Why this is in the vault
Rogojan's year-end read on what data teams will actually spend 2027 doing — not writing more pipelines, but proving their analysis is trustworthy and building the business-context layer AI can't scrape from source systems.
The core argument
Rogojan frames 2026 as the year LLM-assisted work (he names Claude Code specifically) went from novelty to daily reliance across his consulting client base — startups, established SMBs, Fortune 500s. His central claim: the bottleneck in data work is shifting from creating analysis (now fast, commoditized by AI) to trusting analysis. Faster queries and charts don't make the business believe the numbers, especially once "everything your data team creates looks like Claude made it." He expects 2027 data teams to spend more time proving conclusions right — talking to operators, sourcing corroborating data — than writing the first pass of a query.
Second thread: data modeling doesn't go away even as AI writes more of the transform code, because modeling requires business understanding that lives in people's heads, not in ERPs. He's explicit that dumping everything into a warehouse and pointing an AI at it (schema-on-read nirvana) isn't happening in 2027 — business politics and inconsistent source systems prevent it.
Third thread: the metadata/context that lives only in people's heads is where he expects teams to build real leverage in 2027, and he flags the predictable failure mode directly — teams build a knowledge graph that functions as a glorified data catalog instead of something wired into the actual analysis workflow, so the captured nuance never resurfaces when someone asks a question.
Mapping against Ray Data Co
The knowledge-graph failure mode he names — a graph that's "only partially attached to your data warehouse," disconnected from the analysis workflow — is the exact trap RDCO's own typed knowledge graph (~/.claude/state/graph.duckdb, built and queried via the graph-reingest and graph-query skills) is designed to avoid: it's wired directly into vault search and decision-audit citation trees, not a standalone catalog nobody queries.
The stronger hit is his "bottleneck shifts from creating analysis to trusting analysis" thesis. That's precisely the problem RDCO's fresh-eyes critic layer exists to solve — verify-vault-write, verify-strategic-output, verify-dispatch, and the broader "workflow agent output integrity" discipline (seven named failure modes, one gate per chain must hit the primary source) are structural answers to "how do we know this conclusion is right," not just faster generation. Rogojan is describing the demand side of exactly the capability RDCO is building credibility on for phData sales: agents in production that can be trusted, not just agents that are fast.
Curation section
Articles Worth Reading
- "The Pendulums of Data" (Joe Reis, Practical Data Modeling) — decentralization (data mesh) swinging back to centralized governance/semantic layers/AI agent governance; third-party, same author RDCO already tracks via the PDM backfill.
- "Rethinking AI Coding: A Three-Layer Architecture for Software Development" (CodeStrap) — proposes software-architecture layers around AI coding models rather than better models alone; third-party, net new source.
- "Your Data Warehouse Isn't Integrated Just Because the Tables Are in One Place" — self-cross-promo; this is Rogojan's own 2026-07-18 post, already filed as [[2026-07-18-seattle-data-guy-data-warehouse-integration]].
⚠️ Sponsorship
Top-of-article CTA pitches Rogojan's own data-migration consulting ("if you're considering a 9+ month long migration project, then let's chat") via a Substack redirect link — the self-consulting form of the known SDG sponsor pattern (not the Estuary adviser-relationship form this time, and not an explicit "sponsored by me" phrase, but functionally the same first-party pitch). Treat the "bottleneck shifts to trusting analysis" framing as genuine analysis rather than ad copy — the CTA is scoped narrowly to migrations and doesn't bleed into the predictions argument. Curation section this issue was 2/3 third-party, 1/3 self-cross-promo — back to the documented ~50/50 norm after the 2026-10-15 issue's 100%-self-promo outlier.
Related
[[verifier-as-epistemology]] — external corroboration added 2026-11-08: this note is one of seven converging on the LLM-verifying-LLM failure mode
[[2026-07-18-seattle-data-guy-data-warehouse-integration]] — the cross-promoted post in this issue's curation slot; same "integration vs. centralization" vocabulary he reuses here.
[[2026-10-15-seattle-data-guy-solving-the-right-problems]] — prior issue in the same late-2026 run; curation-slot self-promo ratio comparison point.
[[2026-05-26-seattle-data-guy-ai-consultants-utility-thesis]] — his earlier "AI utility for consultants" argument that this issue's "trusting analysis" thesis extends.
[[feedback_workflow_agent_output_integrity]] — the RDCO-side answer to his "how do we know this conclusion is right" question.