Practical AI: Stop Teaching AI SQL. Start Letting It Talk to Your Analytics Engine

Over the last six months, we’ve repeatedly seen the same AI analytics solution pattern emerge across organizations of all sizes. Driven by consulting firms, technology vendors, and platform providers alike, the Text-to-SQL pitch has quickly become one of the most dominant narratives in enterprise AI. The promise is simple and compelling: “I ask a question. I get an answer.” No need for SQL, data models, or the governance and semantic layers that typically sit underneath enterprise analytics. For many business leaders, it understandably feels like the long-awaited silver bullet for data accessibility.

And to be fair, Text-to-SQL absolutely has value—it lowers the barrier to exploration and makes interacting with data more intuitive. But as organizations embed AI into operational workflows, executive reporting, regulatory contexts, and customer-facing experiences, the challenge shifts from generating answers to generating trusted answers.  At that point, governance, definitions, lineage, auditability, and consistency become central again—not optional extras. In our latest article, we explore why the real future may not be Text-to-SQL replacing analytics platforms, but AI becoming the interface to governed analytics engines and the semantic layers enterprises have already spent years building

The Industry Is Solving the Wrong Problem

Much of the excitement around AI in analytics has centered on Text-to-SQL. The concept is easy to understand, and many demonstrations by consulting companies and platform vendors alike are compelling. A user asks a question in natural language, an AI model generates a SQL query, and the answer is returned from the data warehouse. For organizations that have spent years trying to democratize access to data, it feels like a breakthrough. Suddenly, anyone can ask questions of enterprise data without understanding schemas, joins, or query languages.

The problem is that generating SQL was never the hard part.

For decades, organizations have struggled with a very different challenge: how to ensure that everyone asking the same business question receives the same answer. The challenge has never been finding data. The challenge has been creating trusted, governed, and consistent analytical responses. While AI vendors are racing to teach language models how to generate SQL, and many data teams are trying to prompt their way to modeling data models, many organizations are overlooking the fact that they have already spent years building systems designed to solve exactly this problem.

We've Been Building Analytics Engines for Years

Long before the arrival of generative AI, enterprises invested heavily in analytics engines. These appeared in many forms: semantic layers, OLAP platforms, financial reporting systems, risk engines, pricing engines, performance attribution platforms, business intelligence environments, and decision support systems. Although the technologies varied, their purpose was remarkably consistent. They existed to translate business questions into governed analytical outcomes.

When an executive asks for adjusted EBITDA by region, a mature analytics platform does not simply retrieve rows from a database. It applies approved business definitions, enforces security rules, calculates metrics according to agreed formulas, and produces a result that can be trusted across reports, dashboards, board papers, and regulatory submissions. The value of these platforms was never their ability to access data. Their value was their ability to produce consistent answers. 

In many ways, these systems have been generating governed analytical responses for decades. The terminology may be new, but the architectural challenge is not. The rise of AI has not created a new requirement for trusted analytical outcomes. It has simply created a new interface through which those outcomes can be requested and consumed.

Why Text-to-SQL Eventually Hits a Wall

Text-to-SQL is an important capability and should remain part of the analytics toolkit. It enables exploration, accelerates discovery, and allows users to investigate questions that may never have been considered before. For experimentation and ad hoc analysis, it is enormously valuable.

The challenge emerges when organizations attempt to use Text-to-SQL as a governed execution layer. At that point, the conversation changes from data access to analytical consistency. Different prompts can generate different SQL. Different models can interpret the same question differently. Changes to schemas, prompts, or model versions can alter behavior in unexpected ways. The resulting answer may look reasonable, but reasonable is not the same thing as correct.

The reality is that Text-to-SQL does not create most enterprise analytics challenges. Those challenges have always existed. What Text-to-SQL does is make them significantly harder by democratizing access to analytics without necessarily providing the governance layer required to support that access.

The Six Challenges Text-to-SQL Makes Harder

1. Consistent Business Definitions

Every organization has metrics that appear simple but carry significant business complexity. Customer churn, adjusted EBITDA, active customer, portfolio return, risk exposure, and customer lifetime value are rarely represented by a single database column. They are business concepts that have evolved through years of operational, financial, and regulatory discussion.

A Text-to-SQL approach requires the AI to infer these definitions from physical schemas, prompts, examples, or historical usage patterns. As adoption grows, small differences in interpretation can produce materially different outcomes. The challenge is not generating SQL. The challenge is ensuring that every user receives the same answer to the same question regardless of how it is phrased.

2. Governance and Auditability

As analytical outputs become inputs into executive decisions, regulatory submissions, customer communications, or operational processes, organizations must be able to explain how those outputs were produced.

That requires versioned definitions, lineage, approval workflows, reproducible calculations, and traceability. While Text-to-SQL can generate an answer, it struggles to demonstrate which approved metric definition was used, which version of a business rule was applied, and whether the same result could be reproduced six months later after model updates, schema changes, or prompt modifications.

Governance is not simply about obtaining an answer. It is about proving where the answer came from.

3. Security and Entitlements

Enterprise analytics is rarely limited by database access. The more difficult challenge is determining which metrics, dimensions, records, and business concepts a particular user is authorized to see.

In a Text-to-SQL architecture, security often depends on the AI generating the correct filters and predicates within the query itself. This introduces unnecessary risk because entitlement enforcement becomes dependent on probabilistic behavior. In governed environments, security is most effective when it is enforced deterministically and independently of the conversational interface.

4. Large-Scale Analytics and Data Mining

Many of the questions that create real business value extend well beyond simple aggregations. Risk calculations, customer segmentation, performance attribution, forecasting, optimization, fraud detection, recommendation engines, and behavioral analysis often involve sophisticated computational logic that has been refined over many years.

These capabilities typically reside within dedicated analytics engines rather than in individual SQL queries. Attempting to recreate them dynamically via generated SQL introduces complexity, duplication, and governance challenges, and often yields inferior outcomes. The more advanced the analytical requirement becomes, the more important a governed analytics engine becomes.

5. Change Management

Business definitions evolve. Regulations change. Product structures change. Source systems are modernized. Every mature organization has processes designed to manage these changes and maintain analytical consistency over time.

In a pure Text-to-SQL environment, those changes must be reflected through prompts, examples, schema documentation, and model behavior. As adoption scales, maintaining consistency becomes increasingly difficult. What begins as a flexible exploration capability can gradually become a critical dependency that is difficult to govern and even harder to test.

6. Trust at Scale

Perhaps the biggest challenge is trust.

Most organizations can tolerate occasional inconsistencies when a small number of analysts explore data. The challenge changes dramatically when thousands of employees, partners, customers, and AI agents begin relying on generated analytical responses.

At scale, governance becomes less about enabling access and more about ensuring consistency. The ability to ask questions is important. The ability to trust the answers becomes critical.

The Answer Is Not Better SQL

The natural response to these challenges is often to improve prompts, provide more schema information, add additional context, or deploy more capable models. While these approaches can improve query generation, they do not address the underlying issue.

The challenge is not SQL generation.

The challenge is governed by analytical execution.

This is precisely the problem analytics engines were built to solve.

For decades, semantic layers, business intelligence platforms, OLAP environments, risk engines, and analytical services have provided a governed computational layer between business questions and enterprise data. They exist to ensure that business definitions remain consistent, security policies are enforced, calculations are reproducible, and analytical outputs can be trusted.

The emergence of AI does not remove the need for these capabilities. If anything, it makes them more important.

The Dashboard Is Not the Thing Being Replaced

One of the most common misconceptions in the AI analytics market is that AI will replace existing analytics platforms. In reality, something else is being replaced.

For decades, users interacted with analytics engines through dashboards, reports, and visualization tools. Those interfaces required users to understand where to navigate, which reports to open, and how to interpret the results.

AI introduces a new interaction model. Instead of navigating dashboards, users ask questions.

The underlying analytics engine does not disappear. It continues to enforce business definitions, security policies, lineage, governance controls, and calculation logic. What changes is the interface. The dashboard is replaced by a conversation.

This distinction is important because it allows organizations to build on existing investments rather than bypassing them.

From Text-to-SQL to Text-to-Analytics

The future architecture is unlikely to be built around AI models generating SQL directly against enterprise databases. A more mature architecture separates intent interpretation from governed execution.

When a user asks a question such as "Show customer churn by region for the last twelve months," the AI's role should be to understand the intent behind the request. The AI identifies the requested metric, the dimensions involved, the required time period, and the context of the user making the request.

The analytics engine then performs the work it was designed to do. It resolves approved metric definitions, applies security policies, selects appropriate data sources, executes governed calculations, and records the lineage of the result.

The AI interprets the question. The analytics engine produces the answer.

Together, they create something far more powerful than either capability can provide independently.

Analytics Engines Are Becoming AI Engines

The most successful AI initiatives over the next decade are unlikely to be built on raw data access alone. They will be built on trusted analytical foundations. Semantic layers, data context models, business ontologies, metric registries, governed analytical services, and advanced analytics capabilities will increasingly become the control plane through which AI interacts with enterprise information.

This represents an evolution rather than a replacement. The analytics engine becomes the trusted computational layer behind every AI interaction. AI provides the conversational experience, while the analytics engine provides consistency, governance, reproducibility, trust, and large-scale analytical computation.

In effect, organizations are not building entirely new AI capabilities. They are extending capabilities they have already spent years developing.

The Future Is Built on Existing Foundations

Every major technology shift creates pressure to start over. Enterprise AI is no exception. Yet history suggests that the most successful transformations rarely discard existing capabilities. Instead, they build upon them.

The organizations best positioned to succeed with AI will not necessarily be those with the newest models. They will be those who have invested in business context, semantic meaning, governance, analytics, and trusted decision-support capabilities. These foundations provide the structure AI requires to move beyond experimentation and into operational decision-making.

The future of enterprise analytics is not AI replacing analytics engines.

The future is AI becoming its next interface.

How M&A Operating System Helps

At M&A Operating System, we are helping organizations build the next generation of AI-enabled Analytics Engines. Rather than treating AI as a replacement for existing analytics investments, we help clients integrate trusted data foundations, business context, semantic models, governance frameworks, and advanced analytical capabilities into platforms that AI can safely and effectively consume.

We believe the future of enterprise analytics is not built on AI generating better SQL. It is built on AI interacting with governed analytical capabilities that organizations have spent years developing. By combining modern AI technologies with proven approaches to data management, semantic modeling, and analytics governance, organizations can move beyond experimentation and towards trusted, repeatable, and explainable decision-making at scale.

Our practical approach to designing and managing data ensures that AI interactions are grounded in trusted business definitions, governed by analytical logic, and supported by transparent decision-making processes. The result is not simply better access to information, but more consistent, explainable, and trustworthy AI-augmented outcomes.

Whether you are modernizing analytics platforms, establishing semantic layers, building business context models, enabling advanced analytics and data mining capabilities, or developing AI-powered decision support solutions, our focus remains the same: helping organizations transform decades of analytical investment into the trusted intelligence layer that powers the next generation of AI.