generation, and synonym suggestions to every report user — but it does not run on a Power BI Pro license, it consumes Fabric capacity, and it carries real prompt-injection and oversharing risks that most tenants never configure for. Petronella Technology Group, Inc. helps Raleigh and North Carolina organizations size Fabric capacity correctly, prepare semantic models to actually answer questions, govern Copilot through Microsoft Purview, and stand up a private-AI alternative for the workloads where Copilot is not the right answer.
If you have already tried Copilot inside Power BI Desktop and gotten the “Copilot isn’t available” banner, the cause is almost always one of three things: your tenant is on Pro or PPU only (no Fabric or Premium capacity), the Fabric admin portal has the tenant setting turned off, or the workspace your model lives in is not assigned to a Copilot-eligible capacity. None of those are bugs — they are intentional gating that Microsoft built so that organizations would think about governance before letting a generative-AI assistant loose on their analytical surface. The work of standing up Copilot properly is the work of buying the right capacity, turning on the right settings, prepping the model so Copilot can actually answer, and writing a policy so people use it correctly.
That is the engagement Petronella Technology Group, Inc. runs. The rest of this page walks through what we do, how Copilot in Power BI actually behaves, what to watch out for in regulated environments, and the private-AI alternative we offer when Copilot is not the right answer.
What Copilot in Power BI actually does
Copilot in Power BI is not one feature, it is a family of generative-AI capabilities that show up at multiple places in the Power BI experience. They share a common Azure OpenAI Service backend and a common set of capacity, tenant, and regional requirements, but each capability is enabled, billed, and governed slightly differently. Knowing the shape of the surface area is the first prerequisite to writing acceptable-use policy that covers it.
Natural-language Q&A on a semantic model
Business users ask questions in plain English and Copilot translates them into measures, dimensions, and filters against the published semantic model. Available through the report Copilot pane and the standalone Copilot experience in the Power BI service.
Narrative summaries of a report or page
Copilot reads a report or a page and writes a paragraph of plain-language insights. Useful for executive subscription emails, daily briefings, and dashboards where the reader does not have time to interpret the visuals manually.
DAX generation and explanation
Copilot writes DAX measures from a description, explains what an existing DAX measure does, and helps debug DAX that is slow or returning unexpected results. Significantly raises the productivity floor for analysts who are not full-time DAX engineers.
Semantic-model help
Copilot suggests measure descriptions and synonyms so the linguistic schema actually understands the words your team uses. The biggest unlock for getting Copilot to answer business questions well.
Report-page creation from a prompt
Copilot drafts an entire report page from a natural-language prompt, choosing visuals, filters, and slicers. Treat the output as a first-draft accelerator that an analyst then refines, not as a finished report.
App-scoped Copilot (preview)
Inside a published Power BI app, Copilot scopes its answers to the curated content of that app. Useful for surfacing answers to consumers who should not browse the underlying workspace directly.
All six capabilities run on the same plumbing: Fabric F2+ (or Power BI Premium P1+) capacity, tenant admin enablement, supported region, no sovereign cloud, no trial SKU. If any one of those conditions fails, none of the Copilot capabilities work for that workspace.
Prerequisites: Fabric capacity, admin settings, and a workspace that qualifies
Microsoft publishes the requirements in plain language on Microsoft Learn, and Petronella Technology Group, Inc. has translated them into the checklist we walk every client through during the readiness engagement. There are four moving parts and they all have to line up.
- Paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher). A Pro license, a PPU license, or any trial SKU is not enough. The capacity must be assigned to the workspace where your semantic model lives.
- Tenant admin enablement. In the Fabric admin portal, the setting “Users can use Copilot and other features powered by Azure OpenAI” must be on. Some preview experiences (standalone Copilot, app-scoped Copilot) have their own additional toggles.
- Supported region. Your Fabric capacity must be in a region where Microsoft has deployed Azure OpenAI for Fabric. If you are outside the US or the EU data boundary, the tenant admin must also explicitly allow cross-geo data processing.
- No sovereign clouds. US Government cloud and other sovereign environments are not supported — due to GPU availability per Microsoft — and will not be in the short term.
- Time delay after provisioning. When you buy new capacity or scale up an existing one, Microsoft warns it can take up to 24 hours for Copilot to recognize the change and become available.
These are not opinions, they are Microsoft’s published prerequisites. The Microsoft Learn page on Copilot for Power BI lists them under the “Requirements at a glance” section and again under “Copilot requirements.” Our value in the engagement is mapping them onto your tenant, your budget, and your actual usage projection — not just reading them to you.
How to prepare your data for Copilot in Power BI
Microsoft is explicit on this point in their own documentation: “Model owners need to invest in prepping their data for AI… Without this prep, Copilot can struggle to interpret data correctly — leading to generic, inaccurate, or even misleading outputs.” The semantic-model preparation work is what separates Copilot deployments that delight users from deployments that quietly get turned off three months in. Below is the six-step preparation checklist Petronella Technology Group, Inc. runs before any client turns Copilot on for end users. It is also encoded as HowTo structured data on this page so search engines can surface the steps as a rich result.
Provision Fabric F2+ or Power BI Premium P1+ capacity
Confirm your tenant has paid Fabric F2 or higher, or Power BI Premium P1 or higher, and that the capacity is assigned to the workspace your semantic model lives in. Pro and PPU licenses alone are not sufficient; trial SKUs are not supported. For most small and mid-sized firms F2 or F4 is the right starting size — we project Capacity Unit (CU) burn from your current report views, refresh schedule, and expected Copilot prompts per day before recommending a tier.
Enable Copilot tenant settings
In the Fabric admin portal turn on “Users can use Copilot and other features powered by Azure OpenAI.” If your capacity is outside the United States or the EU data boundary, enable the cross-geo data-processing setting as well; otherwise Copilot stays disabled by default. If you want the preview standalone Copilot or app-scoped Copilot experiences, enable those toggles too.
Rename tables, columns, and measures to plain business language
Copilot answers questions in natural language. If your fact table is FCT_REV_USD_2023_AMT and your measure is m_Rev_PY_LCYD, Copilot is going to struggle. Rename to readable forms (Revenue, Revenue Prior Year), and add measure descriptions explaining what each measure represents. The descriptions become the model’s ground truth when an ambiguous prompt arrives.
Add synonyms and Q&A linguistic schema
Power BI exposes a linguistic schema where you can teach the model that revenue, sales, bookings, and topline all map to the same measure. Without synonyms Copilot will refuse questions that use vocabulary your team actually uses. We work with your business stakeholders to build a synonym map that matches how they talk, not how the data team labels things.
Apply sensitivity labels and DLP policies
Tag every semantic model with a Microsoft Purview sensitivity label and configure data-loss prevention (DLP) policies that block Copilot from answering against highly restricted content. This is the single most important step for regulated organizations. It is also where most non-regulated organizations make the largest oversharing mistakes — HR salary data, finance forecasts, and customer PII end up readable by anyone who can prompt a chat pane.
Pilot, monitor, and review at 60 days
Roll out Copilot to a single business team first. Monitor capacity-unit (CU) burn through the Microsoft Fabric Capacity Metrics app. Review every question Copilot refused or answered poorly, refine the semantic model, then expand the rollout. Re-evaluate F-SKU sizing at 60 days — most teams over-buy by 20-30% at the start, and a right-size review pays for itself.
Governance: DLP for Copilot, prompt-injection, and oversharing risk
Petronella Technology Group, Inc. comes at this from a cybersecurity practice first. We have been doing CMMC, HIPAA, and incident-response work since 2002, and that lens is unusual in the Power BI consulting market. The governance section of a Copilot deployment is the section most generalist BI shops skip past because it does not feel like “data work.” In our experience it is the section that decides whether the project survives the first audit.
Microsoft Purview sensitivity labels and DLP for Copilot
Microsoft Purview is the data-classification, sensitivity-labeling, and data-loss-prevention layer that runs across Microsoft 365, Power BI, Fabric, and now Copilot. Purview sensitivity labels apply to semantic models the same way they apply to a Word document or a SharePoint site. When a label is applied and an associated DLP policy is in place, Copilot will refuse to answer questions that would produce output above a configured sensitivity threshold, or for users who are not authorized to view content at that label.
Three configuration choices matter most:
- Default labels for new semantic models. Without a default, models are unlabeled, and unlabeled means uncovered by DLP. Push a default at the workspace level so every new model inherits at least a baseline classification.
- Mandatory labeling for shared workspaces. If a workspace is published to consumers, require a sensitivity label before publish-to-app is allowed. This is the simplest single control to prevent unlabeled regulated data from flowing into a Copilot-readable model.
- DLP rules that block Copilot on restricted labels. Configure Purview DLP rules so that any Copilot prompt against a model labeled “Highly Confidential” or “CUI” or “Regulated” returns a refusal rather than an answer. Then test it with a known-bad prompt and confirm.
Prompt-injection: what it is and why Power BI is exposed
Prompt-injection is the class of attack where malicious or attacker-controlled text inside the data Copilot reads is interpreted by the language model as an instruction, not as content. The attacker does not need to talk to Copilot — they only need to plant text somewhere Copilot will read it. In Power BI the attack surface is wider than people expect:
- Measure descriptions and field descriptions (often editable by anyone with write access to the model).
- Annotations and metadata in the semantic model.
- Text columns in source data — for example, a free-text “Notes” column on a Salesforce object, where a customer-facing rep could paste arbitrary content.
- Translated string tables, glossary tables, and any other user-editable lookup data Copilot can read.
Mitigations include treating measure descriptions as code-reviewed content, restricting write access to semantic models to a small group, sanitizing or excluding free-text columns from Copilot-reachable models, and monitoring Copilot prompts through the Capacity Metrics app for unusual output patterns. None of these mitigations is exotic, but each one has to be deliberately configured.
Oversharing — the bigger day-one risk for most tenants
Prompt-injection gets the headlines. Oversharing is the bigger day-one problem in nine out of ten tenants we audit. Copilot answers what the user can already see; it does not invent new access. But because Copilot makes data discoverable in natural language, it accelerates exposure of anything that was technically shared but practically buried. Examples we have walked into in the last six months: an HR semantic model published to a workspace where the “Members” role had quietly grown to include the entire IT department; a finance forecast workspace with viewer access granted to a vendor mailbox three years ago; a customer-account table containing call-recording transcripts no one realized was in scope.
The fix is an access-review pass before Copilot turns on, plus a workspace-by-workspace decision about which models Copilot is allowed to read. We run this as part of the readiness engagement.
The private-AI alternative for regulated workloads: Penny
For some Petronella Technology Group, Inc. clients, Copilot in Power BI is the wrong answer regardless of how well it is governed. The data classification is too sensitive, the contractual constraints rule out third-party processing, the cyber-insurance carrier objects, or the workload simply needs to stay inside a tightly drawn boundary. For those cases we offer Penny.
Penny is the Petronella natural-language analytics layer that runs entirely on Petronella fleet hardware. It connects to your Power BI semantic model, your data warehouse, or your file share, and answers questions in natural language — the same experience Copilot delivers — without sending data to Microsoft, Azure OpenAI, or any other public service. Inference runs on Petronella-managed GPUs based in our facilities, the model weights are operated by Petronella, and the conversation history stays inside the Petronella boundary.
For CMMC L2 and L3 contractors
When Controlled Unclassified Information (CUI) flows through analytics, Penny keeps inference inside a boundary you can describe to an assessor. Pairs with Petronella’s CMMC L2 program for end-to-end coverage.
For HIPAA-regulated practices
Patient information stays on hardware you can audit. No prompts leave the boundary. Inference logs become part of the HIPAA audit log retention, not a third-party privacy review.
For law firms and accounting firms
Privileged communications and client-confidential financial data stay inside the firm. Penny answers the partner’s natural-language question without that question entering a public LLM’s training, evaluation, or telemetry path.
For organizations that have not decided yet
Run both. Use Copilot for unrestricted models, use Penny for restricted ones. We help draw the line, and we configure the Microsoft Purview rules that enforce it.
Penny is a Petronella own brand. The model selection, the GPU fleet, and the integration architecture are reviewed quarterly. The point is not to recreate Copilot feature-for-feature — it is to give a regulated client the natural-language analytics experience without the third-party-processor risk profile. For more on the broader practice, see AI services from Petronella Technology Group, Inc..
F-SKU right-sizing: the licensing arbitrage most Copilot rollouts miss
Most clients arrive having bought F-SKU capacity through a partner or a Microsoft account team that defaulted to a comfortable size. After 30-60 days of real Copilot usage data, almost every one of them is over-bought by 20-30%. The Capacity Metrics app makes the gap visible if you know which tabs to read. The F-SKU Right-Sizing Audit is a standalone deliverable where Petronella Technology Group, Inc. reviews your current Fabric or Power BI Premium spend against your actual usage and recommends a target SKU.
For a typical mid-sized client, the audit savings cover the rest of the readiness engagement. We treat it as a separate, optional engagement — happy to be hired just for the audit, happy to bundle it into a larger Copilot or pillar Power BI engagement. Either way, we publish the math, and the client owns the decision.
Comparison: Copilot in Power BI vs. private-AI vs. doing nothing
| Dimension | Copilot in Power BI (governed) | Penny private-AI | Do nothing (status quo) |
|---|---|---|---|
| Natural-language Q&A | Yes — broad coverage of Power BI workloads | Yes — same UX, scoped to permitted models and data sources | No — users open tickets, wait, or build the report themselves |
| Data residency | Azure OpenAI region (US or EU data boundary by default; configurable) | Petronella fleet (Raleigh, NC), inside a tightly drawn boundary | N/A |
| Licensing | Fabric F2+ or Power BI Premium P1+ required; Pro/PPU alone insufficient | Petronella service subscription; no per-seat Microsoft AI add-on required for Penny | Existing Power BI Pro / PPU continues to work |
| Regulatory fit | Suitable for most non-restricted data with Purview DLP; case-by-case for CUI/ePHI | Designed for CUI, ePHI, privileged data, attorney work product | No new AI risk added; existing operational and reporting bottlenecks unchanged |
| Time-to-value | 4–6 weeks for SMB readiness; 8–12 weeks for enterprise tenants | 6–10 weeks for integration with existing data sources | Immediate — no project |
| Ongoing cost | Fabric CU burn (variable); F-SKU base cost; governance hours | Petronella managed service (fixed) | Internal headcount on report-writing requests |
For about 60% of the prospects we talk to, the right answer is governed Copilot in Power BI. For about 30%, the right answer is Penny for restricted models plus governed Copilot for the rest. For the remaining 10%, the constraints rule both options out for now — usually because foundational classification work has not been done, and we recommend that work before any AI rollout. We tell clients which bucket they are in during the readiness assessment.
Engagement model
Petronella Technology Group, Inc. delivers Copilot for Power BI work through three engagement shapes. Pricing is custom per client — we use “Request a Quote” until we have scoped the semantic-model count, workspace count, Purview maturity, and pilot population. Quotes are valid for 30 days from issue.
Copilot Readiness Assessment
A focused 2-3 week engagement: capacity review, tenant-setting audit, semantic-model preparation gap analysis, Purview readiness, and a written go/no-go recommendation per workspace. The deliverable is a plan; implementation is a separate engagement. Request a Quote.
Power BI Foundation + Copilot Bundle
The Foundation tier of our AI services practice plus the Copilot-specific governance and semantic-model preparation layer. Covers F-SKU sizing, tenant settings, Purview labels + DLP, semantic-model prep, pilot with one business team, and a 60-day right-size review. Request a Quote.
F-SKU Right-Sizing Audit
Standalone engagement: review current Fabric or Power BI Premium spend against Capacity Metrics data, identify idle or oversized capacity, recommend a target SKU, project monthly savings. Typically saves 20-30% within 30 days. Request a Quote.
For clients also evaluating the broader Power BI consulting practice — dashboard development, vertical packs, governance, migration — the Copilot readiness engagement folds into our pillar Power BI consulting offering. (Pillar page link is the canonical starting point; see Power BI consulting for the full service catalog.)
Why Petronella Technology Group, Inc. for Copilot in Power BI
There are plenty of Power BI consultancies. Very few of them come at the work with a cybersecurity-first lens, and almost none of them operate their own private-AI fleet. Three things make Petronella Technology Group, Inc. different on this engagement:
CMMC Registered Provider Organization (RPO #1449)
We hold the credential that lets us speak with authority on CUI handling, NIST 800-171, DoD supplier requirements, and the controls that govern whether Copilot is even an option. Four CMMC-RP practitioners on staff. Our team bench has the depth most boutique BI shops cannot match.
Private-AI fleet
Penny is not vapor. We run our own GPU servers in Raleigh and have the inference architecture, the integration plumbing, and the operational practice to deliver natural-language analytics on a client’s data without that data leaving a boundary you can describe to an auditor.
Founded 2002, A+ BBB Accredited
23 years of incident-response, forensics, and managed-IT engagements in North Carolina. Cybersecurity services is a delivery line, not a side gig. We deploy Copilot the way we’d deploy any sensitive-data system — with logs, controls, runbooks, and an exit plan.
Built where you are
Raleigh-based, serving the Triangle and the broader US market through managed IT services in Raleigh, North Carolina and remote engagements. Penny is hosted in Raleigh on Petronella infrastructure.
Frequently asked questions about Copilot for Power BI
Does Copilot in Power BI require a special license?
Yes. Microsoft requires a paid Fabric capacity of F2 or higher, or Power BI Premium P1 or higher, plus tenant admin enablement of the “Users can use Copilot and other features powered by Azure OpenAI” setting. A Power BI Pro or Premium Per User (PPU) license alone is not sufficient. Trial capacities and free SKUs are not supported.
What does Copilot in Power BI do?
Copilot for Power BI generates natural-language report summaries, builds report pages from a prompt, writes and explains DAX queries, suggests semantic-model measure descriptions and synonyms, and answers business questions about your data through a chat pane or a standalone full-screen experience.
Is Copilot in Power BI safe for CMMC or HIPAA data?
It depends on classification, controls, and where your Fabric capacity is hosted. Microsoft does not use your prompts to train foundation models, but data your prompt references is sent to Azure OpenAI for inference. For CUI, ePHI, or attorney-client privileged matter we typically restrict Copilot to non-restricted datasets using Microsoft Purview sensitivity labels and data-loss prevention, or move the regulated workload to Penny — the Petronella private-AI alternative — so inference stays inside the Petronella boundary.
What does Petronella Technology Group, Inc. do in a Copilot readiness engagement?
We size the Fabric F-SKU capacity, configure tenant admin settings, prepare the semantic model (descriptive names, measure descriptions, synonyms, hierarchies, sensitivity labels), implement DLP for Copilot through Microsoft Purview, document acceptable-use guidance for prompt-injection and oversharing mitigation, and run a pilot with one or two business teams before tenant-wide rollout. We also right-size F-SKU capacity at 60 days, when usage data has accumulated.
Can Copilot read sensitivity labels and respect data-loss prevention?
Yes. Microsoft Purview information-protection labels and DLP policies apply to Copilot interactions. You can prevent Copilot from answering questions about datasets above a sensitivity threshold and block Copilot for users in groups that should not access classified content. The configuration is non-trivial and is the single highest-leverage governance step for regulated organizations. We set it up before turning Copilot on for end users.
What is the private-AI alternative to Copilot in Power BI?
Penny is the Petronella Technology Group, Inc. private natural-language analytics layer. It connects to your Power BI semantic model, data warehouse, or file share and answers questions in natural language without sending data to Microsoft, Azure OpenAI, or any other public service. Inference runs on Petronella fleet hardware in Raleigh, NC. It is the recommended path for CMMC L2/L3, HIPAA-restricted, attorney-client privileged, or other regulated workloads where Copilot is not appropriate.
How do you right-size Fabric capacity for Copilot?
We project Capacity Unit (CU) burn based on report views, refresh schedule, semantic-model size, and expected Copilot prompts per day, then start at the smallest F-SKU that comfortably handles the projection (often F2 or F4). We review again at 60 days using the Microsoft Fabric Capacity Metrics app to confirm or resize. Most small and mid-sized firms over-buy F-SKU capacity by 20-30% when sizing without data — the 60-day review usually pays for the engagement.
What are prompt-injection risks in Copilot for Power BI?
Prompt-injection is when malicious or attacker-controlled text in a data column, report description, or external content is interpreted by the language model as an instruction, causing it to disclose, redirect, or fabricate output. In Power BI, attack surface includes user-editable measure descriptions, semantic-model annotations, free-text columns in source data, and any other text the model can read.
Mitigations include constraining Copilot to vetted semantic models, treating measure descriptions as code-reviewed content, sanitizing free-text columns, monitoring Copilot usage through the capacity metrics app, and applying sensitivity labels so the model refuses out-of-scope questions.
How long does a Copilot for Power BI readiness engagement take?
Typically four to six weeks for the assessment, semantic-model preparation, governance configuration, and a pilot with one business team. Larger tenants with multiple workspaces, sensitivity-label rollouts across Microsoft 365, Purview DLP rules, and multiple business teams take eight to twelve weeks. The Copilot Readiness Assessment alone runs 2-3 weeks and produces a written plan without committing to implementation.
Does Copilot work with Power BI Desktop as well as the service?
Yes, but you still need a Fabric F2+ or Power BI Premium P1+ workspace to publish to. In Power BI Desktop, Copilot requires write access to a workspace that is on a paid Fabric capacity or Power BI Premium. There is no free or Pro-only path to Copilot in Desktop either.
What is the F-SKU right-sizing audit?
It is a one-time engagement where Petronella Technology Group, Inc. reviews your current Fabric or Power BI Premium spend against the Capacity Metrics app data, identifies idle or oversized capacity, and recommends a target SKU. Most clients save 20-30% on Power BI / Fabric licensing within 30 days of the audit, often offsetting the rest of the Copilot engagement.
Why work with Petronella Technology Group, Inc. on Copilot for Power BI?
We combine a CMMC-Registered Provider Organization (RPO #1449) cybersecurity practice with an AI-native delivery team running our own GPU fleet for private inference. That means you get Copilot governance designed by people who also run Microsoft Purview, vCISO programs, and CMMC enclaves day in and day out — plus a private-AI fallback for the workloads where Copilot is not the right answer. Founded 2002, A+ BBB Accredited, based in Raleigh, North Carolina.
About the author
Next steps
If your team is wondering whether to turn Copilot on, whether your current F-SKU is right-sized, whether to use Purview to gate Copilot, or whether a private-AI alternative is a better fit for your regulated workloads — Petronella Technology Group, Inc. answers that question for a living. Request a quote and we will scope a path that matches your tenant, your budget, and your regulatory posture.
Petronella Technology Group, Inc. — 5540 Centerview Dr., Suite 200, Raleigh, NC 27606 — 919-348-4912 — contact us. CMMC Registered Provider Organization #1449. BBB A+ Accredited since 2003. Serving the Raleigh-Durham-Cary Triangle and the broader US market.