CData launches Connect AI Gateway as one control point for enterprise agents
CData Software said Tuesday it launched CData Connect AI Gateway, an evolution of its Connect AI platform that gives organizations a single control point for the models, tools, data, and actions used by AI agents and the people who work with them.
Companies are moving from chatbots that answer questions to agents that change records in SAP, Salesforce, and data warehouses. Connect AI Gateway is CData’s bet that those agents need one governed path — tools they can call, the company’s own definitions, a permission check down to a single record, and a cheaper model when a cheaper one will do — before anyone hands them that job.
On Tuesday, 29 September 2026, CData Software launched CData Connect AI Gateway. The company calls it the next evolution of the Connect AI platform, and a single control point for the models, tools, data, and actions used by AI agents and the people who work with them. An agent, here, means software that can take a step inside a business system, not only answer a question. The company press page is datelined Chapel Hill, North Carolina, and dated September 29, 2026. It does not print an hour. PR Newswire carried the same release and stamps it Sep 29, 2026, 08:00 ET, which is 8:00 a.m. Eastern. Those lines are CData’s.
What the Gateway does, as CData states it. It connects agents and people to enterprise systems through governed tools. Governed means the company sets the rules the software has to follow. It applies company context at every step. Context, here, is the company’s own definitions, so a word such as revenue means the same thing each time. It enforces permissions down to the record. A record is one row, such as one customer or one invoice, not the whole database. It routes each request to the most efficient model for the task. A model is the AI system that reads the request and writes the answer. CData says the point is to help enterprises move from AI that answers questions to AI that takes action. Those lines are CData’s.
What it already knows. CData says the Gateway is built on its data layer, which it says already powers Palantir, Google, hundreds of enterprise products, and thousands of customers worldwide. Those names, and that “thousands,” are CData’s description of the data layer. They are not a count of companies in the Gateway’s early access program. The Gateway understands the schemas, objects, relationships, and operations of connected systems from the day it is connected. A schema is the shape of the data: the fields, and how they relate. It imports knowledge a company already holds, including metric definitions, business terminology, and existing semantic models. A semantic model is a map of what those business words mean. CData says it keeps what resolved each request, so that context piles up inside the customer’s own environment, under the customer’s control. Schema-aware tools send a model only the information a request needs. Schema-aware means the tool already knows that shape, so it can narrow the rows before the model sees them. CData says that cuts the AI bill while keeping answers accurate across systems. Those lines are CData’s.
The first of five things CData says the design makes possible is enterprise MCP, built in and useful on day one. MCP, the Model Context Protocol, is a shared plug so an agent can call an outside system. CData says hundreds of its connectors are exposed as governed tools over live systems the moment an agent connects. It names SAP, NetSuite, Salesforce, the warehouses and databases, and systems that sit on a company’s own machines, including older ones still in use. One managed platform replaces what CData calls server-per-system sprawl, a separate server for every system. MCP servers a customer already runs are governed under the same policy. Those lines are CData’s.
The second is context that compounds, under the customer’s control. The Gateway starts from the schemas and relationships the connectors carry. It adds the metric definitions, terminology, and semantic models a company already has, so Finance’s “revenue” and Sales’ “enterprise segment” resolve the same way every time. Approved corrections become context that every authorized agent and user inside that environment can reuse. Those lines are CData’s.
The third is security and governance on the model, the tool, and the data. Agents act with exactly the access their users have. Every request carries the identity behind it, a person or an agent. Entitlements, the slice that identity is allowed to see, are enforced where the data lives, down to the record. A change is checked against that system’s own business rules before it is saved. One audit trail traces a figure in an answer back to the prompt, the model, the tool, and the policy behind it. A prompt is the instruction that started the request. Those lines are CData’s.
The fourth is a lower bill where the text adds up. A token is a small piece of text a model is billed for. Schema-aware tools filter, join, and aggregate at the source, so only the answer set enters the context window, the text the model reads for that request. Routing sends each request to the most efficient model that can serve it. Spend is attributed and budgeted by team, agent, model, and tool. Those lines are CData’s. The release does not print a price for the Gateway.
The fifth is that the customer does not have to steer the AI into one stack. CData says there is no warehouse to fill, no model to favor, and no platform to route deeper into. Context lives in a portable graph outside any one model or data platform. A graph, here, is a store of those definitions and the links between them. Leading general models, open-weight models a company can run itself, and models hosted on the company’s own machines all inherit the same definitions. Switch models, and the context comes with the switch. If a model provider fails, CData says the request can move to another provider with the context intact. Those lines are CData’s.
Padmakar Pasala, AI software lead at Adobe, is quoted in the release. He said CData Connect AI gives Adobe’s agents live, governed access to enterprise applications such as SAP, and that onboarding a new application went from weeks to hours. He said a test-automation agent uses it to work inside SAP within the permissions Adobe set, and to drive regression testing on its own. Regression testing is re-running checks to see whether a change broke something that used to work. He said test delivery went from 10 weeks of manual effort to about one, which he calls a 10 times acceleration. He said the Gateway’s policy, reaching every agent action down to the record, is what Adobe needs as more agents come online. That quotation is his, in the release. The 10-week figure is his account of that test agent on Connect AI. The release does not say Adobe is already in the Gateway early access program, and it does not print a second customer’s result.
Amit Sharma, chief executive and co-founder, said CData started by connecting every system a business runs on to whatever needed the data. He said a year ago customers asked how to connect AI to those systems. Now, he said, agents are taking action inside them, and the questions are about control: whether the answer is right, whose permissions applied, and what each request cost. He said Connect AI Gateway answers those questions from one control point. He said that because the data layer is built in, that control and that context reach from the prompt to the record an agent changes, so a company can hand agents, and the people working with them, real responsibility and stand behind every answer and every action. That quotation is his, in the release.
CData Labs research is cited in the same release. Prompts run through CData’s connectors returned correct results 98.5 percent of the time across CRM, project management, data warehouse, and ERP systems, against 65 to 75 percent for competing MCP approaches. CRM is customer-relationship software, such as a sales system. ERP is the software that runs core operations such as finance and inventory. 98.5 percent is about 197 correct results in every 200 prompts in that study. A second study ran 22 models, from cheaper economy models to leading general models, against live enterprise systems. Through governed tools, every model returned the same correct answer, and the least expensive did it at 178 times lower cost than the most expensive. That 178 times is CData’s figure for the gap: the same correct answer, a much smaller bill on the cheap model. CData also says governed tools kept models to the records they were allowed to change, and that no model managed that on raw access, meaning access without those tools. Those figures are CData Labs’, as the release prints them. The release does not name the competing MCP products, the 22 models, or the prompts in the test.
CData says a company moving from an assistant to an agent that carries out a business process needs to know which systems the agent opened, what data it used, which policies applied, and what it did. The Gateway, it says, puts that on one path: grounded in company context, checked against policy at the model, the tool, and the data, and logged from the prompt to the record it touched. Because the path uses CData’s own connectors, CData says a company gets that control without standing up a separate server for every system an agent reaches. Those lines are CData’s.
Availability, as the release states it. CData Connect AI Gateway is available through an early access program beginning September 29, 2026. Early access means a limited program. The release does not say the product is on general sale to every customer today. It does not print a price, and it does not say how many companies are in the program.
The about box describes the company, not a Gateway customer total. It says CData powers AI workloads for Anthropic, Databricks, Microsoft, Google, Palantir, and more than 10,000 customers worldwide. More than 10,000 is that about-box figure. The opening of the release says the data layer already powers Palantir, Google, hundreds of enterprise products, and thousands of customers. Both sentences are CData’s, and they do not use the same count. Neither sentence says those customers are running Connect AI Gateway.
In plain terms, CData told customers on Tuesday that Connect AI Gateway is one control point between an agent and the systems that run the business. The tools are CData’s connectors, exposed over that shared plug, with the company’s own definitions attached. Permissions apply down to a single record. Each request can go to a cheaper model when a cheaper model can answer it. The program that opened on Tuesday is early access. The 98.5 percent figure and the 178-times cost gap are CData Labs’ research, cited in the release. The 10-week testing example is Adobe’s, about a test agent on Connect AI.
The picture is CData’s Connect AI Gateway release graphic. A dark field holds a white card that reads “Introducing Connect AI Gateway,” with the line “One control point for enterprise AI.” The Chapel Hill dateline, September 29, 2026, 8:00 a.m. Eastern, is on the card. It is the announcement image from the company press page.