Universal data schemas are dead. Build ontologies at decision time.

The argument. Data infrastructure was built on the assumption that transformation is expensive. Olickel argues cheap intelligence inverts that cost, so teams should stop chasing a single universal schema. He describes four schemas, source, storage, decision, and model prior, that only align at the moment a decision is known. A customer means something different to billing, support, and compliance, so a schema divorced from a question is fiction.

The case study. Southbridge built an entity resolution pipeline for Ohio campaign finance records. Before a new model class called Jev launched, the work was expected to cost about $10,000. The team split 800,000 rows into 200,000 specialized cases, ran 90,000 calls to Jev, processed roughly 6 billion tokens, and produced ground truth for every donor case in Ohio for $7.95 on an $800 budget. The number is impressive, but it is one vendor example.

The role split. Models play three distinct roles: designing pipelines, working inside them, and hill climbing. Pipeline design wants slower, heavier models that can reason about the problem. In-pipeline workers should be cheap, fast, and narrow. The third category, called burrito surfing, self-optimizes a workload across cost, speed, and accuracy. The useful point is that organizations often compress these roles into one context window, which pushes cost and error in the wrong direction.

The hidden risk. Model priors, the fourth schema, are a growing error source. One pipeline looked for "in-kind" with an underscore while the file used a capital I and hyphen, so 15,658 contributions were miscounted. Another run turned $784,000 of debt into $5.3 million by storing absolute values of negative loan amounts. Larger models defend priors more strongly. Cross-model checks help: Jev and Luna agreed on 194 of 200 families, and every error sat in the disagreements.

The caveats. The work depends on a stable harness, bounded experiments, methodology, and a defined failure surface. The speaker acknowledges models cannot self-optimize natively yet. What is missing from the talk is engineering effort: 10 days of agent runs, more than 4,000 harness executions, and a team that can build and debug these systems. The price tag excludes that labor. Teams should take the mental model and treat the cost number as a bound, not a promise.

The implication. Ontologies are not universal; they are intermediates between intelligence and result, built when needed and reused with receipts. That reframes data platform work: schema migration projects become less valuable than decision-time transformations, verification rigs, and model-role design. Leaders should watch for agent cost, error budgets, and prior failures, not just model capability.