Enterprise AI is not waiting on better models. It is waiting on governed data and bounded agents

The data has to arrive before the agent can work

Databricks and SAP announced the next phase of their partnership at SAP TechEd: the SAP Data Foundation makes data from SAP applications available for consumption in the SAP Data Cloud, and SAP data products can be shared with SAP BDC Connect for zero-copy access. The pitch is the elimination of custom extraction pipelines, so enterprises can build AI agents and models on live business semantics rather than stale copies. The companies name Adidas, Zalando, Mondi, and CAF as customers using SAP Databricks and SAP BDC Connect, and say SAF-Holland unified more than 100 million records and generates analyses five times faster than with traditional reporting approaches. m:…7080

The partnership also folds in Genie, described as an AI coworker providing access to governed data with enterprise context, Lakeflow for ingestion and orchestration, Unity Gateway for multi-AI governance and cost control across agents, tools, models, and MCPs, and Lakebase, a serverless Postgres database built for AI apps and agents. A planned bi-directional integration between Genie and Joule is in development. Note the pattern: every component is framed as governance or context, not raw capability. Databricks and SAP are selling the plumbing that makes an agent's answers traceable to business data, because that is the part enterprises cannot build with a model alone. m:…7080

Faster tasks do not make a faster journey

Databricks makes the same point from the other side in a post on wealth management onboarding, where it says onboarding often stalls due to fragmented systems and manual handoffs between teams, even when individual tasks are accelerated by AI. Its numbers, drawn from a Capgemini survey of 144 wealth executives in 2026, are blunt: 60% lacked a unified client view, more than half of global asset managers involve at least 10 teams in onboarding, only 15% report strong integration with their CRM systems, and only 27% report substantial business impact from GenAI, even as 95% of firms scale generative AI across multiple use cases. Opening a private banking account in Singapore takes approximately six weeks, and relationship managers spend 41% of their time on operational work. m:…3438

Databricks proposes what it calls governed case context: a current, versioned, source-linked view of each onboarding case, with a governed supervisor agent routing work across specialist agents and approved tools while involving human reviewers when judgment is required, and Unity Catalog centralizing access controls, lineage, and auditability. Firms should measure success across the whole journey, including cycle time, manual touches, and control effectiveness. The company concedes the specific performance gains of this architecture in a live production environment remain to be demonstrated, which is the honest caveat on an otherwise confident prescription. m:…3438

Bound the agent, not just the user

Snowflake's announcement addresses the other half of the problem: what an agent is allowed to do once it has the data. Its Session Scope feature, now generally available, lets administrators and end users define a privilege ceiling that applies whenever an AI agent is active, operating at account-wide, per-user, and per-session levels, and covering its own agents, Snowflake CoCo and CoWork, MCP integrations, and external agents identified via OAuth or a SERVICE_AGENT user type. It can block administrative roles like ACCOUNTADMIN, SYSADMIN, and SECURITYADMIN, and offers a built-in SNOWFLAKE$DATA_READ_WITH_AI scope for read-only access. Crucially, it can only restrict agent privileges and cannot expand user privileges, and it requires no changes to the user's actual role grants. m:…3208

The feature integrates with Snowflake's Agent Identity model to provide audit columns in QUERY_HISTORY and ACCESS_HISTORY, so security teams can see what agents did. The underlying argument, voiced by Snowflake's Luis Leon, is that agents differ from human users because they reason dynamically, chain actions across systems, and decide what to access at runtime, which is why traditional role-based access control is insufficient. That is the same governance gap Databricks is describing, stated as a security control rather than a data architecture. m:…3208

The roadmap is the remediation

Invisible Technologies states the diagnosis most bluntly: legacy infrastructure is the primary cause of enterprise AI implementation challenges, and data governance is the most pervasive blocker. Legacy systems, in its definition, are defined not by age but by their inability to support modern AI requirements: they generate fragmented data, lack API capabilities for modern workloads, and use access controls designed for human users rather than continuous AI querying. Data silos become hard blockers when training or fine-tuning models at scale, and regulatory frameworks like GDPR and NIST require data handling practices that legacy systems often cannot demonstrate. m:…2125

The company's prescription is to treat data readiness as a parallel workstream to AI deployment rather than a downstream task, noting that organizations that invested in data infrastructure two or three years ago are now scaling AI at production, and that API modernization from batch-oriented to event-driven interfaces is necessary for agentic workloads. Databricks and DataIQ's report Deliberate Decisions, based on peer discussions at the DataIQ Guild 2026, lands in the same place for leadership: it identifies five leadership disciplines and four foundations for AI leadership, context, control, cost, and choice, and argues a successful pilot is not the same as a production-ready capability. It also says organizations should define which decisions remain human-owned. m:…2125 m:…4385

The decision for an executive is therefore not which model to pick. It is whether the organization's data is governed, its agents are bounded, and its legacy systems can feed both. The vendors agree on that sequence and all sell a piece of it. The risk is buying the agent before the governance. m:…3438 m:…2125