Coginiti 26.6 is out and you can now call an LLM directly inside CoginitiScript. New llm block type, sits next to your SQL. Write a prompt, declare a typed schema, get structured rows back, then query them downstream like any other table. The model's a pipeline step now, not a bolt-on.
Workday's planned acquisition of Honeydew is another strong validation of the semantic layer market. We know AI needs more than access to data via MCP, It needs governed business meaning. The question is whether that semantic infrastructure belongs to application vendors or spans the enterprise.
I'm concerned about the rampant Chronic Intent Attribution Syndrome (CIAS) — the tendency to interpret a model’s output as evidence of private motives, elaborate plans, and a burning desire to fool the evaluation team — running amok in AI labs.
finance ARR ≠ sales ARR ≠ marketing ARR none of them are wrong. they're all "correct" against three different definitions. now ask an AI agent for revenue and watch it pick a fourth.
CrowdStrike saw 2.5x more detections tied to AI agent behavior than to human activity in parts of Q1 2026. Where your semantic layer runs is a security decision now: your cloud, your data center, or air-gapped.
Weekend question for data folks: if an auditor asked about one number your AI agent produced, which part would take longest to explain? The definition, the transformations behind it, the lineage, or who had access?