Data lakehouse provider Databricks is introducing four new updates to its portfolio to help enterprises have more control over the development of their agents and other generative AI-based applications.
One of the new features launched as part of the updates is Centralized Governance, which is designed to help govern large language models, both open and closed source ones, within Mosaic AI Gateway. The feature is currently in public preview.
“Our research shows that governance is one of the top concerns enterprises have about their AI initiatives as it is complicated by the fact that there are multiple components to the process,” said David Menninger, executive director at advisory firm ISG.
The Centralized Governance ability, according to consulting firm West Monroe’s director of technology and experience Doug MacWilliams, is a “pretty big simplifier.”
“It ensures consistent security, access controls, and compliance, while also cutting costs by eliminating duplicates and streamlining licensing fees. Plus, it makes monitoring and fixing issues like drift or bias easier,” MacWilliams explained.
“All in all this should also simplify the approval process for legal, compliance, and security teams, letting them review and approve models through a single interface,” MacWilliams added.
Single SQL query to run batch inference
In order to help enterprises run an AI query without the need to setup infrastructure, Databricks is adding a new capability named Provision-Less Batch Inference.
The new ability, which is in public preview, is a novel way to run a batch inference via Mosaic AI with a single SQL query and enterprises pay for the infrastructure they use, the lakehouse provider said.
“Provision-Less Batch Inference is a big step forward for AI deployment as it makes it easier to scale AI and saves costs by only using resources when needed,” MacWilliams said.
ISG’s Menninger sees the new ability as a serverless functionality which eliminates the need to set things up in advance.
“Without this capability, developers need to do extra work – they need to provision, or set up, some resources to process the inferencing requests,” Menninger explained.
Additionally, MacWilliams believes that the SQL-based interface makes batch inference accessible to data analysts who don’t have MLOps expertise.
“This opens up new possibilities, like processing millions of customer support tickets overnight to spot trends, enriching product catalog data with AI-generated descriptions, running regular compliance checks, and scoring customer databases for churn risk weekly — all without needing special infrastructure,” MacWilliams explained.
Databricks has also upgraded its previously released Agent Evaluation Review App that now allows domain experts to provide evaluations, send traces for labelling, and define custom evaluation criteria — without needing spreadsheets or custom-built applications.
“By making it easier to collect structured feedback, (enterprise) teams can continuously refine AI agent performance and drive systematic accuracy improvements,” the company explained.
In December, Databricks had updated its Mosaic AI Agent Evaluation module with a new synthetic data generation API that was expected to help enterprises evaluate agents faster.
Genie API to extend data analytics to custom and productivity apps
As part of the update, the data lakehouse provider has introduced the AI/BI Genie Conversation API suite in public preview that is expected to help developers embed
natural language-based chatbots directly into custom-built apps or productivity tools, such as Microsoft Teams, Sharepoint, and Slack.
Genie is a no-code tool with an interface that allows users to analyze data by asking questions about it in natural language. The tool is capable of producing visualizations to explain the data.
“With the Genie API, users can programmatically submit prompts and receive insights just as they would in the Genie UI. The API is stateful, allowing it to retain context across multiple follow-up questions within a conversation thread,” the company wrote in a blog post.
The API, according to IDC research vice president Arnal Dayaratna, not only increases the extensibility of conversational assistants that leverage Databricks data but also bridges the gap between data availability and accessibility, thereby enabling faster derivation of actionable insights.
Another advantage of the API is that it democratizes data access by allowing business users to interact with data using natural language, eliminating technical barriers like SQL expertise.
Alternatively, for developers, the API cuts down on work by offering pre-built conversation features, so they can focus on other important tasks instead of building these interfaces from scratch, said West Monroe’s MacWilliams.
Comparing the Genie API to the recently-released Salesforce Agentforce API, MacWilliams said that the Databricks’ version is more integrated with their data lake and BI tools, making analytics a bit more conversational versus the Salesforce approach of creating standalone agents.
This approach, according to Moor Insights and Strategy principal analyst Jason Andersen, is very similar to the AWS approach with Amazon Bedrock.
Databricks’ strategy and the agentic landscape
Analysts also view the updates as Databricks’ strategy to get closer to enterprise users and increase stickiness of its offerings.
“By unifying the data-to-AI pipeline, Databricks is creating a platform that handles everything from raw data to operational AI, cutting out the need for other products,” said West Monroe’s MacWilliams, adding that this strategy makes their platform stickier, reducing customer churn and increasing revenue by expanding the user base within enterprises.
In the agentic space, ISG’s Menninger believes that Databricks does have an advantage over others as it is approach is more technical, “allowing for the creation of more complex agents potentially automating activities” in any domain with data.
But Menninger believes this advantage comes at the expense of who can create these agents — less likely to be business users.
“All vendors are trying to gain the upper hand in the agent wars. However, much of what is going on today is just ‘agent washing’ – calling chatbots agents. True agentic capability is still complicated and technical. It requires programming,” said Menninger. “Salesforce and ServiceNow seem to be very focused on the conversational capabilities, making it easy to create agents, but perhaps at the expense of what types of tasks the agents can accomplish,” Menninger explained.