AI Is Here, But The Hard Parts Haven’t Changed

Publication note

This is from Reis’s personal joereis.substack.com publication, distinct from the Practical Data Modeling newsletter (practicaldatamodeling.substack.com) where the MMA book serialization runs. Same author, different Substack. The survey data overlaps with the organizational crisis piece published on the PDM newsletter — see 2026-02-20-practical-data-modeling-organizational-crisis-89pct.

Core argument

AI adoption is effectively complete — 99.5% of surveyed data professionals use AI tools, 82% daily — but the organizational bottlenecks that have always blocked data teams remain untouched. Legacy systems, unclear ownership, poor leadership direction, and misaligned incentives are human problems that faster code generation cannot solve. Reis frames fundamentals as “gravity” (borrowing from Bill Inmon): you can ignore them temporarily, but they pull you back.

Key findings

The 2026 State of Data Engineering Survey (1,101 respondents) reveals a gap between individual tool adoption and organizational maturity. Claude leads tool usage at 49% among data professionals. 57% say AI makes them write code significantly faster — but faster code is not faster production. A VP/Director respondent captures the tension: leadership expects smaller AI-enabled teams to maintain quality at higher velocity, but the engineers who can actually deliver that are rare.

Data modeling surfaces as a critical priority: 49% identify it as most important for 2027, nearly 90% report at least one modeling pain point (same dataset as the PDM survey piece), and organizations using canonical models fight fires at half the rate of ad-hoc shops (19% vs 38%).

A new form of technical debt is emerging: AI-generated code that nobody fully comprehends. 74% of legacy modernization projects already fail; adding AI-generated systems with thin human understanding compounds the risk.

RDCO mapping

Change management convergence. The Every bundle from yesterday — 2026-04-12-every-missing-layer-ai-adoption — concluded that AI adoption is a people management problem, not a platform purchase. Reis arrives at the same thesis from survey data: the hard parts are organizational. Two independent sources, same finding. This is strong evidence for the consulting positioning.

Weber’s outcome metrics. Reis’s point that faster code generation does not equal faster production delivery maps directly to Weber’s Decision Velocity metric in 2026-04-04-eric-weber-data-team-roi-ai-first. Speed-to-code is an operational metric; speed-to-decision is the value metric. The survey confirms the gap between the two.

Testing framework. The firefighting-tax correlation (ad-hoc 38% vs canonical 19%) reinforces the 01-projects/data-quality-framework/testing-matrix-template. Teams with modeling discipline spend less time on reactive work — the testing matrix is the mechanism that keeps canonical models trustworthy enough to reduce firefighting.

MMA book chapters. Reis is essentially writing the survey companion to his own book. The “pressure to move fast” pain point (59%) is exactly what Ch 12’s synthesis checklist (2026-04-09-practical-data-modeling-mma-ch12-synthesis) is designed to counteract — a structured approach that forces completeness before implementation. Ch 13’s business-process discovery method (2026-04-12-practical-data-modeling-mma-ch13-seeing-the-business) addresses the “unclear ownership” pain point (51%) by mapping actors, bounded contexts, and handoffs before writing any schema.

⚠️ Sponsorship

Sponsored by Ellie.ai (enterprise data modeling tool). The survey data stands on its own — 1,101 respondents is a reasonable sample — but the framing naturally favors the conclusion that data modeling tooling is underinvested, which aligns with the sponsor’s product. The fundamentals argument is Reis’s long-running position, not sponsor-driven.