Overview
In this episode of The Dashboard Effect, Brick Thompson and Landon Oaks confront a statistic that should give any enterprise AI initiative pause: studies from MIT and others suggest that up to 95 percent of AI projects fail. The hosts are not interested in debating the exact number. They are interested in the reason behind it, and their answer is consistent and direct. Most AI projects fail because the data foundation underneath them is not ready to support them.
The episode lays out what a ready foundation actually looks like, covering the architecture, the semantic layer, and the governance practices that separate AI deployments that produce trusted results from those that produce confident-sounding ones that nobody can verify. See how Blue Margin’s Managed Data Platform helps organizations build the consolidated, governed data foundation that makes AI initiatives reliable rather than aspirational.
What This Episode Covers
Why AI Projects Fail (0:00 – 1:20)
The failure rate for enterprise AI projects is not primarily a technology problem. It is a data problem. Organizations that attempt to layer AI onto fragmented, inconsistent, or poorly governed data find that the technology surfaces the inadequacy of the foundation rather than compensating for it. Understanding that the primary investment required to make AI succeed is a data investment rather than an AI investment changes how organizations should be sequencing their initiatives.
Consolidating Data Silos (1:53)
Most organizations operate with data scattered across multiple systems that were never designed to work together. The hosts advocate for data lakehouses as the central architecture for aggregating that data into a single, governed environment. Consolidation is not just an analytics improvement. It is the prerequisite that makes every subsequent layer of the AI stack work reliably rather than inheriting the fragmentation of the source systems it draws from.
The Role of Semantic Models (3:00 – 5:00)
A semantic model acts as the translation layer between raw data in a lakehouse and the end users or AI tools that need to query it. It standardizes naming conventions, defines business logic, and codifies metrics and KPIs so that every tool and every user is working from the same definitions. Without that layer, the same question asked in different ways produces different answers, which is the condition that makes AI outputs unreliable for business decision-making.
Natural Language Querying Through Celeste (4:27 – 6:30)
By building a clean semantic model, organizations can enable tools like Celeste, Blue Margin’s internal natural language querying tool, to convert plain language questions into accurate SQL queries. The quality of that conversion depends entirely on the quality of the semantic layer underneath it. When the model is well-structured and consistently defined, natural language querying produces reliable results. When it is not, the tool produces plausible-sounding answers that cannot be trusted.
Importance of Data Governance (7:00 – 10:30)
Governance is what makes AI results trustworthy at scale. The hosts cover the specific components that governance requires: maintaining data provenance so the origin of every data point is traceable, cataloging and approving specific transforms so the logic applied to data is documented and auditable, defining error-checking processes that catch quality issues before they reach the AI layer, and implementing security measures like row-level access control that ensure AI tools only surface data to users who are authorized to see it. Without governance, there is no reliable way to audit what the AI is doing or verify why it reached a particular conclusion.
Trusting the Data Before Using It for High-Stakes Decisions (10:30 – 11:41)
The hosts close with a principle that should frame every AI initiative: AI tools are evolving rapidly, but their business value depends on the organization’s ability to trust the data before using it to make decisions that matter. Speed of AI adoption is not a competitive advantage when the underlying data cannot be verified. The organizations that invest in rigor, structure, and oversight before deploying AI are the ones that will get reliable returns from it.
Who It’s For
This episode is worth your time if you are a data or technology leader evaluating why an AI initiative did not deliver what was promised and wanting a clear diagnosis of where the foundation was insufficient, an executive trying to understand what your organization actually needs to invest in before the next AI initiative has a realistic chance of succeeding, a data engineering or architecture team responsible for building the infrastructure that AI tools will depend on and wanting a framework for what that infrastructure needs to include, or any organization that has been told it needs to prepare for AI and wants a concrete and sequenced picture of what that preparation actually involves.
Why It’s Worth a Listen
The failure rate framing is useful because it names a pattern that most organizations experiencing AI disappointment have been reluctant to examine honestly. The technology is not failing. The data foundation is failing, and the technology is making that failure more visible. This episode gives organizations the diagnostic framework to identify which specific foundation components are missing and what building them requires.
The semantic model discussion is particularly valuable for organizations that have consolidated data without yet building the interpretability layer on top of it. Consolidation alone does not make data useful for AI. The semantic model is what allows an AI tool to ask a meaningful question and receive a trustworthy answer, and without it, consolidation is infrastructure that stops one step short of delivering its potential value.
And the governance section is the most critical part of the conversation for organizations evaluating AI for high-stakes use cases. The ability to audit AI outputs, trace data provenance, and verify that the right people are seeing the right data is not a compliance consideration that can be addressed after deployment. It is the foundational requirement that determines whether AI results can be trusted for the decisions that actually matter.
