It feels as if you asked your HR manager to help with your PBI project. It's terrible, just like most AI features implemented by the big, old players. I remain deeply unimpressed.
These are the thoughts of many of those online who tried to enable Copilot in Power BI and waited for a transformation that never came. The dashboards look the same, the numbers conflict, and analysts, who were supposed to be saved by AI, spend more time explaining why the AI got it wrong than actually making decisions.
Technology isn’t to blame; the weakness is in the base it rests on. What organisations consistently skip is the foundational work that makes those tools worth using.
After years of building enterprise analytics solutions, we keep spotting one truth that keeps surfacing: AI amplifies clarity and chaos alike. Before you introduce AI into your Power BI environment, you need to ask a different question: is our data architecture ready for it?
What "Architecture First" Means
A data model is an information model of your company, a representation of what your business does, the key attributes it relies on, and the relationships between them.
In practice, architecture-first means that before any report is built or any AI feature is switched on, the following are clearly defined:
- Core dimensions and facts that describe business activity
- Critical data elements that form the basis of key performance indicators
- A clear mapping between those data elements and your actual business processes
This last point is often overlooked, and it's the one that costs the most. Architecture does not treat data as a collection of numbers or text strings. It views data through the lens of the value it carries, and asks: what exactly are we trying to measure, and why does that measurement matter to the business?
When this foundation exists, AI becomes extraordinarily useful, helping structure categories, identify segments, validate data quality, and accelerate complex calculations that previously took days to write and test. Without it, AI produces inconsistency at scale.
How Skipping Architecture Dooms Projects
The failure mode is predictable. Companies skip the architectural phase, enable AI features, and find themselves with dashboards showing different numbers depending on who built the report. No shared definition of a customer, a transaction, or a revenue figure.
When AI is applied on top of this kind of environment, it accelerates the chaos. Different models produce different outputs. Metrics become unreliable. Business leaders stop trusting the data, which means they stop using it. The investment in analytics delivers nothing.
When Is Copilot Useful, And When Should You Wait?
Copilot becomes valuable when:
- Data is structured with clear relationships between entities
- Identifiers exist that connect records across tables: customers to transactions, orders to products, contracts to revenue lines
- The data model has no significant duplications or ambiguities
- Dimensions and facts are properly separated
When the model is poorly structured, Copilot begins to hallucinate — drawing connections that are not there, aggregating incorrectly, and generating measures that look plausible but are wrong. The output is confident and unreliable, a dangerous combination for executive decision-making.
The honest advice: do not rush to enable Copilot. Assess your data maturity first. If your model does not yet have clean identifiers and clear dimension-fact relationships, the time spent enabling AI features will cost you more in confusion than it saves in productivity.
Case Study: Predictive Analytics at the Largest Norwegian Dental Group
The client came to us with a problem that many companies share: customer retention was deteriorating, but the data environment could not explain why. The signals were strong — declining engagement, early churn — but without the architecture to surface them clearly, the team had no reliable way to understand how to act.
The engagement began with business processes. We mapped their data to their actual operations, identifying where the gaps were between what was happening in the business and what the data was capturing. That mapping exercise revealed exactly which data elements were missing, duplicated, or misaligned.
From that foundation, we built a data model that powered three specific analytical capabilities:
- Churn rate analysis: identifying which customers were likely to leave before they did
- Frequency-drop detection: surfacing customers whose engagement had declined and who needed re-engagement
- Prioritization signals: giving the team a clear view of which client relationships to address first
The predictive layer was the result of the architectural work that preceded it. The AI and analytics simply made that foundation visible and actionable.
The Roadmap: 3, 6, and 12 Months to AI Readiness
Moving from a basic Power BI environment to a powerful AI-ready one has a certain path, and the milestones are concrete.
Months 1–3: Foundation Define your data model. Identify core dimensions and facts. Map critical data elements to business processes. Resolve duplication and establish clean identifiers. This phase is the most important and the one most often skipped.
Months 3–6: Intelligence Layer Begin using AI to accelerate analytics development — DAX formula generation, automated calculations, pattern detection. With a clean data model in place, AI can reduce report development time from 5–15 days to 1–5 days while maintaining high quality. Introduce structured KPIs and validate them against business outcomes.
Months 6–12: Advanced AI Readiness Evaluate Copilot and AI-driven reporting. Explore MCP integrations that allow language models to connect directly to Power BI and run calculations on demand. Introduce predictive analytics where the data foundation supports it. At this stage, AI is a genuine operational advantage.
The companies that skip to month twelve without doing months one through six will find themselves rebuilding.
Why Expertise Still Matters
You may ask the question: If AI can generate DAX formulas, build measures, and draft reports, why does expert involvement remain valuable? The answer lies in the role of AI: it handles execution, but does not supply judgment.
The decisions that determine whether an analytics environment works, which data model to build, how to map business processes to data elements, how to avoid the duplication and misalignment that cause AI to fail. These are not something AI can make for you. They require understanding data management practice, experience with how these systems work and fail in a real environment, and the ability to see not just the data a client has, but the structure they need.
AI has meaningfully reduced the time required to act on good decisions; it has not replaced the need to make them.
If your organization is investing in analytics and AI-driven reporting, the highest-value question to ask right now is: "Is our data foundation strong enough to make AI work?". The rest is architecture.