AI analytics: semantic context moves into data

Since 7 July 2026, AWS has described how Amazon Quick moves business context from legacy Topics into the dataset itself. For DACH organisations, this is not cosmetic BI housekeeping. It is about where terms, rules and AI context are maintained as binding operational assets.

What is changing

According to AWS, column descriptions, synonyms, calculated fields, custom instructions and business rules can now live directly next to the data through Dataset Enrichment. Previously, legacy Topics and datasets had to remain synchronized. When column names, permissions or calculations changed, business context could drift.

AWS is also repositioning Topics: the dataset carries its own semantics, while Topics become more of a multi-dataset layer for relationships, metrics and business terminology. That matters because analytics assistants do not merely generate SQL. They interpret business terms, metrics and filter logic.

Why this matters for AI analytics

Microsoft makes a similar point for Copilot in Power BI: organisations must prepare their data, semantic models and users. Without that preparation, Microsoft says Copilot is more likely to produce low-quality results. In Fabric, Copilot is positioned as an LLM-based assistant across analytics workloads, with explicit attention to security, privacy and regional availability.

The shared message is clear: AI analytics is only as reliable as the controlled semantic layer beneath it. If “revenue”, “active customers” or “contribution margin” are defined differently in each dashboard, an assistant only scales the ambiguity faster.

What DACH organisations should check now

CIOs and CFOs should stop treating semantic models as pure reporting artefacts. Check which business terms are binding, who approves changes, how permissions are inherited and which datasets are suitable for AI-assisted queries at all.

A pragmatic starting point is one critical management report: which definitions does it use, where are they maintained, and can a Copilot or agent reproduce the same logic reliably? The guiding question is simple: do you have a semantic layer — or only many dashboards with similar names?