From Question-First to Analyst-First: Domain-Expert Skills and Verified Knowledge Compilation for Proactive Enterprise Analytics

TEXT START: Conversational analytics systems assume the user already has a well-formed question, leaving a non-expert facing a blank query box on an unfamiliar enterprise schema.

The Dissection

The paper describes an automation architecture that packages domain expertise into reusable skill folders, compiles dataset knowledge offline, generates standing reports, and re-verifies every metric through executable evidence SQL. Its real function is to remove the analyst from the repeated loop of schema discovery, report construction, metric checking, and question formulation.

The “analyst-first” label disguises the direction of travel. The analyst is not elevated; the analyst’s methods are being decomposed into prompts, references, templates, routing rules, and validation routines so they can be executed without the analyst present. The system converts tacit labor into portable infrastructure.

The Core Fallacy

The central error is confusing verifiability with validity and domain packaging with durable human indispensability.

Re-executing evidence SQL proves that a reported number is reproducible under the selected query. It does not prove that the metric is conceptually correct, that the join preserves business meaning, that the data is representative, or that the resulting interpretation supports a sound decision. Value overlap can detect an empty join; it cannot establish semantic identity, grain compatibility, causal relevance, or absence of selection bias.

Under the Discontinuity Thesis, this architecture is not a defense against cognitive automation. It is an implementation of it. It advances P1 by turning analytical work into a repeatable pipeline and advances P3 by making the analyst’s productive participation less necessary. The report-question-click loop may improve usability, but usability is not economic sovereignty.

Hidden Assumptions

• Schema matching can reliably identify the correct domain skill despite naming collisions, schema drift, undocumented business rules, and client-specific meanings.
• Prompt facets, references, templates, and optional compute are an adequate representation of domain expertise rather than a brittle approximation of it.
• Offline parquet snapshots remain sufficiently current and faithful to production data.
• Join validation by value overlap is meaningful evidence of a valid relationship rather than merely evidence that two columns share tokens.
• Critic gates and self-healing retries catch substantive errors instead of reproducing the same model’s blind spots with more procedural confidence.
• Re-executed SQL validates the metric itself, when it only validates the query’s output against the available data.
• Standing reports can anticipate valuable questions without live business context, changing incentives, exceptional events, or political constraints.
• An open marketplace of expert packs can preserve quality, provenance, security, version control, and accountability at scale.
• Enterprises will trust automated outputs even though the paper supplies no user-study or benchmark evidence.
• The humans who author, govern, and verify the skills remain necessary after enough of their work has been compiled into the catalogue.

The last assumption is the most important and the least defensible. Once the system has accumulated domain packs and verified analytical patterns, the remaining human role becomes narrower, more supervisory, and easier to automate.

Social Function

Primary classification: transition management and partial truth, wrapped in prestige signaling.

The paper identifies a real friction point: non-experts often cannot formulate useful queries against unfamiliar enterprise data. Its engineering responses—offline probing, deterministic skill resolution, critic gates, and executable evidence—are materially better than pretending that fluent prose equals reliable analysis.

But the social narrative is managerial anesthesia. It frames the conversion of expert labor into infrastructure as empowerment and proactivity, while the economic consequence is labor compression. “Domain experts” become authors of depreciating software assets; analysts become maintainers of systems that increasingly encode their own replacement. The architecture does not preserve productive participation. It manages the transition from human-led analysis to machine-mediated reporting while presenting the transition as a better interface.

The Verdict

This is useful analytics plumbing and weak evidence for a durable moat. Its strongest contribution is operational: compile knowledge once, automate recurring analysis, and attach reproducibility to outputs. Its systemic consequence is harsher: it accelerates the severing of expertise from employment.

The paper’s explicit refusal to make benchmark or user-study claims is intellectually honest, but it also means “defensibility” remains an architectural assertion. The skill folder is not a sovereign asset; it is a container waiting to be standardized, copied, improved, and absorbed into a larger agent platform.

In DT terms, this is a transition-management mechanism, not a survival mechanism. The temporary winners are the owners of the data, compute, distribution, governance, and enterprise relationships. The analysts who merely author the packs are servitors at best—and increasingly replaceable ones.