Most enterprise AI projects don't fail because of the model. They fail because the data underneath it isn't reliable, accessible, or well-governed. That's the argument CData Software is making — and at the Gartner Data & Analytics Summit this week in Orlando, the company is putting benchmarking data behind it.
CData announced significant enhancements to its Connect AI platform, a managed Model Context Protocol (MCP) server designed to give AI agents accurate, governed access to live enterprise data. The announcement centers on three areas: expanded connectivity through a new Connect Gateway, smarter agent tooling with a new scoped architecture, and tighter enterprise security controls.
The Accuracy Gap Is Bigger Than Most Teams Realize
CData ran an internal benchmark comparing five MCP providers — including unified API, native, gateway, and iPaaS approaches — across 378 queries spanning CRM, project management, data warehouse, and ERP platforms. Every response was scored against a predetermined correct answer. No partial credit.
Connect AI scored 98.5% accuracy (67 out of 68 correct). Competing providers ranged from 65% to 75% — failing on roughly one in three queries.
The gap was modest on simple queries. It widened significantly as complexity increased. The biggest failures across competing providers clustered in four areas: relative date logic, multi-filter queries, semantic interpretation of business terms, and write operations.
Jerod Johnson, CData's Director of Technology Evangelism, explained the date issue clearly: ask an LLM what Q3 means, and it may guess — or get it wrong. CData avoids that problem because its connectors have access to the full source system, including calendars and notes, so the agent doesn't have to infer fiscal calendars. It can look them up.
The write operation failures were telling in a different way. Many competing MCP providers simply don't support write-back actions. As organizations move from AI assistants toward autonomous agents that take real actions, that limitation matters a lot.
Why a 75% Accuracy Rate Is Actually a 75% Failure Rate
The compound effect is where this gets serious. At 75% accuracy per step, a five-step agentic workflow succeeds less than 24% of the time. At 98.5%, that same workflow still completes successfully more than 92% of the time. CData cited Demis Hassabis, CEO of Google DeepMind, noting that even a 1% error rate compounds like interest across thousands of sequential steps.
Johnson put the accuracy thresholds in practical terms:
- Knowledge retrieval needs to exceed 85% to be useful
- Decision support needs to clear 95%
- Autonomous agents need to exceed 98% — because the cost of an error isn't just a wrong answer; it's a wrong action taken without a human in the loop
What's New in Connect AI
The platform enhancements fall across CData's three core pillars: connectivity, context, and control.
Connectivity: A new Connect Gateway extends live data access to on-premises sources behind the firewall — including SAP, SQL Server, and PostgreSQL — without requiring data replication. Connect AI already supports more than 350 data sources.
Context — Agent Tooling: This is the most significant part of the release. CData introduced three types of tools for agents:
- Universal Tools — Eight tools that work consistently across all 350+ connected systems using a SQL abstraction layer. Agents don't get buried in hundreds of system-specific options, which reduces tool sprawl and context window bloat.
- Source Tools — Tightly defined operations tied to a specific platform like Salesforce or Jira, giving teams control over exactly which actions are permitted.
- Custom Tools — Purpose-built queries and operations scoped to a specific workflow, limiting token usage and reducing the risk of unintended data exposure.
Workspaces and Toolkits bundle these tools together. A Workspace defines which data an agent can access. A Toolkit defines which actions it can take. Each combination deploys as a dedicated MCP server — so agents only see and do what they're supposed to.
Control: New governance features include SCIM 2.0 for automated identity lifecycle management and Custom OAuth Applications so organizations can use their own credentials for compliance. The platform enforces per-user authentication with permissions applied at runtime and maintains full audit trails.
The Bigger Problem: Most AI Infrastructure Isn't Ready
CData's own research is blunt: only 6% of enterprises are satisfied with their current data infrastructure for AI. More than half still rely on custom-built connectors and pipelines, and 71% of AI teams spend more than a quarter of their implementation time on data integration alone — time spent on plumbing rather than building anything useful.
CData's identity passthrough approach means if a user connects through Claude or ChatGPT, the system knows who they are inside Salesforce and enforces those permissions automatically. Organizations don't have to rebuild a separate permission layer for AI access. They inherit what they already have.
Bottom Line
MCP is quickly becoming the standard interface between AI agents and business software. But not all MCP providers deliver the same results. CData's benchmark data — self-reported, though the company said it is pursuing third-party validation — shows a substantial accuracy gap that widens as query complexity increases.
For developers and engineers who've spent time wiring together custom integrations and watching AI fail on anything beyond a basic lookup, this is a familiar pain point. The argument CData is making — that the data layer, not the model, determines whether enterprise AI actually works — is one that tends to resonate with the people doing the implementation work.
The full benchmark methodology and results are available at cdata.com/lp/ai-accuracy-whitepaper.