AI can generate the pipeline; it cannot own the metric definition

The bottleneck moves upstream

dltHub argues that agent-written data pipelines do not remove the slowest part of analytics work: deciding what the business means by “customer,” “revenue” or “active user.” Its August 25 essay separates three jobs that teams often collapse into one: checking whether code matches a specification, choosing which specification to adopt, and remaining accountable for that choice. The first is increasingly automatable; the latter two require authority and a person who can explain the result.

That distinction matters directly to natural-language analytics. An NL2SQL agent can produce syntactically valid SQL against a pipeline whose definitions were never explicitly approved. Faster generation then amplifies an organizational ambiguity rather than resolving it.

A green test is not a business decision

The vendor gives concrete examples: whether a refund reverses revenue in the sale month or the refund month; what inactivity window defines an active user; and how to count a company that signs multiple contracts through subsidiaries. Each can have several internally consistent implementations. Tests can confirm that one implementation behaves as specified, but they cannot decide which policy finance, sales or leadership intended.

dltHub describes one warehouse project that took two weeks to build while a single customer-definition decision took two months of meetings among executives. That is an anecdote from the author, not a benchmark, but it illustrates the operational asymmetry: coding agents can compress implementation time without compressing the time required to reconcile competing business definitions.

Treat definitions as release artifacts

The practical response is not a generic human-in-the-loop approval button. Teams should attach a small decision record to every metric or semantic definition an agent can use:

• the exact business rule and effective date;
• the accountable owner with authority to change it;
• edge cases such as refunds, subsidiaries and late-arriving events;
• the test cases that encode the approved interpretation;
• downstream models, dashboards and agents affected by a change.

That record should travel with the pipeline or semantic model through review and deployment. An agent may propose the SQL and generate tests, but publication should remain blocked until the named owner approves the definition—not merely the diff.

This also changes what teams should measure. SQL execution success, schema validity and regression tests show whether the implementation works. They do not show whether the organization chose the right meaning. A production acceptance check therefore needs two gates: technical conformance and business authority.

The job becomes stewardship, not typing

dltHub’s commercial framing is explicit: it sells infrastructure intended to encode repeatable engineering decisions. Its broader point is still useful beyond that product pitch. As agents automate connectors and orchestration glue, senior data work shifts toward maintaining the theory of the system—why definitions exist, who can change them and how changes propagate.

For NL2SQL deployments, that means the semantic layer is not complete when a model can query it. It is complete when every consequential definition has an owner who can defend it after the generating context window is gone.