Investment in AI continues to grow at dizzying rates, but there are also clear signals emerging that many organizations are questioning how this will ultimately turn into business value. In fact, recent research suggests this problem is widespread. MIT's State of AI in Business 2025 report found that while 60% of organizations evaluated generative AI tools, only 20% reached pilot stage, and just 5% successfully reached production deployment.
In their 2026 predictions brief (Predictions 2026: AI Moves From Hype To Hard Hat Work), Forrester Research suggested that enterprises will delay 25% of their AI spend through 2027 due to the fact that only 15% of CFOs reported measurable EBITDA lift from their AI investments.
Of course there are dozens of really good reasons why these numbers are where they are, but in the world of analytics it seems one strong focus is the area of trust and confidence in responses from AI systems.
Over the past two years, enterprises have rushed to invest in AI analytics platforms, conversational BI tools, text-to-SQL interfaces, and AI copilots.
Executives see exciting and flashy demos that show natural language queries generating instant charts and insights. It promises a future where business users simply "ask questions of the data" without relying on analysts or dashboard developers.
For many organizations, it feels like the long-promised democratization of analytics has finally arrived, but a large percentage of enterprise AI analytics initiatives never fully scale into production decision-making systems.
Some remain trapped in pilot mode. Others become "assistive" tools used only for exploratory analysis. Many generate early enthusiasm but slowly lose organizational trust and adoption.
The issue is not that AI isn't working. The issue is that most AI analytics systems understand language, but not the business context behind the language.
And in enterprise analytics, that distinction matters enormously.
The Cool AI Demo Problem
One reason AI analytics initiatives gain early momentum is because demos can work exceptionally well. In controlled environments, with carefully curated data, conversational analytics appears almost magical:
- clean datasets,
- simplified schemas,
- curated business logic,
- well-defined metrics,
- and predictable questions.
Under these conditions, natural language interfaces can generate impressive results very quickly.
But enterprise environments are not controlled environments.
Real organizations operate across:
- fragmented data platforms,
- inconsistent KPI definitions,
- undocumented business logic,
- siloed departments,
- and years of accumulated semantic drift.
Even something seemingly straightforward like "customer churn" may mean entirely different things to:
- finance,
- customer success,
- product teams,
- or operations.
And the AI system, using its core probabilistic based reasoning powers, will always run the risk of coming up with an incorrect interpretation and response.
And this is where organizations start running into trust issues, stalling the roll-out and ultimately not getting anticipated value.
The challenge is not simply language interpretation. It is business specific contextual interpretation and the standardization of that across the enterprise.
The Industry Confused Language Fluency with Business Fluency
Much of the current AI analytics market has focused heavily on the interface layer:
- chat interfaces,
- prompt engineering,
- text-to-SQL generation,
- conversational UX,
- and LLM integration.
These capabilities are important. But they are not sufficient for enterprise-grade analytics.
Why?
Because enterprise analytics is fundamentally about semantic consistency and business interpretation, not simply query generation. And being able to consistently repeat that across hundreds or thousands of users to always provide the right answer.
Generating SQL is relatively easy compared to understanding:
- how your company defines revenue, active customer, or platinum customer segment,
- which metrics are officially governed,
- what exclusions apply,
- which data sources are trusted,
- or how business policies alter calculations.
This is where many AI copilots quietly struggle.
Most copilots are optimized for productivity assistance:
- helping analysts create queries faster,
- summarizing dashboards,
- or accelerating report generation.
That is valuable. But productivity assistance is not the same thing as trusted decision intelligence.
There is a major difference between:
"Here is a SQL query."
and:
"Here is the correct business answer according to your organization's approved logic, definitions, governance rules, and contextual relationships."
The first is language generation.
The second requires business understanding.
The Silent Trust Collapse
The most dangerous part of failed AI analytics projects is not technical failure.
It is trust erosion.
At first, users are impressed. Business leaders begin experimenting with conversational interfaces. Teams test language analytics capabilities. Executives ask ad hoc questions previously routed through analysts.
Then subtle inconsistencies begin appearing.
- A revenue number differs from the dashboard.
- A margin calculation excludes returns.
- A customer count changes depending on phrasing.
- Two departments receive different answers to the same question.
Eventually users begin double-checking everything. Once that behavior starts, adoption slows dramatically.
Business users revert to:
- ask the analysts,
- dashboards,
- spreadsheets,
- and manually validated reports.
The AI system becomes "interesting," but it cannot be rolled out to the end users.
This pattern is appearing across many enterprise AI initiatives. Gartner has projected that more than 40% of AI initiatives may ultimately be abandoned due to unclear business value, governance concerns, and risk management challenges.
The critical insight is this: in analytics, trust matters more than novelty.
A dashboard that is slow but trusted will almost always outperform a system that is fast but inconsistent.
Why Copilots Alone Don't Solve the Problem
The market currently uses the terms:
- copilots,
- conversational analytics,
- AI assistants,
- and AI agents.
But these are not the same thing.
Most copilots are fundamentally productivity tools attached to some traditional database or technology platform to assist with helping to make the platform easier to use. That, at least, is the promise to the buyer. The reality is probably more along the lines of the CEO in the vendor org demanding "we need to add AI to our platform."
They help users:
- generate SQL,
- summarize information,
- navigate applications,
- or automate tasks within a platform.
And yes, they do allow conversations, but since their "context" is limited to the technical metadata or user supplied prompts at run time, the AI has no ability to consistently reason correctly about your business. Real, effective conversational analytics requires something much deeper in addition to the technical database metadata:
- business knowledge,
- semantic consistency,
- contextual understanding,
- and governed interpretation.
This is why many copilots potentially perform well during demonstrations but struggle in enterprise-wide deployments.
They are schema-aware, not context-aware. A schema-aware system understands tables and columns.
A context-aware system understands:
- business definitions,
- organizational vocabulary,
- policy logic,
- approved metrics,
- lineage,
- and semantic relationships.
That distinction becomes increasingly important as organizations attempt to scale AI analytics across departments and decision-making processes. Without shared contextual understanding, every AI interaction becomes vulnerable to inconsistency.
And inconsistency destroys trust.
The Missing Layer: Governed Context
This is the point many enterprises are now beginning to recognize. AI analytics systems need more than data access and data understanding or AI that simply helps make the platforms easier to use.
They need a deep, certified foundation of contextual understanding.
That contextual understanding often includes:
- business terminology,
- semantic relationships,
- governed metrics,
- organizational policies,
- data lineage,
- domain knowledge,
- synonyms,
- validation rules,
- historical business meaning,
and of course it needs to know and understand your database, relationships, tables, and columns.
This is where contextual semantic layers are becoming strategically important. Traditional semantic layers were designed primarily for BI tools and dashboard modeling.
Modern AI-driven analytics requires something more sophisticated: a context-aware semantic foundation capable of helping AI systems reason consistently across structured and unstructured business knowledge.
Platforms like Codd AI are positioning around this exact problem, focusing not just on natural language interaction, but on creating contextual semantic layers that help AI systems understand business meaning, governance, and organizational context.
Because ultimately: AI analytics does not fail because the model cannot generate language. It fails because the organization has not taught the AI how the business actually works.
The Organizations Succeeding with AI Analytics Are Doing Something Different
The enterprises successfully scaling AI analytics tend to share several characteristics.
They are:
- standardizing business definitions,
- governing semantic models,
- integrating structured and unstructured knowledge,
- enabling human validation workflows,
- prioritizing explainability,
- and investing heavily in contextual consistency.
In other words, they are building AI trust layers.
The strategic conversation is slowly shifting from:
"How do we query data using AI?"
to:
"How do we ensure AI understands the business consistently?"
That shift is profound. Because it moves the focus away from prompts and interfaces, and toward organizational knowledge architecture.
The future winners in AI analytics will likely not be the companies with the most copilots. They will be the organizations with the strongest contextual understanding of their business.
But Where Do We Find the "Business Knowledge" to Incorporate Into the Contextual Semantic Layer?
This is turning into the million dollar question for many companies. What do we do when our business does not have all our rules and logic clearly documented? The truth is most companies are in this position.
From working with data and analytical systems over many decades, the one thing is super clear: no amount of database metadata, database content or database query logs will expose much about your business logic and rules. The reason is that most of the business knowledge lives in local calculations in BI reports, dashboards, Excel spreadsheets, and in people's minds.
Numerous so-called contextual semantic layers are now emerging that claim to automate the collecting of your "business knowledge" via scraping database query logs. The end result is pretty disappointing and what's more it generates more noise than what is actually useful beyond the basic database schema and relationships.
While there is no magic bullet to just automatically discover all your business rules, there are a few approaches to getting there:
- Don't boil the ocean. Do not try to solve for every use case in your entire enterprise from the start. Pick a use case and focus on that single use case. Then move to the next.
- Start with what you do have. Start documenting the most important rules and logic that drive the most impactful business metrics. That way, you build your knowledge foundation in an iterative manner.
- Divide and conquer. Crowd sourcing from your subject matter experts is another avenue to explore.
- Moderate inputs. Have a moderation process whereby knowledge definitions and rules need to be reviewed and certified before they are allowed into the system.
- Rinse and repeat. By continually expanding on the knowledge base you can grow your trusted foundation over time.
The Future of AI Analytics Depends on Context
GenAI is here to stay. It is opening a new paradigm with its ability to produce insights in a very consumer friendly manner. It can interpret natural language questions and take the prompts and produce amazing insights and results.
What is increasingly clear though is that AI alone is not able to consistently be accurate, yet. We are seeing evidence of this in multiple facets of using AI, from writing articles to coding and most certainly analytics. Whether it is an expert reviewing every response from the AI or whether we develop trust frameworks (such as the contextual semantic layer for AI analytics), we will need to have a governed layer to ensure results are accurate and consistent.
Getting to the business knowledge that ultimately will sufficiently enrich the semantic layer to correctly reason about my business and data might not be trivial, but platforms like Codd AI are focused on this exact challenge.
Introducing Codd AI
Codd AI is the leader in the emerging category of contextually-aware AI platforms. Codd AI automatically builds a governed Contextual Semantic Layer comprising your technical metadata, business knowledge, and rules. This semantic layer then becomes the foundation for AI to drive conversational analytics, reasoning about your data and your business like one of your own subject matter experts.
If you are interested in learning more, visit us at www.codd.ai or schedule a quick chat!