Good AI ideas usually stall waiting on other teams, separate systems and security reviews, Google Cloud's Shan Kulandaivel said in a COLLIDE Data + AI Conference 2026 keynote. He showed how forecasting, prediction, document processing, graph and vector search inside BigQuery let SQL users take ideas to production.
Key takeaways
- Ideas aren't the bottleneck. Kulandaivel said skills gaps, separate systems, governance and process slow AI to production.
- Forecasts in one line of SQL. AI.FORECAST uses the pre-trained TimesFM model, so no model training is needed.
- Documents stay in the governed platform. Gemini models parse PDFs inside BigQuery and join results with existing tables.
- Token costs can be managed. An optimized mode scores a subset with an LLM and trains a classifier for the rest.
- Context lives with the data. Native graph and vector search avoid copying data to purpose-built databases.
Most enterprises don't lack good AI ideas. They lack a fast path from idea to production, Google Cloud's Shan Kulandaivel told the audience in a keynote at COLLIDE Data + AI Conference 2026 in Atlanta. "Ideas have never been the problem," said Kulandaivel, Group Product Manager at Google Cloud, who works on BigQuery.
In "From Raw Data to Running Predictions: Accelerate Your ML Workflow from Weeks to Minutes," he argued that the obstacles sit between ideas, data and production: separate systems to stand up, skills gaps, governance gaps and process gaps. His answer was to bring AI models directly into the data platform.
Why do enterprise AI ideas stall before production?
Kulandaivel told the story of three employees at Harbor and Hearth, a fictional retailer.
Maya in demand planning had a proof of concept for AI forecasting ahead of the holidays. Her request landed at number 47 in the upstream team's queue, and the forecast arrived two weeks late, after buyers had placed orders using the manual method they had relied on for 10 years.
Dave in loyalty and marketing wanted a churn model to feed a retention campaign. The ML team liked it but gave the familiar answer: "Let's revisit next quarter because we are busy with deliverables this quarter."
Rosa in accounts payable processes 40,000 PDF invoices against purchase orders. Her pilot showed processing time could drop from nine hours to nine minutes, but it required copying data out of the governed platform, which meant a security review that could take a quarter or two, and she had no idea what production would cost.
The common thread, Kulandaivel said, was waiting on other teams and systems. The bigger loss is the compounding effect of every idea that never ships.
How can SQL users build AI forecasts and predictions?
Google Cloud's approach is to bring Google's and partners' AI models into BigQuery so analysts can work with the SQL skills they already have, inside the governed platform, according to Kulandaivel.
In his telling, Maya forecasts every product for every store with one line of SQL using the AI.FORECAST function, backed by TimesFM, a pre-trained, zero-shot foundation model from Google Research that requires no training. Connected Sheets then puts those forecasts in front of buyers who work in spreadsheets, so they don't have to change their workflow. "Maya's 6 weeks became 6 minutes," he said.
"Maya might be fictional, but her problems are not," Kulandaivel said. He cited a Croatian retailer that used BigQuery forecasting to reduce produce wastage by 7% and increase operating profit by 8%, which he said can be the difference between a loss-making and a profitable quarter in a thin-margin business.
Dave's churn model uses AI.PREDICT, which Kulandaivel said recently launched in preview and uses a zero-shot foundation model for tabular data built by Google Research and BigQuery. He said Duolingo used BigQuery AI to improve its ad placement engine, with revenue impact in the tens of millions of dollars, and has since expanded to customer retention and churn prediction.
How does BigQuery process unstructured documents?
For Rosa's invoices, Kulandaivel described a single SQL statement that reads PDF files in Google Cloud Storage through a managed, governed capability called ObjectRef, uses a Gemini model to extract structured data, and joins the results with tables such as purchase orders.
For longer documents, such as a 200-page supplier contract, he said Google is adding parse document and chunk document functions in SQL that store document chunks as table rows for analytics, generative AI and combined lexical and semantic search.
Unstructured data matters more as companies adopt agentic AI, he said, because structured warehouses capture only part of an enterprise's knowledge. He cited a Gartner estimate that 80% of a typical enterprise's data is unstructured.
What about LLM token costs at scale?
Leaders worry about running LLMs across millions or billions of rows, Kulandaivel acknowledged. He described an optimized mode that uses an LLM to score a subset of data, then trains a classification model on the fly for the rest. He said the technique delivers up to 230 times fewer tokens on classification tasks.
He also outlined BigQuery's wider set of functions, including generation, classification and scoring with generative models, text, vector and hybrid search, and support for ML engineers training custom models without moving data. Letting data users handle routine AI work frees ML teams for harder problems, he said.
Why bring graph and vector search into the data platform?
"AI is great, but sometimes AI needs more context," Kulandaivel said. Graph analytics and vector databases used to require copying data into purpose-built systems. With native graph capabilities, he said, Rosa could map suppliers, bank accounts and transactions to trace payments and spot fraud, and some large financial services customers use BigQuery graph for fraud detection. Native vector and hybrid search support semantic search for agentic AI use cases.
What stood between his three characters and production, he concluded, was not people but skills, processes and compliance requirements.
See more enterprise AI keynotes 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. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Between conferences, Data Science Connect's free AI Demo Days show data and AI platforms in action.
Frequently asked questions
What is BigQuery AI.FORECAST?
A SQL function that, according to Kulandaivel, produces forecasts using TimesFM, a pre-trained zero-shot foundation model from Google Research, without any model training.
Can BigQuery process PDF documents with AI?
Yes. Kulandaivel described a single SQL statement that reads PDFs from Google Cloud Storage via ObjectRef and uses a Gemini model to extract structured data.
How can teams control LLM token costs on large tables?
Kulandaivel described BigQuery's optimized mode, which uses an LLM on a subset of rows and trains a classifier on the fly, cutting tokens by up to 230 times.
Why do AI pilots get stuck before production?
In Kulandaivel's examples, teams waited in upstream queues, lacked ML skills, or needed security reviews because pilots copied data out of governed platforms.
Watch: Keynote: From Raw Data to Running Predictions — Accelerate Your ML Workflow from Weeks to Minutes on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
