TL;DR: AI doesn’t hesitate when meaning gets murky. It commits, scales, and keeps going. The real AI risk isn’t bad models but the semantic chaos underneath them. A governed semantic layer for AI is what makes data AI-ready, keeps performance high, and prevents financial disaster.

On September 23, 1999, NASA lost a $327 million spacecraft. Nine months into its journey, the uncrewed probe approached Mars, fired its thrusters to enter orbit, and vanished, never to be heard from again. The best explanation scientists could offer was that it burned up in the Martian atmosphere or skipped off it and drifted into space.

The reason, when they found it, was almost insultingly simple. One engineering team had been measuring thrust in pound-force seconds. Another in newton-seconds. Neither had any idea the other was speaking a different language. The numbers looked perfectly normal on both ends. The math was right and the data was correct.

The failure occurred in the interpretation layer, producing the most expensive unit conversion error in history.

NASA spent $327 million to learn that shared data and shared meaning aren’t the same thing. I worry about how much organizations are currently spending to learn this lesson the hard way.

Because this is exactly what I see happening in enterprise AI. Your AI will generate thousands of confident answers across every business unit. But one team’s “customer” is another team’s “account” is a third team’s something else entirely. And AI won’t stop and ask questions. It’ll pick a definition, commit at scale, and keep going.

NASA learned a very expensive lesson, so how do you prevent your organization from making a similar mistake?

The problem hiding underneath the platform

Knight Capital Group’s trading algorithm lost $440 million in 45 minutes in 2012. The software was functioning correctly, but it was executing logic from a previous deployment, activated by a flag that meant one thing in the old codebase and something completely different in the new one. Another “same word, different meaning” scenario that drove one of the largest US equity market makers effectively bankrupt in under an hour.

The problem always lives underneath the platform in the layer nobody migrates, no vendor upgrade touches, and no cloud migration solves: the accumulated weight of what everyone assumed the data meant.

A financial services firm recently shared something terrifying. They had more than 100 distinct definitions of “customer” living across their data estate. Not 100 tables. One hundred definitions. Each one a small variation created by an analyst who needed something slightly different, pulled a local copy, made a tweak, and moved on.

Every one of those decisions made sense in the moment. Together, they built a semantic minefield. And that was before anyone turned AI loose on it. I explored how this plays out in practice in my recent webcast with Snowflake and Databricks expert Adam Morton, “From data swamp to data product: Building a data-driven culture.” Our conversation gets into what it takes to move from fragmented, one-off data delivery to a governed, reusable foundation AI can trust. So, let’s get into the fix.

A semantic layer for AI is essential

Every major data platform has converged on the same answer: a governed semantic layer for AI is the foundation that makes AI-ready data possible. The biggest constraint on enterprise AI is no longer model quality. It’s whether the model understands your business.

Databricks reached general availability of Unity Catalog Business Semantics in early 2026, bringing Metric Views (governed, portable definitions of business logic) into the core data layer. Access controls and lineage now apply to semantic definitions the same way they apply to tables. Open-source implementation means definitions are portable without vendor lock-in.

The practical impact is measurable. AI agents operating on governed semantic context answer structured business questions with dramatically higher accuracy than generic models working without it. The Mercedes-Benz Korea “Talk to Data” deployment demonstrated this directly: AI agents can resolve questions correctly when every KPI, dimension, and business rule is defined once and inherited everywhere.

The semantic layer for AI is the foundation that makes AI-ready data possible. The question is whether your organization has built it into your data estate.

AI doesn’t ask clarifying questions

Human analysts bring something genuinely valuable to ambiguous data: skepticism. They pause. They ask which version of “revenue” the report is using. They send a Slack message: “Wait, is this the old customer definition or the new one?”

AI skips that part entirely. Feed it 100 definitions of “customer” and it won’t surface a warning. It identifies patterns across all of them, selects whichever interpretation fits the context, and delivers a beautifully formatted answer at a speed no analyst could match. Then it does it again. Thousands of times before anyone looks closely at what it actually learned.

There’s no single catastrophic moment that announces the error. There’s just drift that compounds across every query, report, and forecast a leadership team builds strategy around.

This lack of shared meaning also drains productivity. New Omdia research surveying more than 250 enterprise data leaders found that the typical data delivery project takes more than 800 person-hours of effort, and 38% of total data team effort is spent on non-reusable, one-off work. That’s not a pipeline problem but a meaning problem at scale.

And the cost shows up in the platform bill: AI compensates for semantic uncertainty with compute, running more reconciliation queries, spinning up larger clusters. Bills that look like infrastructure problems are actually data definition problems. I’ve worked with organizations running 21 separate profitability calculations across business units. The contradictory interpretations were a disaster. So were the costs.

You didn’t hire AI to be confident. You hired it to be right. The semantic layer is what closes that gap.

How a semantic layer for AI improves performance and costs

Most teams frame the semantic layer as a governance project: documentation, catalog cleanup, taxonomy work. It’s not. It’s an economic control. Ambiguous definitions produce competing transformation logic, which generates reconciliation queries, which requires larger compute, which creates budget volatility. The cascade is predictable and entirely preventable. The semantic layer for AI stops this cycle proactively.

It directly improves performance and costs in three ways:

• Reduced compute waste. Standardized definitions mean transformation logic converges. You stop paying for redundant pipelines and the compute to run them.
• Faster, cheaper AI output. Agents with governed context answer accurately on the first pass, without burning tokens resolving contradictory definitions.
• Reuse at scale. One certified data product supports many use cases. Every team that reuses instead of rebuilding reduces costs.

Omdia’s research backs this up as well. Of organizations that have adopted structured data products, 75% have achieved faster time to insight and 73% have improved data accuracy. Reuse isn’t just a good governance practice, but it’s also a measurable performance strategy.

Building a semantic foundation for AI

“Establish shared meaning across the enterprise” sounds like a project that will outlast several leadership cycles. It doesn’t have to. The teams that move fastest treat meaning as infrastructure, not documentation.

Start with the model, not the data. Capture business intent as a governed schema before a single pipeline runs. Embed “customer” as a structural fact, not a note in a wiki nobody has opened in fourteen months.

Let lineage do the translation work. Once the model is in place, every downstream consumer (analyst, application, or AI agent) works from the same trusted structure. When an AI output surprises you, trace it back to the definition that powered it. Without that chain, debugging AI is archaeology.

Close the dialect loop with accountability. Assign ownership over definitions. Require sign-off before data earns certified status. Unity Catalog Business Semantics encodes this into the platform, but the platform can only surface the governance you’ve established.

AI needs an interpretation layer

The Mars Climate Orbiter engineers weren’t careless. The Knight Capital developers weren’t incompetent. They were talented people working on sophisticated systems with shared data pipelines but with no shared language to run through them.

That’s the layer enterprise AI is missing. Not more data, faster platforms, or better models. AI needs a verified agreement about what the data means, embedded in the infrastructure itself.

Your AI will translate whatever language you give it. If that language is a hundred dialects of the same word, it will translate all of them simultaneously at enterprise scale with complete conviction.

A governed semantic layer for AI provides agreement on terms, definitions, and contextual meaning before it enters the system. Without that, even the most capable AI will devolve into the most confident source of the wrong answer.