AIエージェントは準備完了。あなたのデータは？Your AI agents are ready. Is your data?

Google has published a new State of AI Infrastructure report arguing that the biggest obstacle to deploying agentic AI is not the models themselves but the readiness of the data that feeds them. The report frames a shift many enterprises are now confronting: while autonomous agents capable of reasoning, planning, and acting have matured rapidly, most organizations lack the underlying data foundation to let those agents operate reliably at scale.

The central claim is that agentic systems place fundamentally different demands on data infrastructure than earlier waves of analytics or even generative AI chatbots. An agent that books travel, reconciles invoices, or triages customer issues needs current, governed, and well-structured access to operational data, and it needs that access continuously rather than in scheduled batches. According to Google's framing, data quality, freshness, lineage, and access controls become the limiting factors once models are capable enough to take actions with real consequences. In other words, the bottleneck appears to have moved downstream from model capability toward data plumbing.

To address this, Google positions its Agentic Data Cloud as the proposed solution for enterprises. The concept bundles together the company's existing data platform components, including BigQuery for analytics, AlloyDB and Spanner for transactional workloads, and its broader governance tooling, and orients them toward serving AI agents as first-class consumers. The idea is that agents should be able to query, retrieve, and act on enterprise data through consistent interfaces, with governance and security applied uniformly rather than bolted on per application. Google also ties this to Gemini models and its Vertex AI platform, where agents are built, evaluated, and deployed.

Several technical themes recur in this kind of positioning. One is retrieval and grounding: agents need reliable ways to pull relevant, up-to-date context, often through vector search and semantic indexing layered over existing databases, so that responses are grounded in company facts rather than model guesses. Another is metadata and governance, since an agent acting on behalf of a user must respect the same permissions, data residency rules, and audit requirements as a human employee. A third is interoperability, reflected in the industry's growing interest in open protocols such as the Model Context Protocol and Google's own Agent2Agent effort, which aim to standardize how agents connect to tools and to one another.

The report should be read in the context of a broader competitive push. Rival cloud providers are making similar arguments. Microsoft has emphasized Fabric and its Copilot agents, Amazon has promoted Bedrock agents alongside its data services, and vendors including Snowflake and Databricks have leaned heavily into positioning their data platforms as the substrate for enterprise AI. Google's messaging that data readiness is the deciding factor is therefore both a genuine technical observation and a strategic argument that favors companies, like itself, that sell integrated data and AI stacks. Readers should weigh the vendor's incentive to define the problem in terms its products are designed to solve.

The underlying diagnosis, however, is consistent with what many practitioners report. Enterprise data is frequently fragmented across silos, inconsistently labeled, and wrapped in legacy access patterns that were never designed for autonomous software. Efforts to deploy agents often stall not on reasoning ability but on the unglamorous work of consolidating sources, cleaning records, defining semantic layers, and establishing guardrails. Prerequisite concepts such as data catalogs, feature stores, governance frameworks, and observability all become more important, not less, as agents take on more autonomy, because errors can propagate into actions rather than remaining confined to a text answer.

For organizations evaluating agentic AI, the practical takeaway is that infrastructure investment likely needs to precede or accompany agent deployment. That includes assessing which data is trustworthy enough to expose, instrumenting systems so agent actions can be monitored and rolled back, and clarifying accountability when an agent makes a mistake. Whether or not enterprises adopt Google's specific Agentic Data Cloud offering, the report reinforces a message that is becoming widely shared across the industry: the agents may be ready, but realizing their value will depend on how well the surrounding data foundation is prepared.