What shipped, and why it is broader than a permissions tool
CData opened early access to its Connect AI Gateway on September 29, a product that centralizes registration of models, MCP servers, and agents in one place, then enforces permissions when any of them retrieve data or take action. Row and column-level access control with full audit trails is the governance layer most enterprises would expect from a data access vendor, and CData has offered versions of that for years through its existing Connect AI platform and managed MCP server offering.
What makes this release different is the context engine sitting underneath the permissions layer. It pulls business definitions directly from tools like dbt and Power BI, so when an agent is asked a question about revenue, it answers using the metric definition your finance team already agreed on rather than inferring one from raw table names and column headers. SVP of product marketing Marie Forshaw put the problem plainly: "You don't want AI redefining revenue every time you ask a question about revenue."
The accuracy numbers, and why they deserve a second look
CData claims 98.5 percent accuracy across 378 enterprise queries, compared to a 65 to 75 percent range it attributes to competing approaches. That is a meaningful gap if it holds up, enough to change the build-versus-buy math for any team currently stitching together its own semantic layer for agent access. But the comparison set, which competitors, which query types, which definition of a correct answer, was not fully disclosed in the announcement, and a vendor-run benchmark against unnamed competitors is a starting point for due diligence rather than a conclusion.
Any enterprise evaluating this category should ask CData to run the same 378 query set, or a representative sample of it, against their own data estate rather than accepting the published number as a proxy for how it will perform on messier internal schemas. Benchmark accuracy on clean demonstration data and accuracy on five years of inconsistently maintained production tables are reliably different numbers, and the gap between them is exactly what a proof of concept exists to surface before a contract is signed.
The cost claim is the more interesting number
CData says it measured up to a 175-fold cost difference between different models producing an identical correct answer to the same query, a figure that reframes model selection as a line item rather than a technical preference left to individual engineers. For an enterprise running agent workloads at any real volume, routing every query to the most capable, most expensive model by default is the kind of decision that looks reasonable in a pilot and becomes an uncomfortable budget conversation once usage scales past a few hundred users.
The gateway's model routing and token budget management features exist specifically to address that gap, and CData says a future release will add automatic model selection based on cost optimization rather than the static routing policies most teams currently hand-write and rarely revisit. That roadmap item matters more than it might sound: static routing rules degrade as new models ship and pricing shifts, and a system that re-optimizes routing automatically is solving a maintenance problem most data platform teams have not budgeted time for.
This is part of a broader shift in what governance vendors sell
CData's own framing describes a move from governing data access alone to overseeing the complete request-to-action pipeline: the prompt, the model selection, the data retrieval, and the execution that follows. That is the same direction Collibra, Immuta, and the native governance layers inside Databricks and Snowflake are all moving, which suggests the market has converged on a shared diagnosis even where the products differ. Buyers evaluating one vendor in this category should assume the others are building toward the same scope within a year.
For PE-backed SaaS and retail organizations running lean data platform teams, the practical question is sequencing: whether to standardize on one vendor's gateway now while the category is still maturing, or wait for the dust to settle around a smaller set of serious contenders. Early movers get influence over the roadmap and faster internal learning; those who wait get a more mature product and clearer pricing once the market has had a year to consolidate around what actually works in production.
What this means for your AI data access strategy
If your organization has agents or copilots querying production data today without centralized model routing or a shared semantic layer, you likely have some version of the 175-fold cost problem already happening quietly inside your cloud bill, distributed across individual engineering teams who each picked whichever model was convenient when they built their integration. Finding that number inside your own usage data is worth an afternoon before evaluating any vendor, because it tells you whether this is a cost problem, a governance problem, or both.
Put CData, Collibra, and the native options from your existing data warehouse vendor on the same evaluation list rather than treating this as a single-vendor decision, since the category has not yet settled on a dominant pattern. Whichever you choose, insist on testing accuracy and cost claims against your own schemas and your own query patterns before signing, because the gap between a vendor's demo environment and your production data estate is where most of these evaluations go wrong. Build the business case around the cost line specifically, since a CFO will approve a governance purchase far more readily when it is framed as closing a measurable cost leak rather than as a compliance expense with no visible return.
