In a hands-on lab at COLLIDE Data + AI Conference 2026, Matillion's Arawan Gajajiva showed how the Maia agent builds, tests and documents data pipelines from plain-language requests. He argued data engineers' work is shifting from writing code to governing, planning and evaluating what AI builds.
Key takeaways
- Data engineering is becoming collaboration. Gajajiva said teams shift from writing code to governing, planning and evaluating an AI agent's work.
- Plan mode reduces rework. Having the agent present a plan before building is the safer choice for production pipelines.
- Permissions keep a human in the loop. Maia asks before running pipelines, committing code or deleting files.
- Visual pipelines help people check AI. A low-code canvas lets engineers validate generated work without reading a wall of code.
- LLMs can run inside pipelines. The lab used Snowflake Cortex to classify free-text transaction notes and add sentiment in minutes.
Data teams are buried in backlogs and production maintenance, and agentic AI is changing how they work, Matillion's Arawan Gajajiva told attendees at COLLIDE Data + AI Conference 2026 in Atlanta. In a 45-minute hands-on lab, "Agentic Data Engineering in 45 Minutes: Maia from Plan to Pipeline," Gajajiva, Principal Architect, Sales Engineering COE at Maia by Matillion, walked participants through building data pipelines by describing what they wanted in plain language.
"I've yet to talk to a data team that has told me that they don't have a backlog," he said. Most teams, he added, spend the majority of their time maintaining what they have already released to production.
What is agentic data engineering?
Gajajiva described it as a shift from hands-on-keyboard work, whether writing SQL and Python or dragging components onto a canvas, toward collaboration between people and AI agents. The data team's role becomes governing, planning and evaluating the AI's work.
The goal is to "work alongside AI, not replace, but to help scale that data team," he said. In his framing, the agent does the heavy lifting while people supply the critical thinking.
He argued that coding assistants mostly address one stage of the data product life cycle, the build. Releasing a data product also involves intake with stakeholders, design, testing, deployment, documentation and operations. Matillion positions Maia as an AI data automation platform meant to support that entire life cycle, according to Gajajiva.
How does an AI agent build a data pipeline?
In the lab, each participant typed a request into a chat panel. Gajajiva's first prompt asked Maia to join a budgets table with departments, aggregate budgeted and actual amounts, calculate variance and write the results to a new dashboard table. Maia built the pipeline, sampled the data to check it, asked permission to run it, and then summarized the results.
The agent works in two modes. In act mode it starts building immediately. In plan mode it first presents a plan for review, which Gajajiva recommended for real production pipelines because it reduces iterations. "Even the most expert data engineer is not going to get everything right the first time," he said.
Permission settings add a human in the loop. Maia can ask before every action, or run in a balanced mode that Gajajiva said he uses most, where it can sample data and query the warehouse freely but still asks before running pipelines, committing code or deleting files.
Why keep a visual interface when AI writes the code?
"In an AI world, high and low code really doesn't matter," Gajajiva said, since writing SQL or Python is not hard for AI. Matillion keeps its low-code visual canvas anyway. "This is for the humans that are governing the AI," he said.
Instead of reading a wall of generated code, an engineer can look at each component, sample the data at any step and confirm the pipeline does what was intended. Engineers can also take over and configure components themselves. Gajajiva compared it to a car's self-driving mode: sometimes you want to drive.
How can LLMs be used inside data pipelines?
The second exercise put an LLM inside the pipeline itself. Participants used a Snowflake Cortex component to classify free-text approval notes on financial transactions into categories such as compliance risk, cost optimization and vendor issue. Gajajiva estimated that kind of task might take a day or two in many tools.
"AI building AI pipelines," he said of the exercise. Maia picked Mistral Large 2 for the task, but users can choose other models from a drop-down. Gajajiva said the platform also has components for Google Vertex, Amazon Bedrock and OpenAI, an approach he called "bring your own AI." He then added a sentiment field to the same step during the demo.
The lab ran on Snowflake, and Gajajiva said the platform also supports Databricks, BigQuery and Redshift.
What else can AI agents automate in data engineering?
Gajajiva highlighted several other capabilities:
- Ingestion: connectors for a couple hundred source systems, from on-premises databases such as Oracle and SQL Server to SaaS platforms. For a REST API without a connector, Maia can read the documentation and build one, he said.
- Synthetic data: generating test data for outlier conditions that are hard to reproduce before a pipeline goes to production.
- Skills and rules: codifying patterns such as medallion architecture, slowly changing dimensions or star schemas. Maia can also detect patterns in pipelines a team has built and turn them into skills, so future work follows them.
- Documentation: Maia added notes to pipelines as it built them.
See more agentic AI sessions 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, panels and hands-on sessions like this one. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Between conferences, Data Science Connect's free Agentic AI Showcases and AI Demo Days show agentic tools in action online.
Frequently asked questions
What is agentic data engineering?
As Gajajiva described it, an approach in which AI agents do the heavy lifting of building, testing and documenting data pipelines while data teams govern, plan and evaluate the work.
What is Matillion Maia?
Maia is Matillion's AI data automation platform, with an agent that builds data pipelines from natural-language requests and asks permission before key actions.
Which data platforms does Maia support?
Gajajiva said the platform supports Snowflake, Databricks, BigQuery and Redshift, with components for LLM providers such as Google Vertex, Amazon Bedrock and OpenAI.
Can AI agents build data connectors?
According to Gajajiva, if no connector exists for a REST API, Maia can read the API documentation and build one.
Watch: Agentic Data Engineering in 45 Minutes: Maia from Plan to Pipeline on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
