At COLLIDE 2025, Google Cloud's Suda Srinivasan ran a live demo of a data science agent planning and building a demand forecast in a BigQuery notebook, including fixing its own error. Srinivasan described agents as an exoskeleton for data scientists, not a replacement.
Key takeaways
- The role has expanded. Data scientists now handle RAG, embeddings, multimodal data and content generation alongside classic ML.
- Plumbing eats time. Srinivasan said moving data between separate SQL, Python and BI tools takes time away from data science.
- Agents plan, ask and fix. In the demo, the agent proposed a multistep forecasting plan, asked for approval and corrected its own library error.
- Four good uses: automating the mundane, scaling the possible, eliminating the cold start and moving faster.
- An exoskeleton, not a replacement. Srinivasan said agents make data scientists faster rather than replacing them.
Data scientists are being asked to do far more than they were three or four years ago, from building RAG applications to analyzing video, and AI agents can take over much of the routine work, according to Suda Srinivasan, Group Product Manager, Data Analytics at Google Cloud. At COLLIDE Data + AI Conference 2025 in Atlanta, in a session sponsored by Google Cloud, Srinivasan ran a live demo of an agent building a forecasting model inside a notebook and explained what agents are, and aren't, good for.
How is AI changing the data scientist's job?
Classic data science meant forecasting, classification and regression, Srinivasan said. Now the role also includes agents, semantic search, vector embeddings, multimodal data and content generation. Citing Gartner, Srinivasan said data scientists are becoming AI engineers and data analysts are becoming data scientists.
Srinivasan described three ways AI is expanding what data teams can do:
- Multimodal data at scale. Documents, images and videos that companies stored for years can now be analyzed inline, such as scoring sentiment on more than a million customer reviews from one to five, or extracting damage analysis from videos in cloud storage.
- Agents and assistants that take routine work off data scientists and data engineers.
- Blending real-world knowledge with enterprise data. A single query can combine structured data, such as schools in certain zip codes, with what Gemini knows, such as which schools are highly rated.
AI also amplifies the problems already in your data, Srinivasan warned. "Governance has to be front and center."
Why do data scientists lose time to tool sprawl?
Writing Python in one tool, SQL in another and dashboards in a third forces data scientists to move data between systems, a process Srinivasan called slow and error-prone. "A lot of the time that data scientists spend is spent in plumbing and stitching rather than on data science," Srinivasan said.
Google's answer is an AI-first Colab notebook in BigQuery that mixes SQL, Python and Spark in one place, with built-in visualization and AI help for code generation, completion and error fixing. Teams that prefer Jupyter or VS Code can connect those tools to Google's data services instead, according to Srinivasan.
Google has also built what Srinivasan called a family of first-party data agents. The data science agent turns a natural-language request into a multistep plan and carries it out, while the data engineering agent builds data pipelines from prompts and troubleshoots them when they break. Srinivasan said the data science agent can even field a question like whether a root mean square error looks acceptable, explaining why it might be high and how to fix it.
What did the live agentic AI demo show?
The demo started with a BigQuery public dataset of Austin bike-share trips. Srinivasan asked the agent to pull 10 rows, and it located the right table and generated the query. The next prompt asked it to switch to BigQuery DataFrames, which Srinivasan described as data frames that run on BigQuery's scale behind the scenes, and to build a time-series model forecasting daily demand by station for a specific subscriber type, so marketing could focus on the busiest stations.
The agent laid out a plan: load the data, aggregate daily demand by station, train the model, forecast, rank stations and produce a report. It asked for approval at each step, and Srinivasan said users can edit the plan at any point.
The demo didn't go perfectly. The agent hit an error caused by an outdated library version, then diagnosed and fixed it without being asked. "Some of this is not deterministic," Srinivasan said, noting that each run can vary slightly. When model training ran long, Srinivasan switched to a prepared tab to continue.
What are AI agents good for in data science?
Srinivasan listed four uses:
- Automating the mundane: installing libraries, cleaning data and routine transformations
- Scaling the possible: for example, trying several models in parallel and reporting which fits best, work Srinivasan said would otherwise take about half a day
- Eliminating the cold start: giving data scientists starting code to edit instead of a blank page
- Moving faster: troubleshooting, debugging and code completion
What agents won't do, according to Srinivasan, is replace data scientists. "Think about it like an exoskeleton, like a suit that you can wear that lets you get faster and stronger," Srinivasan said.
How do you bring AI models to data in BigQuery?
Srinivasan described three paths that avoid moving data out of the warehouse: building and training models in SQL with BigQuery ML, scaling training with serverless Spark from the notebook, and managing custom models in Vertex AI, then using them in BigQuery.
For forecasting without training, Srinivasan pointed to TimesFM, a pretrained time-series foundation model from Google DeepMind. Vertex AI also offers third-party models, including Claude, Mistral and some OpenAI models, plus a connection to Hugging Face, Srinivasan said.
Go deeper at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for keynotes, technical sessions and live demos. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center; see the COLLIDE event page. Between conferences, Data Science Connect's free webinars, including AI Demo Days and Agentic AI Showcases, feature live product walkthroughs.
Frequently asked questions
What is a data science agent?
An agent that turns a natural-language request into a multistep data science plan, such as loading, aggregating, training and forecasting, and runs it with user approval at each step.
Will AI agents replace data scientists?
Srinivasan said no, comparing agents to an exoskeleton that helps data scientists work faster and take on more.
What are AI agents good for in data science?
According to Srinivasan: routine tasks, trying many models in parallel, giving starter code instead of a blank page, and speeding up debugging and coding.
Can you forecast time series without training a model?
Srinivasan pointed to TimesFM, a pretrained time-series foundation model from Google DeepMind available for inference with BigQuery data.
Watch: Supercharge Your Next Project With Agentic AI, brought to you by Google Cloud on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
