Google publicly releases 'Deep Research'... The competition to integrate enterprise internal data with AI agents officially kicks off

Google has rolled out two new artificial intelligence (AI) agents that can automatically generate survey reports based on topics specified by users. They can not only conduct simple searches, but also perform analysis by combining data from the public internet and even enterprise internal systems, which is expected to further intensify competition in enterprise AI.

On the 22nd local time, Google released “Deep Research” and “Deep Research Max.” These two products are follow-up models to the existing AI research tools launched last December. At the time, the tool was developed based on “Gemini 3 Pro,” while the new products use the more advanced large language model (LLM) “Gemini 3.1 Pro,” released in February this year.

The performance improvements are also substantial. Google said that, according to OpenAI’s benchmark test “BrowseComp,” which compares the two generations of models, Gemini 3.1 Pro scored 85.9. This is more than 25 points higher than the existing Gemini 3 Pro. The benchmark evaluates the LLM’s online research capability across more than 1,000 tasks.

Data Access Scope and Use Cases

The standout feature of the new AI agents is their data access scope. “Deep Research” and “Deep Research Max” can access not only the public internet, but also call data from enterprise internal systems. When connecting to internal systems, they use “MCP” (Model Context Protocol), and users can also directly upload spreadsheet or video files to supplement the analysis data set.

Google has proposed healthcare and finance as application examples. For instance, researchers can quickly generate reports on new compounds that show therapeutic potential, and financial experts can hand over research work for companies they are considering investing in to AI. This means it can significantly reduce the time needed for the information collection and organization stages.

These agents also provide functions to visualize the collected data. Visualization can be implemented in the form of HTML code, or by using Google’s image generator “Nano Banana.” According to Google, Nano Banana includes a built-in general knowledge database that can relatively accurately interpret the provided information and present it in image form.

Working Method and Product Differences

The working method is also designed so users can adjust it in advance. Before generating a report, the AI will first propose an outline of how to conduct the research. Users can modify this plan to improve the quality of the final output. For example, researchers can specify a particular scientific database as the priority target for search.

The two products are positioned differently. “Deep Research” is designed to run using relatively fewer computing resources. Google explained that, compared with the December version, this model is cheaper, faster, and improves result quality. This means it is suited to application scenarios that require quick responses.

By contrast, “Deep Research Max” focuses on “maximum comprehensiveness.” Its structure involves investing more time and hardware resources to generate more in-depth reports. This has been interpreted as a product tailored for work tasks that value completion and the breadth of research more than speed.

Significance and Future Plans

In a blog post, Lucas Hasse and Srinivas Tadepalli of Google DeepMind wrote: “‘Deep Research’ reports in and of themselves are valuable, but they can also serve as the first step in complex agent workflows that begin with deep-context data collection.” This indicates that AI is moving beyond simple question-and-answer, becoming the “starting point” for actual work processes.

At present, “Deep Research” and “Deep Research Max” are available through the Gemini API in a public preview format. Future plans include expanding to Google Cloud. Google also said it plans to add MCP interworking capabilities to make it easier to access data sources such as FactSet and PitchBook.

This release shows that competition in generative AI is rapidly shifting from “interactive chatbots” to “practical agents.” In particular, Google’s attempt to integrate research, analysis, visualization, and internal data connectivity into one system could have a significant impact on the enterprise AI market.

TP AI Notes: This article uses a language model based on TokenPost.ai to generate a summary. The main content may be omitted, or may differ from the facts.