An executive who used to file a ticket for a custom report can now type "why did revenue drop in the Northeast last quarter" into Tableau Pulse, Power BI Copilot, or Databricks Genie and get an answer in seconds, generated SQL and all. This is a different pressure than the one covered in AI Doomsday's look at business analysts, who bridge stakeholders and requirements. This is about the person who builds the dashboards, writes the queries, and models the data those requests actually run against, and that person's core deliverable is now something a non-technical stakeholder can generate directly.

Business intelligence analysts are rated Moderate risk, with a horizon toward 2029, one of the nearer-term horizons in this series. The underlying capability, translating a plain-English question into a correct database query, has improved fast and is now embedded in every major BI platform. Whether it's reliable enough to remove the analyst depends entirely on something most executives asking the question never see.

Text-to-SQL accuracy is not one number. It swings from roughly 90% on a well-defined, human-built semantic model to something closer to a coin flip on a raw, undocumented schema. That gap, not the headline capability, is where this profession's future is actually being decided.

Key Points

  • Business intelligence analysts are rated Moderate risk with a horizon toward 2029, as natural-language query tools now embedded in every major BI platform let non-technical stakeholders generate reports without a request to an analyst.
  • Text-to-SQL accuracy depends heavily on a human-built semantic layer: Snowflake reports over 90% accuracy on well-modeled data but roughly 50% without one, and a 2026 study of complex real-world schemas found the best AI attempts succeeding only 30 to 36% of the time.
  • Failures are often silent: a wrong join or misapplied business rule doesn't crash the query, it returns a plausible, confidently wrong number that looks like a correct report.
  • Gartner has predicted that by 2030 the semantic layer, the human-defined map of what a business's data actually means, will be treated as critical infrastructure comparable to cybersecurity.
  • No BLS category tracks "BI analyst" directly; the closest proxies, operations research analysts and data scientists, are both projected to grow well above the average occupation, suggesting the field is bifurcating rather than shrinking outright.

What a Business Intelligence Analyst Actually Does

The job is building the infrastructure that turns raw transactional data into something a business can act on: writing the SQL and ETL pipelines that clean and join source data, designing the data models and semantic definitions that decide what "revenue" or "active user" actually means across the company, and building the dashboards and reports that make all of it visible. It is a technical role, closer to data engineering than to the stakeholder-facing work of a business analyst.

What AI Is Already Doing

Text-to-SQL has become a standard benchmark category, with academic test sets like Spider and BIRD showing steady gains, BIRD-level execution accuracy around 73% on its curated real-world databases. In production, Snowflake's Cortex Analyst, Databricks' AI/BI Genie, general availability since mid-2025, and Google's Gemini integration inside Looker all let users query data conversationally, generating and executing SQL behind the scenes. Looker's own data suggests a well-built semantic layer cuts natural-language query errors by roughly two-thirds compared to letting the model work against raw schema. dbt's Copilot goes a layer further, generating SQL models, documentation, and even draft semantic-layer definitions from project metadata.

THE SEMANTIC LAYER MOAT

Gartner has predicted that by 2030 the semantic layer, the governed, human-authored definition of a company's metrics, joins, and business logic, will be treated as infrastructure on the same level as cybersecurity. That prediction is really a statement about where this job's value concentrated once AI could write SQL: not in the query itself, but in the layer of judgment that tells an AI system what a "customer" or a "completed order" is actually supposed to mean, work that has to happen before any natural-language tool can answer reliably.

Where AI Still Fails

The failure modes are well documented and mostly invisible to the person asking the question. A 2026 study testing AI query systems against complex, real-world schemas, some with more than a thousand columns, found the best three-attempt accuracy landing at just 30 to 36%. Analysts cataloging these failures point to the same recurring causes: ambiguous joins, domain-specific metric definitions the model never saw, and business logic, a subscription model instead of a simple order table, say, that isn't visible from the schema alone. The dangerous part is that these queries don't error out. They return a row count and a chart that look exactly as credible as a correct one.

The Labor Market Reality

There is no dedicated BLS category for business intelligence analyst; the role is split across adjacent classifications. Operations research analysts are projected to grow 12% through 2035 with a median wage around $88,940. Data scientists, the more technical and often AI-adjacent end of the same spectrum, are projected to grow 35%, with a median wage above $120,000. Read together, that pattern looks less like a profession in decline and more like one splitting: the reporting- and-dashboard end of the work is the part AI is absorbing, while the end of the field building and governing the semantic layer, and doing genuine data science, is growing quickly.

How to Use AI as a Business Intelligence Analyst Now

For dashboard and report requests: let AI handle the routine, well-modeled queries directly. Treat requests against messy or undocumented data as your work, not the tool's, until the underlying model is cleaned up.

For your highest-leverage work: invest in building and maintaining the semantic layer, clear, documented, business-approved definitions of what every core metric means. That layer is what determines whether the self-service tools your stakeholders now have access to give correct answers or convincing wrong ones.

For trust: audit AI-generated queries the way you'd audit a junior analyst's work, especially on anything touching a metric with known edge cases. The failure mode here doesn't announce itself.

What I Think

The 2029 horizon feels right because the underlying technology is already deployed at scale; the constraint isn't capability, it's trust, and trust in this field runs entirely through the semantic layer. An analyst whose job is running one-off SQL queries against a schema no one has cleaned up is doing work AI will keep eating. An analyst who owns the definitions that make an AI system's answers trustworthy is doing work that gets more valuable as more of the company starts asking questions directly.

What I'd watch is whether companies actually invest in that semantic layer or just deploy the natural-language tool and assume the accuracy numbers on the demo apply to their own messy data. The 30 to 36% figure on complex real-world schemas is the number that should worry executives more than it currently seems to. It's also the reason this job isn't disappearing on the timeline the flashiest product demos imply.

"The AI will give an executive a confident, wrong answer just as fast as a correct one. Someone has to have already decided what 'revenue' means before that question gets asked. That's the job now."