About This Episode
In episode 12 of Data and AI Heads, host Ian Allison sits down with Bill Inmon, widely recognized as the godfather of data warehousing and a pioneer of data analytics. With more than 60 books to his name and decades of practitioner experience, Inmon traces the evolution of data from simple operational systems to the unstructured frontiers of textual and analog data, arguing that the industry's obsession with new tools while neglecting data foundations is its most dangerous blind spot.
Inmon details a 20-year journey developing industry-specific language models as an alternative to the dead ends of NLP, large language models, and small language models. He explains how his company DataVox has built composable, industry-specific taxonomy libraries covering sectors from airlines and banking to medical and telecommunications, which can be assembled into a working taxonomy in under two minutes. He also shares a vivid case study involving Nike, sentiment analysis of public customer comments, and a shoe defect that ultimately cost the company two billion dollars in market capitalization in a single day.
The conversation is a candid, experience-driven argument that data quality and structured analytical foundations are prerequisites for any AI or analytics investment, and that the next peaks for the profession are textual data and, eventually, analog data.
Key Takeaways
- Data quality is a prerequisite for analytics value: no tool, however sophisticated, overcomes garbage-in, garbage-out, and most corporations invest heavily in new products while spending nothing on cleaning their foundational data.
- The computing industry is historically immature compared to fields like medicine, engineering, and accounting, which have thousands of years of accumulated practice, and this immaturity explains the profession's repeated infatuation with new tools over fundamentals.
- Separating operational transaction processing databases from analytical databases was initially heresy opposed by IBM, Ted Codd, and academia, but proved essential and eventually became the foundation of modern data warehousing.
- NLP as an academic discipline set the development of practical text taxonomies back by approximately 50 years because it is full of nonproductive rabbit holes and was never designed to produce a commercial product.
- The correct progression for handling textual data is not LLMs or SLMs but industry-specific language models scoped to commonly used vocabulary, further customized to the individual company, assembled from modular taxonomy libraries in roughly two minutes.
- Medical research is a high-value, underdeveloped application for text analytics: converting doctor's notes into a database format allows analysis across hundreds of thousands of patients simultaneously, yielding insights impossible through manual review.
- Sentiment analysis of public customer-facing text, done properly before the buy decision rather than after the sale, can surface critical product or service failures early, as demonstrated by identifying Nike's shoe defect months before a two-billion-dollar stock drop.
- Analog data represents a third and even less explored peak beyond structured and textual data, with distillation of the small percentage of meaningful signal from vast volumes of analog content being a major unsolved and commercially valuable problem.
What We Cover
Frequently Asked Questions
Why does Bill Inmon say most companies still fail at data analytics despite investing in new tools?
Inmon argues that every analytics product on the market depends on a solid data foundation, but most corporations invest billions in new technologies while spending nothing on cleaning and governing their underlying data. He cites a large insurer that allocated one billion dollars for new products with zero budget for data quality, calling this a direct application of the GIGO principle: garbage in, garbage out, regardless of how advanced the tool is.
What is the difference between an operational database and an analytical database, and why does it matter?
Bill Inmon pioneered the argument, initially considered heresy by IBM and leading academics, that the database structure needed for operational transaction processing is structurally different from the one needed for analytical processing. Attempting to run analytics against OLTP systems produces unreliable results, and building a separate dedicated environment for analytics was the foundational insight that led to the modern data warehouse.
Why did NLP fail to deliver practical text analytics, and what is the alternative?
Inmon argues that NLP was an academic exercise never intended for commercial production and is filled with nonproductive rabbit holes that misled practitioners for decades, setting the field back roughly 50 years. After 20 years of work, his team concluded that the correct approach is industry-specific language models built from commonly used vocabulary, scoped to a specific company within a specific industry, and assembled from modular taxonomy libraries rather than attempting to build universal large or small language models.
How does DataVox build a working taxonomy, and how long does it take?
DataVox maintains a library of modular taxonomy components covering industries such as airlines, banking, insurance, medical, telecommunications, oil and gas, and others, as well as cross-industry libraries for accounting, financial, marketing, and sentiment terms. When a client needs a taxonomy for a specific use case, the relevant modules are selected and assembled in approximately two minutes. The libraries are proprietary, protected with custom security measures, and available as a charged service through DataVox rather than sold outright.
What is the Nike case study that Bill Inmon describes, and what does it demonstrate about voice-of-customer analysis?
Inmon's team collected and analyzed 5,000 public customer comments each for Nike and Adidas after a marketing contact reported Nike was losing share without knowing why. Roughly 40 percent of Nike comments described shoes failing within two months of purchase and Nike refusing to replace them. When this finding was reported to Nike management and then to the company's president, it was dismissed. Months later, a nationally televised college basketball game showed Duke star Zion Williamson's Nike shoe sole separating during play, causing injury. Nike's stock dropped two billion dollars the following day. The root cause turned out to be a change to defective shoe glue, which the public comments had been signaling all along.
Why is medical research such a promising use case for text analytics?
Medical records contain doctor's notes written in free-form text, and those notes hold a significant portion of the clinical value in a patient record. No human analyst can read and synthesize notes across 100,000 or more patients, but by converting doctor's notes into a structured database, analytical tools including knowledge graphs and dashboards can surface patterns across large patient populations. Inmon's team applied this approach to COVID-19 research, identifying relationships between COVID outcomes and factors such as smoking, drug interactions, and comorbidities that would have been invisible without converting text to structured form.
What does Bill Inmon mean when he says the computing industry is immature?
Inmon uses a historical comparison to illustrate his point: Roman engineering walls built 2,000 years ago still stand, accounting traces back to ancient Egyptian hieroglyphics, and evidence of medicine has been found dating back 10,000 years in South America. Computing as a discipline only began around 1960, making it an infant profession by comparison. He argues this immaturity explains why the industry repeatedly chases new tools rather than building the disciplined data foundations that older professions take for granted.
About the Guest
Bill Inmon
Bill Inmon is widely recognized as the godfather of data warehousing, the practitioner who first articulated and fought for the architectural separation of operational transaction processing systems from analytical databases at a time when that position was considered heresy by IBM, Ted Codd, and the academic establishment. He has authored more than 60 books on data management and analytics. Beyond data warehousing, Inmon spent approximately 20 years developing proprietary industry-specific language models and taxonomy libraries for processing unstructured textual data, work that culminated in a commercial offering through his company DataVox. His research spans medical text analytics, customer sentiment analysis, and the emerging challenges of analog data, and he is recognized as a continuing visionary on where the data and analytics profession needs to evolve next.
Connect on LinkedIn →More Episodes