Original: Does your AI agent really understand your enterprise context? You need more than a Knowledge Graph, says Google Cloud’s Andi Gutmans by Phil Wainewright. Summarized by Claude AI on September 24, 2026.
The Take
Enterprise AI succeeds or fails on context, and the context layer can’t be a knowledge graph built by humans for humans. It has to be assembled by agents across 100% of the data estate, including the unstructured mess, and it has to keep correcting itself as users reveal what they actually mean. Because agentic work runs end-to-end across many systems, application-specific ontologies from SaaS vendors will fall short, which is Google’s pitch for platform providers owning the context layer. Nobody, Google included, has fully solved it yet, so enterprises are left stitching together context from several vendors with no guarantee those vendors’ agents agree on what anything means.
Summary
Context is now central to enterprise AI: it cuts token spend and raises accuracy by letting a model narrow its search and validate its answers.
Without it, models burn extra inference cycles and still return incomplete or misleading results.
Context covers organizational processes and policies, industry and team vocabulary, the requester’s role and access rights, interaction history, relevant transactional data, and any constraints on scope. A generic LLM knows none of this unless told, much like a new employee who hasn’t been briefed.
Unlike employee onboarding, there are no established disciplines for building the “Systems of Knowledge” that brief agents, so most enterprises are turning to vendors for ready-made context frameworks.
Google, which sells its own such product called Knowledge Catalog, recommends against enterprises building specialized models. Models improve so fast that bespoke work goes stale, and better results come from surrounding the model with the right context, RAG1, and other harnesses.
Existing knowledge graphs don’t solve the problem, because they were built for humans and don’t translate well to agents.
Much enterprise context has never been mapped at all. Knowledge Catalog extends Google’s earlier Dataplex2 with automation, connectivity, metadata analysis, and context curation, plus agentic feedback loops that detect poor outcomes and revise the underlying knowledge automatically.
Humans are too slow to build an enterprise-wide ontology3; only agents building knowledge for agents can cover the whole data estate.
Roughly 90% of enterprise knowledge sits unstructured in PDFs, contracts, and object storage, and none of it has been activated. The approach aggregates metadata from everywhere, including Oracle, SQL Server, and non-Google clouds, then infers meaning from semantic models4 in Looker or Power BI and from query logs showing what users actually ask and how.
Agentic workflows cross many systems end-to-end, so vendor context frameworks rooted in a single application stack such as ERP, HCM, or CRM will fail for enterprise-wide use cases.
Protocols like MCP5 let agents interact but don’t give them shared meaning. A question like “how much revenue did we have last year?” hinges on fiscal versus calendar year and net versus gross revenue.
Context is not a one-time mapping; it requires continuous hill-climbing6, a learning engine that notices when an answer missed and updates the catalog so the next answer is right.
That complexity is why no one does this well yet, and why claims of already having a usable knowledge graph trivialize the problem.
The bet is on platform providers over SaaS vendors, since cross-system agents need cross-system context.
SaaS vendors can still win by opening up to external agents, which drives more workload to them. Salesforce’s headless7 strategy is the example: the value of its accumulated processes and data persists even as its UI matters less.
No vendor is fully there, Google included, though Knowledge Catalog has customers in production.
Enterprises are therefore stuck assembling context from multiple vendors, each strong in its own domain, with no assurance that their agents will share a common understanding.
Links
🤖 Does your AI agent really understand your enterprise context? You need more than a Knowledge Graph, says Google Cloud’s Andi Gutmans – Human-built knowledge graphs won’t give agents usable enterprise context; agents have to build it across the whole data estate and keep correcting it as users reveal what they mean. Google bets that cross-system agents favor platform providers over siloed SaaS vendors, while conceding nobody has fully solved it.
- Retrieval-augmented generation: fetching relevant documents or data at query time and feeding them to the model alongside the prompt. ↩
- Google Cloud’s data management product for organizing, cataloging, and governing data spread across distributed storage. ↩
- A formal model of an organization’s concepts, entities, and how they relate, which software can reason over. ↩
- The layer in BI tools that defines business meanings for raw data, such as how “revenue” or “active customer” is calculated. ↩
- Model Context Protocol, an open standard for connecting AI agents to tools, data sources, and other agents. ↩
- Iterative, incremental improvement toward a goal, here meaning repeated adjustment of the knowledge base to push accuracy upward. ↩
- Exposing an application’s data and business logic through APIs so other systems and agents can use it without its own user interface. ↩
Leave a Reply