Maia is Matillion’s agentic AI assistant for data engineering. It helps teams build, inspect and debug data pipelines using plain language, while keeping outputs constrained, visible and production-safe.
In this session, Ethan Cohen, UK Sales Director at Remobi, spoke with Julian Wiffen, Chief of AI and Data Science at Matillion, about how Maia works in practice, what sits under the hood, and what enterprise teams can learn from building and shipping agentic AI in real environments.
If you work in enterprise data, AI governance or platform leadership, this discussion explores both the technical architecture and the organisational impact of AI assistants in the data stack.
Why AI Data Engineering Matters for Enterprise Teams
AI in the data stack is now being evaluated on delivery, not novelty.
Enterprise teams face more data sources, greater demand from analytics and GenAI initiatives, and rising expectations around speed and governance.
The question is no longer “Can AI help?”
It is “Can it deliver safely, visibly and at scale?”
Maia is positioned as one answer to that challenge.
What is Maia and How it Changes Day-to-Day Work
Maia is designed as a practical assistant for building and maintaining data pipelines. It’s core promise is to reduce time spent translating requirements into working pipelines, and make it easier to inspect what has been built before anything runs in production.
In the live demo, Julian built a pipeline using plain language (filtering, ranking, sampling), then introduced an intentional error and used Maia to diagnose and correct it. The workflow was built around visibility; the pipeline components were displayed, sampling was available, and changes could be committed.
5 Practical Takeaways from the Session
1. Relieving the data engineering bottleneck
Teams are dealing with more sources, more demand, and rising expectations driven by GenAI and machine learning. Maia is positioned as a way to reduce that pressure by enabling faster delivery and broader self-serve.
“Data engineering teams are already at a pinch point… [and] the world of Gen AI and machine scale is just going to make that worse… So what we’re offering is a way in which [GenAI] actually helps shift that load by making them way more productive.”
2. How Maia builds and validates pipelines
The session goes beyond a surface-level “AI assistant” description and explains how Maya is implemented. Maya runs on frontier models through Amazon Bedrock, using Anthropic Claude models, and relies on a loop where the model decides whether to call tools (inspect tables, sample data, validate steps) or complete the task.
A major design choice Matillion emphasised: constrain the model’s output so it configures deterministic components that can be validated and iterated.
As Julian explained:
“There’s a lot of power in constraining the output… asking [the LLM] to configure something that then goes back into a deterministic box you can check and push back into the cycle… Getting that feedback loop is super powerful.”
3. How teams should use Maia
Julian avoided positioning Maia as “hands-off autonomous.” Instead, he described a review model similar to working with a skilled colleague, where either the human or Maia can be the reviewer, but validation still matters.
“I wouldn’t regard it as cast iron… It’s very similar to the way you’d work with a colleague… you still want a second pair of eyes on it.”
He also linked trust to explainability. Because Maia is configuring a visible pipeline (and can sample outputs), it is easier to validate than long blocks of generated SQL.
“Because we’re a low code platform, it’s very easy to see what it’s done… to sample what it’s got and say, does this match what I was expecting?”
4. Raising the floor and moving teams “up the stack”
This part of the discussion is directly relevant to leaders hiring or reshaping teams. Julian’s point is not that engineering goes away; it shifts. Juniors start at a higher baseline, seniors take on more complex work, and leads operate at a larger scale.
“The technical lead is operating at just an unprecedented scale… the amount of new work that’s being created is huge.”
5. Lessons for building agentic products
Julian shared practical advice for teams building agentic systems:
- Build around feedback loops
- Keep outputs constrainable and testable
- Do not abandon projects if early iterations are inconsistent
“If it kind of half works now, don’t abandon it… park it, come back to it in three months… the rate at which the quality of the frontier models advances…”
He also shared a quality lever that many teams underestimate: documentation.
“Well-written documents also really drive it… good documentation for a human is also good documentation for an AI.”
Clear structure improves both human understanding and model performance.
Key Benefits of Maia for Enterprise Data Teams
- Faster translation from requirement to pipeline
- Built-in validation through visible components and sampling
- Reduced engineering bottlenecks
- Structured, constrained outputs rather than free-text code
- An agentic loop that improves through feedback
GenAI increases demand for data infrastructure. Tools like Maia aim to shift load by increasing productivity while maintaining governance and review.
If you want the full walkthrough (demo, architecture, adoption patterns, and lessons learned), watch the full recording below:
Continue the Conversation
Remobi supports organisations building AI and data teams across 27+ countries.
We bring together senior leaders through the Remobi AI Enterprise Forum to discuss practical AI adoption in enterprise data environments: governance, productivity, operating models, and capability design.
To find out about joining our next event, click here.
Or connect directly with Ethan Cohen to chat.
FAQ
What is agentic data engineering?
Agentic data engineering refers to AI assistants that can build, edit and debug data pipelines using tool-calling and feedback loops, rather than generating free-text code alone. These systems configure structured components that can be inspected and validated.
Does Maia use a custom model?
No. Julian confirmed that Maia uses frontier models (Claude via Amazon Bedrock), rather than a specially customised proprietary model.
Can teams trust Maia fully autonomously?
Julian compared Maia to working with a skilled colleague. It increases productivity, but outputs should still be reviewed and validated before production deployment.