Today, Google Cloud launched Google Agentspace, an out-of-the-box standalone application that essentially acts as a generative AI-powered conversational search (and action) layer atop an enterprise’s various data sources. Agentspace can be embedded into an existing intranet and/or accessed via Google’s NotebookLM, an AI-powered research assistant. Agentspace is not part of Google Workspace and does not have any dependency on Workspace.
One key goal of Agentspace is to reduce the time and effort it takes for desk workers to find the information they need to do their jobs. “An internal Google study found that the enterprise workers use an average of four to six tools just to ask and answer one work question,” said Raj Pai, VP Product Management, Cloud AI at Google in a prebriefing.
Integrates with Third-Party Data Sources
To reduce the amount of ‘alt-tabbing’ among applications and the multiple data sources every enterprise owns, Agentspace includes built-in connectors to pull in data from various third-party applications to build an index, as well as an enterprise Knowledge Graph, that connects entities from, as one example, Jira to Salesforce entities which are then moved into the Google Cloud environment. This ‘grounds’ Agentspace on the enterprise’s own data, which helps the AI generate responses that are more relevant to that enterprise. Translation is built in, so data sources in other languages can also be accessed. And, Agentspace can access structured and unstructured data sources.
“Entities in this context means the various objects in the associated third-party enterprise applications. In essence, how a Jira ticket is associated or related to a Salesforce account/activity/meeting notes. So, when you search for information related to a certain customer, Agentspace can blend and associate data from both applications within Google Agentspace,” wrote Stuart Moncada, Product Management, Google Agentspace at Google Cloud in a follow-up email. “Google Cloud is maintaining an index of the data within our platform to have a truly blended and improved search experience. So, yes, the Salesforce data in this instance would be moved over to Google and that is how we build the enterprise graph and power a Google-quality search across applications instead of just joining search results.”
The prebriefing included a demo of this functionality at work – using Agentspace to find Jira tickets, summarizing them (via Gen AI) and then sending an email. That scenario is illustrated below. Note that Agentspace consolidated several steps into one: it searched for the open tickets (citations on the right-hand side) and then summarized those tickets (left-hand side) and then enabled the user to instruct Agentspace to send an email of the results. Pai noted that this functionality “takes advantage of the planning, the reasoning and the memory that AI agents provide to answer complex questions.”
The pre-built connectors for Agentspace enable access to more than 100 common enterprise applications including Microsoft Sharepoint, Google Drive, ServiceNow, Confluence, Box, Dropbox, SAP, Oracle and many more. Agentspace uses Google Cloud's secure by design principles and approaches; it also provides granular IT controls, including role-based access control (RBAC), VPC service controls, and identity and access management (IAM) integration so that data will remain protected and compliant. (See here for Google’s Secure by Design report.)
AI Agents to Do Your Bidding
Agentspace comes with pre-built “AI agents that offer intelligence, search and advanced research capabilities, and companies will be able to launch their own customized agents,” Pai said. “For example, Agentspace can have an HR agent that helps new employees onboard faster, or a marketing agent that generates targeted content.”
These custom AI agents can be built directly in Google Agentspace, which is part of the Cloud AI suite of products, and in Agentspace itself users can also activate AI agents that have been built using Vertex AI Agent Builder.
During the prebriefing, Kalyan Pamarthy, group product manager for Google Agentspace at Google Cloud, demoed the creation of a press release announcing the launch of a new water bottle. First, Pamarthy asked Agentspace to find examples of other press releases created by the fictious company. One of the results was a ‘notebook’ created in NotebookLM.
NotebookLM allows users to ‘ground’ a language model (the Gemini models in this example) in the sources the user wants to use (this is the same process, conceptually, as what Agentspace does with enterprise data sources). The user uploads those sources and/or links to existing enterprise data sources. The user can then ask ‘natural language’ questions about that data, thus leveraging the capabilities of the underlying large language model (LLM) to generate responses, summaries, etc., about only the data sources it can explicitly access. So, in the demo, Agentspace generated a press release specific to the fictitious company.
“Creating a press release is not a simple single shot search or prompt. If you tried that on Gemini, for example, you would get a press release, but it probably wouldn't be very relevant to what you're doing,” Pamarthy said. “Behind the scenes we’re calling an advanced research agent that is developed for breaking down a complex question or a complex prompt into multiple different sub prompts.”
As Pamarthy described it, this research AI agent planned, did the research, reasoned through the process to make sure it had what it needed (to fulfill the prompt), and then assembled the press release – which the human user could then edit as required. Agentspace is also integrated with multimodal models so it can generate images and videos, which Pamarthy also demoed. And according to Pai, “Agentspace also includes copyright indemnity for peace of mind for anything generated from the Gemini model.”
Google highlighted several potential use cases for Agentspace, including business analysts using the tool to uncover industry trends and create presentations from AI-generated insights, HR teams improving the employee experience by streamlining the onboarding process, software engineers proactively identifying and resolving bugs, and marketers optimize content recommendations, analyzing campaign performance and fine-tuning campaigns.
Currently, Agentspace supports Google’s latest models (the Gemini 2.0 models were announced this week). Pai indicated that in the first quarter of 2025 they intend to open Agentspace up to “some of the other common models.” Support for customers’ own fine-tuned models is also on the roadmap. Agentspace is in early access.
