From systems of intelligence to systems ofaAction: Yasmeen Ahmad on the agentic data cloud

In this fireside-chat-style interview at Google Cloud Next '26, host Stephanie Wong sits down with Yasmeen Ahmad, Managing Director of Data Cloud at Google Cloud, to explore how Google is repositioning its data platform from a "system of intelligence" to a "system of action" through what it calls the Agentic Data Cloud.

Ahmad opens by diagnosing a structural failure in the previous decade of data platforms: despite producing sophisticated dashboards, predictive models, and KPI reports, only 10–20% of data insights ever made it into production. Generative AI and agentic systems, she argues, fundamentally lower the cost of that last-mile "productionization" step, enabling data to participate actively in real-time reasoning loops and drive action directly into operational systems like ledgers, marketing platforms, and supply chains.

A central theme is the inadequacy of traditional data quality efforts. Ahmad explains that clean data with good lineage only gets AI agents to roughly 50% accuracy. The remaining 50% depends on context — what she calls the "invisible work" historically stored only in the minds of skilled data scientists who understood the hidden meanings behind data fields, PDFs, and business processes. Google's response is the Knowledge Catalog, which uses GenAI to infer schema, descriptions, and semantic relationships across thousands of unstructured documents (such as PDFs) without requiring humans to manually annotate them. This allows agents to navigate enterprise data estates efficiently without overloading context windows.

The conversation covers Gemini Enterprise as the new "front door" for business users, abstracting away the complexity of underlying data systems. Google has enabled organizations to publish conversational agents built on BigQuery, AlloyDB, and Looker directly into Gemini Enterprise. A notable integration is the Deep Research Agent, now connected to the Knowledge Catalog, enabling it to reason over enterprise data alongside web data and documents simultaneously — compressing what previously required weeks of IT effort into minutes or seconds.

Ahmad describes a shift away from fixed persona-based agents (e.g., a "data science agent" or "data engineering agent") toward intent-driven engineering, enabled by the new Data Agent Kit. Modern models are capable enough to handle end-to-end workflows — ingesting data, wrangling it, selecting models, building visualizations, and deploying applications — without being constrained to a single domain. This mirrors a move from human-style domain specialization toward goal-oriented multi-agent swarms.

On multi-cloud, Ahmad highlights the Cross-Cloud Lakehouse, built on Apache Iceberg and the Iceberg REST Catalog as an open standard that eliminates proprietary format lock-in. Combined with Cross-Cloud Interconnect, which delivers sub-second latency and largely eliminates egress costs, Google can now federate data across AWS S3, Databricks (Unity Catalog), Snowflake (Polaris), and BigQuery without requiring data migration.

Cost and performance at agent scale are addressed directly. Ahmad notes that one human click now generates 10–20 API calls from agent swarms running multi-step reasoning loops, creating massive infrastructure load. Google's response is vertical integration: BigQuery query speeds improved 35% while costs dropped 40% year-over-year; managed Apache Spark with the Lightning Engine is 5x faster than vanilla Spark and 2x better price-performance than proprietary alternatives; BigQuery AI inferencing achieved a 230x reduction in token usage through stack-level optimization; and the newly announced TPU v8 separates training and inferencing workloads to eliminate silicon-level bottlenecks. Ahmad cites the integration of Google Search's hybrid search stack — including complex re-ranking algorithms — into the Knowledge Catalog's retrieval layer as a key differentiator for ensuring agents receive the right context at the right time.

Ahmad closes by sharing customer examples from the conference where agentic workflows have compressed 45-minute human processes down to under a minute, with Google hosting 80 data cloud sessions at Next '26, each featuring at least one live customer case study.