Your BI Gives Business Information, Not Intelligence

Business Intelligence is a misnomer. The industry has been shipping Business Information for twenty-five years and calling it Intelligence. The difference is structural, and your analysts are paying the price for it.

The original promise was Intelligence

In 2000, two Stanford researchers, Chris Stolte and Pat Hanrahan, published a paper called Polaris. The claim at its heart:

Because of the size of the data sets, dense graphical representations are more effective for exploration than spreadsheets and charts. Furthermore, because of the exploratory nature of the analysis, it must be possible for the analysts to change visualizations rapidly as they pursue a cycle involving first hypothesis and then experimentation.¹

Three years later, Stolte and Hanrahan co-founded Tableau with Christian Chabot, who drove its commercialization. The mission the company still publishes today is "help people see and understand data." Chabot, its first CEO, put it bluntly in a 2012 InfoWorld interview: the goal was to "wrest business intelligence from the grip of specialists and hand it to a broad swath of business users."²

Notice what that promise was about. Not reporting. Not dashboards. An exploratory loop: hypothesis, experiment, refinement, at the speed of human thought. That is what Intelligence was supposed to mean. Our platform is named after that paper. Michael Behrisch (Utrecht University) and Remco Chang (Tufts University) have spent more than a decade in visual analytics research for exactly this reason. The original specification was never finished.

What the industry actually shipped

Look at the BI industry in 2026. Power BI. Tableau. Qlik. Looker. Every product page leads with dashboards, reports, and KPIs. Further down: "governed self-service," which means IT decides what the business is allowed to explore. That is Information, prepared and served. Competent for flat data and shallow joins. Also a ceiling.

BARC and Eckerson Group found in April 2022 that actual adoption of BI tools sits at 25% of employees and has been stuck around 20% for years.³ Benn Stancil, writing in 2021, called the industry's attempt "solving the wrong thing altogether."⁴ The problem is not that dashboards need to be prettier. The problem is that the exploratory cycle never reached the depth most business questions require.

SQL handles tables. It does not handle relationships at depth. Emil Eifrem, who has built graph databases longer than most people in this field have had jobs, said it cleanly: relational, key-value, and document models all treat relationships as second-class citizens. They are an afterthought expressed in foreign keys.⁵ This is not a feature that the next release fixes. It is a floor written into the query language.

Every decision your organization makes that requires relationship-depth is currently being made on a flattened picture. Your dashboards are not lying. They are looking at the wrong shape.

Generative AI is not the rescue

In 2012, Hadoop was going to unlock business insight. By 2018, Cloudera and Hortonworks merged in what the trade press called ecosystem contraction.⁶ And Gartner had already dropped big data from its Hype Cycle three years earlier. The ROI never materialized. Today, the pitch is that a large language model plugged into your data fixes the exploratory gap. It will not.

Spider 2.0, published in late 2024 and accepted as an oral at ICLR 2025, tested frontier language models against 632 real enterprise text-to-SQL problems on BigQuery and Snowflake databases with more than a thousand columns each. The best available agent framework, built on o1-preview, solved just 21.3% of the tasks. On the simpler BIRD benchmark from 2023, ChatGPT reaches 40.08% execution accuracy, while humans score 92.96%.⁷

The deeper problem is structural. Language models are generative. Generation is not interrogation. Yann LeCun, Meta's Chief AI Scientist and a Turing Award winner: language models "cannot reason in any reasonable definition of the term."⁸ Even at temperature zero, the same question asked twice can return different answers; research shows up to a 70% gap between the best and worst runs on identical inputs.⁹ You cannot audit what refuses to be the same thing twice.

Benn Stancil tested AI on actual analytical work and described the result as "30% nonsense, which I personally believe is too much nonsense."¹⁰ Data analysis, he said, is one of the few things he will not trust AI to do.

Retrieval augmentation reduces hallucination. It does not eliminate it. GraphRAG, the latest wrapper, inherits the failure modes of its underlying graph; a 2025 evaluation titled "How Significant Are the Real Performance Gains?" found that the improvements attributed to GraphRAG are far smaller than published results suggest.¹¹

Generation is probabilistic and non-deterministic storytelling and works very well for brainstorming and ideation. True Business Intelligence must be deterministic evidence. These are different categories of software.

What Intelligence actually requires

The requirements for exploratory analytics were named twenty-five years ago: dense graphical representations, and a rapid cycle of hypothesis and experiment. Not niceties. The minimum specification.

Two more are structural today. Deterministic: the same question, the same data, the same answer, today and after an audit. Traceable: the analyst points at the evidence chain and sees every step from question to conclusion.

Visual, exploratory, deterministic, traceable, and built for the relationships that make business questions actually interesting. That is what Business Intelligence was always supposed to be. The industry shipped half of it and called it done.

We kept the name of the paper on purpose.

See it working on real connected data. Watch the demo →