Matillion's field engineering team argued that agentic data management pays off most for the analysts, data scientists and leaders stuck in the data team's queue. A live demo of Maia showed self-service pipeline building, permission prompts, context files that enforce standards and AI-driven troubleshooting.
Key takeaways
- The bottleneck is downstream. Analysts and business leaders often wait days or months on overloaded data teams for simple reports.
- Analysts know the business rules. Agentic tools let them build pipelines without writing Python, the presenters said.
- Inspectable output matters. Lipowitz argued that code a business user can't read forces them to simply trust the model.
- Context files govern agents. Naming rules and connector preferences keep non-engineers' work within company standards.
- Agents can triage failures. AI diagnosis of failed pipelines takes routine troubleshooting off data engineers' plates.
Most talk about automating data work centers on data engineers. In a Data Science Connect webinar, "What Agentic Data Management Unlocks for Data Scientists, Analysts, and Leaders," Matillion's field team argued the bigger payoff is downstream, for the analysts, data scientists and business leaders who wait in the data team's queue.
The session, sponsored by Matillion, featured Liam Morrison, VP of Field Engineering; David Lipowitz, Senior Director of Field Engineering; and Diana Pilzer, Senior Technical Account Manager. Brian Mink, co-founder and president of Data Science Connect, hosted. Lipowitz ran a live demo of Maia, Matillion's agentic AI, which Morrison described as a team of virtual data engineers launched in June.
Why are data teams a bottleneck for the business?
Morrison said he spoke as "the not data engineer." In earlier roles, getting insight meant opening tickets with a data team that everyone else was also waiting on. Data engineering talent is short, data sources are multiplying, and the goal is to drive revenue growth without tripling the size of data teams, he said.
He described trying to define when a customer on a SaaS platform was successful. A small data science team took two or three days per report, mostly because of other requests. Then he realized he had forgotten holiday weekends, which made healthy customers look unhealthy, and the wait started again. Reaching the report he wanted "actually took the better part of 90 days," he said. With self-service, "I believe I could have saved 89 days of effort."
Pilzer, a former data engineer and Matillion customer, described the other side. In daily standups, she said, "all the time the status was blocked and the blocker was me and that feels awful."
How does agentic data management help analysts and data scientists?
Lipowitz said there are two ways to widen the bottleneck: make data engineers more productive, and let analysts and data scientists self-serve their own data engineering. "Data analysts, in my experience, are the ones who often understand the business rules the best," he said, but they are not always Python coders or pipeline builders.
In the demo, using synthetic data for a fictional large bank that Maia itself generated, Lipowitz asked Maia to suggest analyses based on a context file describing the available data. It proposed inflow forecasting, outflow prediction, cash flow optimization and customer segmentation. He then asked for a pipeline to build a data mart showing interest opportunity cost.
Maia drew up a to-do list and began creating a staging table in Snowflake. Partway through, it found the data was already loaded in three existing tables, changed course and asked permission to delete the file it had created. It also asked before sampling data. Lipowitz described these permission prompts as a security control. Data scientists, he said, can spend less time on data prep and more on modeling, while data engineers can run multiple Maia sessions in parallel.
Why not just use a coding assistant?
Morrison raised the question himself. Lipowitz said he enjoys vibe coding on weekends, but a business analyst who can't read the generated code is "sort of just trusting what the large language model is generating."
Maia, he explained, writes pipelines in a human-readable YAML format that Matillion's Data Productivity Cloud renders as a visual pipeline built from a fixed palette of components. That narrows the problem space and makes each step inspectable, with data sampling available at any point. Everything is backed by Git and fits CI/CD patterns, and the platform supplies logging, failover, role-based access control, audit logs and lineage. He said Maia is underpinned by Claude Sonnet 4.5, among other models.
Lipowitz said the platform runs several agents: a data analyst agent that profiles data, a data engineer agent that builds pipelines, a data ops agent that promotes code and writes commit messages, and an operations agent for root cause analysis.
How do you govern agents when non-engineers build pipelines?
Through context files, Lipowitz said. In the demo, files set naming rules, such as a BNK prefix on every table and an STG suffix on staging tables, and an order of preference for database connectors. Even when an analyst does the work, Maia follows the company's standards.
Pilzer said the same files help onboard new data engineers, compared with a standards document in a shared drive that no one may read again. Maia can also generate context files from existing material, such as a PDF, a Word document or a legacy Alteryx workflow, the presenters said.
Can AI agents troubleshoot failed pipelines?
Lipowitz filtered the activity log for failed runs and showed Maia diagnosing one: a Salesforce incremental load that staged no rows, with a detailed write-up on how to fix it. That spares data engineers routine questions and lets analysts and data scientists get past errors themselves, he said.
Morrison called this kind of work "the beginning of true agentic work," built on "trust but verify." Earlier he said one customer cut pipeline build time from about 10 hours to one hour and migrated 3,000 legacy pipelines.
Pilzer tied it back to trust. When people can see what's happening, or talk to Maia directly, they start trusting the data. "You have to trust it to be able to do anything with it," she said.
More demos on the 2027 calendar
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: AI Demo Days, Agentic AI Showcases, the AI × Enterprise Panel Series and the Plug & Play Series. See the webinar schedule for upcoming sessions.
Frequently asked questions
What is agentic data management?
As Matillion presented it, AI agents take on repetitive data tasks such as building, documenting and troubleshooting pipelines, so data engineers handle harder work and analysts and data scientists can self-serve.
How can data analysts build data pipelines without coding?
In the demo, an analyst describes the goal in natural language and the Maia agent builds a pipeline in a human-readable YAML format rendered as a visual pipeline, asking permission before key actions.
How do you keep AI-built pipelines within company standards?
Matillion's David Lipowitz showed context files that set rules such as table prefixes, staging table suffixes and preferred connectors, which the agent follows regardless of who is using it.
Why not use a general coding assistant for data engineering?
Lipowitz said business users can't verify generated code they can't read, and that production data work also needs governance, Git integration, logging, failover and scalability.
Watch: What Agentic Data Management Unlocks for Data Scientists, Analysts, and Leaders on YouTube
Event: Webinars · More coverage: Webinars insights
