CData’s AI gateway governs agents’ access to enterprise data
SiliconANGLE Paul Gillin
CData launched an AI gateway that controls how agents pick models and touch company data. It can also track why an answer happened, which should help with both errors and leaks.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
CData is trying to put a guardrail around the messy middle of enterprise AI. Its new Connect AI Gateway, now in early access, sits on top of the company’s existing Connect AI platform and managed Model Context Protocol setup, and gives IT a single place to register models, MCP servers and agents before setting rules for all of them.
The pitch is not just control, but context. CData says the gateway can pull in business definitions, system structure and information gathered from users’ interactions, then use that material to shape answers. The company is aiming straight at a familiar enterprise problem: one team says “revenue,” another means something slightly different, and the model happily blurs the difference unless somebody pins the definition down.
That context can come from places businesses already use, including Fivetran’s dbt and Microsoft’s Power BI. It can also absorb less formal knowledge from documents, conversations and corrections to agent responses, then keep that material in a context graph outside the model itself. That matters because the graph can travel with the organization even when the model changes. People can inspect it and decide what becomes shared context, which is a polite way of saying humans still get the final cut.
The gateway also tries to govern the mechanical side of agent behavior. It can enforce a person’s permissions when an agent retrieves data or acts on someone’s behalf, apply policies down to specific rows and columns, and keep an audit trail showing which prompt, model, tool and policy fed into a response. CData says its data layer can also filter, join and aggregate records before they ever reach a model, cutting the amount of work sent into the context window and, in theory, the token bill.
CData is also leaning hard into cost control. The gateway can route requests based on policy and set token budgets, and the company says it plans to add automatic selection of the cheapest model. In a company-run test of 378 enterprise queries, CData said Connect AI answered 98.5% correctly, while other MCP providers it tested landed between 65% and 75%. A separate test found a cost gap of up to 175-fold between models that still produced the same correct answer. Those are CData’s numbers, not independent verification, but they make the company’s point pretty loudly: model choice is expensive, and vague governance is worse.
My take — AI-written commentary, not fact-checked reporting
This is the right fight. Enterprise AI does not need more “magic”; it needs a paper trail, permission checks and fewer models making up a definition on the fly. The funny part is that the most useful AI product here may be the boring one that tells the flashy model to sit down and read the rules.
Read more about this at: SiliconANGLE