cognitio analytics

Leveraging AI: How a Top Insurer Migrated Platforms Without Losing a Single Metric Definition

Migrating 30 production dashboards from legacy platforms to a cloud, AI-native Intelligence platform, without letting a single metric definition drift.

Overview

A top personal-lines insurance agency in the United States ran its analytics on Tableau, with more than 30 dashboards tracking policy sales, retention, premium, commissions, and producer performance. As the firm moved its data infrastructure onto Snowflake, leadership set a clear mandate: use AI to lift productivity, speed up decisions, and modernize how the business worked. The pressure was real. Across the industry, executives were watching competitors adopt AI and worrying about being left behind. No one wanted their function to be the one standing still.

Sigma Computing was the chosen answer. Built natively on Snowflake, it promised live queries on current data, a spreadsheet-like interface familiar to business users, governance built into the platform, and native AI integration. On paper, it was faster, cheaper, and more modern than the Tableau estate it would replace.

Buying the platform was the easy part. However, more than 30 dashboards still had to be rebuilt in Sigma, and until they were, every real decision was still being made in Tableau. To switch to the new platform, leaders had to be able to trust every number which meant they could not afford to lose a single metric definition. Hence, the risk in recreating thirty dashboards from scratch, was that if each is built by a different person writing their own formulas, the organization ends up with thirty subtly different definitions of the same metric.

Unpacking the Challenge

The real challenge was never the charts. It was the meaning behind them.

Hundreds of components sat inside those dashboards, from simple aggregations to multi-level LOD (level of detail) expressions, parameter-driven filters, and custom table calculations. Four things made moving them safely hard, and every one of them threatened consistency:

  • The knowledge lived in people’s heads. Whoever built a dashboard knew what each formula was meant to compute, but that intent was never written down. Migrate without capturing it, and you are guessing at the definitions. Assumptions create risk.
  • The two platforms don’t speak the same language. A Tableau calculated field does not transcribe into Sigma; every formula must be reinterpreted. Reinterpretation by hand, dashboard after dashboard, is how one metric ends up defined five different ways.
  • Nothing was accumulated. Each dashboard started from zero. A decision made on the third dashboard did nothing for the fifteenth, so the same formula might be translated one way on Monday and another on Thursday.
  • The target kept moving. Sigma shipped new functions throughout the engagement; any reference written at the start was stale within weeks. There was no fixed standard to migrate toward.


Done manually, each dashboard required two to three days. For the full estate, that meant 60 to 90 working days, with no mechanism to guarantee that two analysts ever made the same call twice.

That Is Where Cognitio Analytics Stepped In

Rather than allocating more analysts to the problem, Cognitio Analytics built a migration system around one principle:

The machine does the translation and the documentation. The human confirms the intent. And every confirmed decision makes the next translation more accurate.

What We Did

The system worked in four moves:

1

Capture

Read every workbook; extract every component, full fidelity

2

Translate

Interpret into Sigma against a living reference

3

Confirm Intent

Subject Matter Expert confirms it still means what it should

4

Compound

Every confirmation feeds the next translation

  1. Capture. The system read every Tableau workbook and extracted every component with full fidelity: chart types and their encodings, calculated fields, group-bys and dimensions, aggregations, joins and relationships, filters and parameters, and the table calculations behind each view.
  2. Translate against a living reference. Each component was interpreted into Sigma, not merely transcribed, using a continuously updated reference of Sigma’s current capabilities. When Sigma shipped a new function, the system already knew. The moving target stopped moving.
  3. Confirm intent with the subject matter expert. For every component, the system wrote a plain-English description of what it was meant to compute and put it in front of the person who built the original. Their job was to confirm intent: does this still mean what it was supposed to mean? The reviewer brought the domain knowledge; the system did the mechanical translation.
  4. Compound the learning. Every confirmation was stored and reused. When the same formula pattern reappeared, the system already had a human-validated definition to apply, so the fifteenth dashboard inherited the decisions made on the first.

Each component came out the other side as a single documented definition: what it computes in business terms, the context it lives in, and the steps to build it in Sigma. One metric, one meaning, captured once and reused everywhere it appeared. That documented definition is the direct answer to the fragmentation problem, and the raw material of a governed semantic layer.

Evaluation by Design, not as an Afterthought

Accuracy was not checked at the end. It was measured from the first dashboard.

Most teams answer the question “is the AI right?” with a feeling: the output “seems better” as the project goes on. Cognitio refused to rely on a feeling. From the first dashboard, a custom annotation tool put the firm’s own Tableau experts in the position of labelling each translation against ground truth, in production: confirmed correct, or corrected with the record kept.

This did two things at once. It produced a continuous, human-verified accuracy signal, and those same labels fed back into the system’s compounding loop. Measuring the system and improving it were now one and the same.

The trend was clear. As the body of confirmed decisions grew, the share of first-pass translations accepted without correction rose materially between the early dashboards and the later ones. The system was measurably improving, and that measurement is what made the claim defensible.

For a large, regulated insurer answerable to auditors and state regulators, this matters more than the speed. “We used AI” is a demo. “We measured every translation, with humans in the loop, and tracked accuracy over time” is a system you can stand behind in an audit.

The Benefits

~95%

Faster per dashboard

2–3 days → under 2 hours

100%

Component coverage

All types, incl. LOD & table calcs

60–90 → 10

Working days (30+ dashboards)

Months of work → two weeks

  • Speed without shortcuts. The first dashboard went from upload to reviewed output in under two hours, against a two-to-three-day manual estimate. The time was saved because human attention went only to confirming meaning, not to research, lookup, formatting, and documentation.
  • Accuracy that improved with use. Every confirmation fed the next translation, so the system grew more accurate as it ran. And because every translation was labelled, that improvement was measured, not assumed.
  • Trust, because the work was auditable. Every component carried its definition, the reviewer who confirmed it, and its history. Nothing was a black box.
  • A standardized definition for every metric. One metric, one documented meaning, captured once, not thirty analysts’ thirty versions. The fragmentation problem, solved at the source.
  • Governance over the data models. With every definition documented and traceable, the firm can see, and enforce, that dashboards draw on the same models and the same logic, rather than discovering divergence months later.
  • The foundation of a semantic layer. The migration did not just move dashboards. It produced a structured, searchable record of what every metric means across the BI estate: the foundation a governed semantic layer is built on.

The firm set out to migrate 30 dashboards. What it gained was a single, trusted definition of every number behind them: the one thing a rebuild from scratch would have destroyed.

Explore the Latest

Discover trends and strategies shaping smarter, faster business processes.