As artificial intelligence (AI) shifts from a novelty to an operational necessity, the nature of how humans interact with data is undergoing a radical transformation. AI agents have made the act of generating code "cheap." However, this abundance of generated code has not diminished the value of Structured Query Language (SQL). On the contrary, SQL has emerged as the essential "common ground": the precise, human-readable interface that allows humans to maintain ownership and responsibility in an era of probabilistic automation.
The paradox of "cheap" generation
In the current AI landscape, writing SQL is virtually free. An AI agent can transform a complex natural language prompt into a 50-line SQL query in seconds. Yet, while the creation of the query is cheap, the cost of an error remains significant. Because the responsibility for data-driven decisions still rests solely on human shoulders, there is a fundamental requirement for the output of an AI agent to be transparent. If an agent produces a "black box" script — the "wall of text" problem — the human cannot truly own the result. To own a result, a human must understand the logic that produced it. To do that, a language that is clear, unambiguous, easy to understand and formally correct is needed.
Back to the caveman language
In the quest for token efficiency, "caveman prompting", the practice of stripping away grammar for blunt, keyword-based instructions, is often treated as a modern "hack" to make requests cheaper by removing the formalities typical of natural language between humans [1].
The result is communication that is direct and efficient. Notably, while caveman prompting describes how humans speak to AI, the same logic applies in reverse: when an AI must communicate data manipulation results back to humans, the ideal language is equally compact, unambiguous, and verifiable. Applying that logic, we would arrive at something very much like SQL: a language that was, fortunately, invented long ago with exactly that goal in mind.
SQL was in fact created as a language that nontechnical people could understand, while ensuring outcomes would be deterministic, verifiable, and explainable.
A language built for humans
SQL's role as the "common ground" was a core design goal from its inception at IBM in the 1970s. Chamberlin and Boyce sought to create a language that "casual users" could use to interact with databases using English-like syntax [2]. This declarative philosophy, telling the computer what you want rather than how to get it, is exactly why it resonates with modern AI, which also operates on user intent. While it looks like English, it is built on the rigorous foundation of E.F. Codd's relational model [3], ensuring that once a query is written, its execution is predictable and verifiable.
A language built for AI agents
A language that is compact and efficient is needed to avoid wasting tokens and time on unnecessary formalities. SQL is something that AI agents can leverage to emit efficient, clear code: easier and less expensive to maintain for both humans and AI agents alike. Given that AI agents have no problems remembering the full set of commands available in SQL, the richer SQL's feature set, the more efficient and precise the generated code will be, without compromising readability. The goal is decidedly not to end up with something like regular expressions: extremely efficient but notoriously hard to read. SQL is the perfect balance.
The future: ownership and understanding
In the AI age, "understanding" is synonymous with "responsibility." If a professional cannot explain why a specific figure was reported, they do not own that figure: The AI does. To reclaim that ownership, humans must be comfortable reading the intermediate SQL that the AI produces. SQL is one of the few query languages to have survived decades of technological churn, precisely because it is an incredibly efficient way to communicate complex data relationships in a way that both machines and humans can agree upon.
Investing in SQL is investing in the future
The evidence is clear, and AI only confirms a trend that has been building for years. Python is rightly celebrated as the dominant language of the AI age, but its dominance is in orchestration: coordinating agents, managing pipelines, gluing systems together. When it comes to actually manipulating data, structured or semi-structured, SQL remains the language of record. SQL is the language for data, full stop. Nearly every major attempt to replace it with something different, whether for structured or semi-structured data, whether in the SQL or NoSQL world, has ultimately ended with those systems embracing SQL. The pattern is too consistent to ignore. The clear consequence is that enhancing SQL with richer semantic capabilities is a highly strategic approach; it augments a universally accepted foundation, enabling both AI models and human developers to execute complex business logic directly at the data layer.
The semantic evolution: giving AI a cleaner shorthand
SQL has actually long been evolving toward this semantic richness; for instance, the powerful MATCH_RECOGNIZE clause [4] was introduced back in the SQL:2016 standard to native identify complex sequential patterns. This capability is exceptionally powerful for advanced use cases like tracking conversion drop-offs in user funnel analysis, spotting security breaches in anomaly detection or monitoring signal health trends through in the most diverse scenarios. By expanding the language to natively express these intricate behaviors, AI agents can replace verbose, token-heavy "walls of code" with concise, declarative queries generated in milliseconds. This semantic density not only can reduce the computational and financial costs of AI token usage but also makes the resulting code significantly more transparent, self-documenting, and remarkably easy for humans to audit and verify.
Standardizing intent with semantic views
Beyond individual query clauses, this semantic shift is materializing at the structural level through innovative database objects like Snowflake's "Semantic Views" [5]. Rather than forcing an AI to deduce obscure physical table schemas and risk disjointed calculations, Semantic Views allow organizations to centrally model business concepts, entities, relationships and key performance indicators (KPIs) directly inside the database catalog using native role-based governance. When an AI agent or a business intelligence tool interacts with these unified, rule-based abstractions, it no longer needs to regenerate complex aggregation logic or guess how physical tables link together. By merging LLM reasoning with centralized semantic models and advanced syntax, we arm AI with the ultimate shorthand to communicate complex data logic back to humans clearly, cheaply and unambiguously.
The path forward
Organizations that invest in SQL literacy, in teams capable of reading and reasoning about AI-generated queries, are investing in accountability. And platforms that place SQL at their core, making it richer, more expressive and more capable, are the natural home for this new era of human-AI collaboration. Choosing a platform where SQL is a first-class citizen is not a conservative technical choice; it is the forward-looking one.
References
- Caveman Skill: caveman/README.md at main
- Chamberlin, D. D., & Boyce, R. F. (1974). SEQUEL: A Structured English Query Language. IBM San Jose Research Laboratory.
- Codd, E. F. (1970). A Relational Model of Data for Large Shared Data Banks. Communications of the ACM.
- Funnel analytics with SQL: MATCH_RECOGNIZE() on Snowflake | by Felipe Hoffa | TDS Archive | Medium
- Overview of semantic views | Snowflake Documentation
