Data connectivity and integration firm CData Software Inc. today introduced Connect AI Gateway, a platform designed to control how artificial intelligence agents choose models, use tools and access company data. The product enters early access today, extending the company’s existing Connect AI platform and its managed Model Context Protocol offering.
The gateway gives IT teams one place to register models, MCP servers and agents, then set rules for what each can do. It also adds a context engine that draws on the structure of connected systems, business definitions and knowledge gathered from users’ interactions. CData said the combination can improve the accuracy of agents’ answers while limiting data exposure and model costs.
CData already provides connectors to hundreds of business applications and databases. The new gateway builds on those connections to enforce a person’s permissions when an agent retrieves data or takes an action on that person’s behalf. Policies can govern access to specific rows and columns, with an audit trail showing which prompt, model, tool and policy contributed to a response.
CData said a central question for companies deploying agents is whether an answer uses the same business definitions their employees use. A request for “revenue,” for example, can return different results depending on which systems and calculations the agent selects. CData’s context engine can reuse definitions maintained in tools such as Fivetran Inc.’s dbt and Microsoft Corp.’s Power BI, alongside information about fields and relationships in source systems.
The gateway can also capture less formal knowledge from documents, conversations and corrections to agent responses. That information is held in a context graph outside individual AI models, so an organization can apply it when changing models. People can review the graph and determine what becomes shared context.
“You don’t want AI redefining revenue every time you ask a question about revenue,” said Marie Forshaw, CData’s senior vice president of product marketing. .
Model routing
The gateway includes separate routing and control functions for models, MCP tools and agents. It can set token budgets and route requests according to policies, allowing a company to use a the most appropriate model for a task based on characteristics like features and cost. CData plans to add the capability for the gateway to automatically select the cheapest model, executives said.
“Initially, it’s going to be based on policy,” said Will Davis, CData’s chief marketing officer. “Eventually, it will be intelligent enough to see the prompt, see the system behavior, and be able to select the lowest-cost model that connects to the task.”
CData said its data layer can also reduce the amount of work assigned to models to save on token costs. It can filter, join and aggregate records before sending a result to a model, potentially reducing the amount of data placed in the model’s context window.
In a company-run test of 378 enterprise queries, CData said Connect AI answered 98.5% correctly, compared with 65% to 75% for other MCP providers it tested. A separate test found a cost difference of up to 175-fold between models that produced the same correct answer. Those are CData benchmarks, not measured independent tests or documented savings from a customer deployment.
The gateway reflects a shift in what CData is trying to manage. Its earlier Connect AI offering concentrated on giving AI tools governed access to enterprise systems. The new product is meant to oversee the request from the initial prompt through model selection, data retrieval and any resulting action. That broader view could help administrators investigate an agent’s behavior when an answer is wrong or an action is blocked.
CData is positioning its context engine alongside existing data management tools. Executives said it can import business definitions from analytics products, but it is not intended to replace a full data catalog or semantic layer. The initial release doesn’t support imported definitions directly from data catalogs, although that capability is planned.
Although the gateway’s semantic graph has elements of a data catalog, CData doesn’t want to enter that market. “We by no means want to be the be-all and end-all of a catalog,” Davis said. “We just want to use that information to make sure that the interactions that people and agents have with AI are accurate, efficient, and well-governed.”
Customers will also have to decide which lessons from agent interactions become shared knowledge. A correction useful to one employee could mislead others. CData said human review will govern that promotion, making the approval process a consequential part of deployment.