In this article 4 sections
On 30 October I was in Bioggio, Switzerland, for the Ticino Data Conference 2026, a community conference organised by Power BI User Group Italy at the Acer EMEA headquarters.
My session was called How to Survive AI-Assisted Code as a Data Engineer, with the subtitle Choosing the Tool, Giving it Context, and Checking What it Wrote.
AI coding assistants generate code fast. Fast and correct are two different things, though. As a data engineer, a hallucinated function or a subtly wrong transformation can break a pipeline in ways that take hours to diagnose. So the question I wanted to help people answer was this: when is the AI helping you, and when is it hurting you?
The session was for data engineers who already use tools like Copilot, Claude, or ChatGPT and want to move beyond trial and error. It followed the three parts of the subtitle.
Choosing the tool
We started with the setup of the agent coding harness. What is each major AI assistant actually good at, where does it fall short, and how do you match it to the task in front of you? Writing Spark transformations, building SQL logic, and generating boilerplate are different jobs, and the tool should fit the job.
If you want to see how two of them compare on a real task, I put ChatGPT and Claude head to head on a Fabric notebook.
Giving it context
Next up was the context we give to the agent. I went through concrete patterns for structuring prompts, including how to provide context, how to constrain the output, and how to iterate. Do this well and you get production-grade code, instead of “close enough” output you have to rewrite anyway.
I wrote more about Claude skills for Microsoft Fabric, one way to give an agent reusable context.
Checking what it wrote
Even with the right tool and good context, the code still needs checking. So the last part was validation, with a practical checklist for verifying that AI-generated code covers edge cases, handles nulls correctly, and doesn’t silently alter your data logic. The checklist is repeatable, so you can run through it every time to catch hallucinations, edge cases, and silent data logic errors.
If you want to automate part of that, I wrote about letting Claude Code test its own work, at your own risk.
How it went
The talk was very well received. I had a big room, which I liked, and there was some nice interaction going on with the audience. I did run out of time a little, because there was just too much nice content to share in the time available, I guess.