Three of the week's larger enterprise announcements had almost nothing to say about models. They were all about where context lives and who is allowed to see it.
Salesforce announced a Trusted Enterprise AI Harness that groups context, agency, action, governance, security, and models into a common architecture so agents share a consistent understanding of the customer and business. Strip the branding and the claim is architectural: agents should stop each carrying a private retrieval stack. If your support agent, your renewals agent, and your forecasting agent each embed the same account notes into separate indexes with separate chunking and separate freshness guarantees, they will disagree about the customer — and you will spend your quarter reconciling three answers rather than shipping a fourth agent.
The shared-context design is the right instinct, but it relocates the hard problems rather than solving them. A single resolved context layer means a single place where permission propagation has to be correct. Row-level ACLs that were merely embarrassing when they leaked into one team's search box become catastrophic when every agent in the company reads from the same substrate. It also centralizes staleness: one ingestion lag now degrades every downstream agent simultaneously, instead of one. Teams building toward this pattern should be measuring permission-check latency and index recency as first-class SLOs, not as ops trivia.
The integrator signal
The second data point is commercial. On September 9, Accenture and Google Cloud announced the Accenture Gemini Enterprise Business Group to accelerate large-scale Gemini Enterprise deployments. Nobody capitalizes a dedicated services organization because prompting is hard. They do it because connectors, entitlement mapping, schema reconciliation, and data residency are hard, slow, and specific to each customer's twenty-year-old systems. That's a useful signal for anyone selling a knowledge product: the labor is migrating toward the plumbing, which means the plumbing is where differentiation and margin now sit.
This isn't a new observation so much as a newly funded one. Gartner's September 2025 Market Guide for Enterprise AI Search made the case that enterprise search has become foundational infrastructure for AI — that every assistant, agent, and automated workflow inside an organization depends on reliably retrieving accurate and trusted information at the moment of need. A year later, vendors are organizing their product architecture around exactly that framing.
What to actually check
The components that decide whether any of this works are unglamorous: semantic search, permission controls, near-real-time updates, and integrations with business tools like Slack, Notion, and Google Drive . When you evaluate a "unified context layer," ask how permissions are resolved at query time versus index time, how quickly a revoked share propagates, and what happens to agent behavior when the connector is an hour behind.
An architecture diagram that names governance as a box is not the same as a system that enforces it. This week produced several diagrams. The implementations are still the open question.