When BI Can't Answer the Question: Building a Decision-Ready View from Raw Data

Data can exist everywhere in a company and still be unusable for a decision.

In one product, behavioural events—opens, clicks and journeys—were captured in Grafana. Customer and financial segmentation lived in a separate Power BI environment. The latter required access to multiple data pools, was slow to work with, and was not built for exploratory analysis.

The business question was straightforward: which customer segments should receive attention first? The existing analytics stack could not answer it quickly or connect the relevant signals.

The problem was not only access

There was no usable specification for the relevant tables. Before any analysis could begin, it was unclear where the required fields lived, how tables related to one another, what their grain was, and which joins could be trusted.

So the first task was not a cohort analysis. It was rebuilding a minimum viable understanding of the data itself.

Build the smallest analytical layer that answers the question

I obtained authorised access to the raw data and extracted only the datasets needed for the decision. With AI support, I then:

1. Reviewed the data index and metadata to map the relevant tables, fields and relationships.
2. Created and deployed a local database for the analysis.
3. Prepared SQL queries, normalised and transformed the data, and validated the links between sources.
4. Connected customer and financial segmentation to behavioural event data.
5. Reconstructed customer journeys, current statuses, activity patterns and cohorts.

AI did not make the business decision. It accelerated the technical work around poorly documented data: exploring the schema, preparing SQL, working with the local database, transforming datasets and forming initial analytical interpretations. The analytical framing, validation and prioritisation remained human-led.

What the analysis revealed

Once the data was connected, two segments stood out.

The first was previously active customers who had become dormant. The second was customers with high financial potential who showed a clear sequence of target actions, signalling intent, but did not consistently reach the business-relevant outcome.

For both segments, the analysis made it possible to see:

• when customers entered the product;
• how their activity changed over time;
• which actions they took and in what sequence;
• their current state; and
• where friction appeared in their path.

The cohort analysis then turned an open-ended research question into a prioritised set of hypotheses and a plan for validating them.

The output was a decision package, not another report

The result was not a generic dashboard or a claim of realised revenue impact. The recommended business and product actions had not yet been implemented, so presenting them as proven outcomes would be misleading.

Instead, the work delivered a practical basis for the next decision:

1. Two customer segments worthy of immediate focus.
2. A clear view of their behaviour, status and barriers.
3. Specific bottlenecks to investigate.
4. A prioritised set of testable hypotheses and next steps.

The validated logic was subsequently moved into the dashboard environment, so the work did not remain a one-off local analysis.

The lesson

When BI is slow, inaccessible or unable to connect the right signals, waiting for a perfect data model can delay an important decision. A carefully scoped, authorised analytical layer can be enough to establish what matters now—provided the calculations are validated and the durable logic returns to the main analytics environment.

For B2B products, that is often the difference between knowing that "the data is somewhere" and knowing where to focus the next commercial or product effort.