Teradata expands AI assistant with tools for governed agents

Teradata is pushing Tera beyond chatty assistant territory and into something closer to a controlled operator for enterprise data work. The company has added a context engine and an execution system that are meant to let AI agents do useful work without wandering outside company rules.

Tera already lets people use natural language instead of SQL or code. With the update, Teradata wants business analysts, data engineers and database administrators to ask for things like analysis, pipeline building or infrastructure management and have the assistant handle the job inside the organization’s access rules.

The new Tera Context Engine is the memory and policy layer. It pulls in metadata, lineage, business definitions and access policies from databases, catalogs, pipelines and other sources, while avoiding the need to move data first. Teradata says that helps the system trace AI outputs back to their sources and keep policies attached as information moves between systems. That matters if you’re trying to let an agent act on real enterprise data without turning the whole thing into a free-for-all.

Tera Harness is the part that actually does the running. It chooses the tools, models and data needed for a task, follows work across multiple steps and can stop for human approval before sensitive actions. Teradata says it can also pick up again after an infrastructure failure and apply controls before an action runs, which is the sort of feature set that tells you the company knows exactly how nervous customers are about autonomous systems.

There’s also a set of Agent Skills for common engineering and data science work, including writing SQL or Python, optimizing queries, tuning workloads and sizing compute resources. Customers can plug in their own tools through the Model Context Protocol. Teradata is leaning hard on cost claims too: in company-reported testing on SWE-bench Pro with the same Opus 5 model, it says Tera used 73% fewer tokens than Claude Code, finished 42% faster and cut total cost by 58% while completing more tasks. On data-eng-bench, Teradata says it posted a 53% lower cost per reliably solved task than Snowflake’s Cortex Code, based on Snowflake’s published results.

The company does add the usual caution that these numbers come from its own tests and benchmark conditions, not every enterprise workload under the sun. The broader pitch is clear enough: let AI agents do more, but keep the brakes, the policy checks and the deployment options in place. Teradata says customers will be able to choose their models and run workloads in cloud, on-premises or sovereign environments, with the new capabilities due in the fourth quarter of 2026.