Note: AI Engineer in 2026
A check-in on the definition of an AI Engineer in 2026 - from fine-tuning to agentic systems, tools, RAG, and MCP.
I was just pulled into a Slack thread talking about the definition of an AI Engineer. And as my thumbs flew in response I realised that it was time for a check-in on the subject. To that end, here is verbatim what I wrote in that thread adding in links for context:
Ref: The definition of the AI Engineer in 2026
Let me throw in my two cents here. The contemporary definition of AI Engineer used by most people in the industry started with the Rise of the AI Engineer blogpost. That went on to be the genesis of the AI Engineer World’s Fair etc.
Back when the term was coined the world was a different place. So yes, AI Engineers worked with API hosted models and also did all the fine tuning etc and working with synthetic data emerged as a thing, model distillation etc etc. This ‘beyond the use of APIs’ was required as back then the models where too expensive and not aligned with various industry verticals and use cases.
The world is in a different place now. While it used to be that someone like Bloomberg would fine tune a model to work with financial data (see BloombergGPT), I would hazard a guess that they don’t use that anymore. Models are big enough and powerful enough and cost effective enough that they can work with agentic systems to do pretty much whatever we need. The cost of developing tools and data sources is far far cheaper and less risky than fine tuning a model.
Fine tuning is still a thing, but it’s for the Cursors of the world who are training very large models with the bet of carrying their whole business.
So an AI Engineer today… someone who uses models via API, understands agentic architectures, creates tools, sets up RAG, knows MCP, and bottom line… manipulates strings.
Editor’s note: I’m the AI that helped add the links to this post (Claude). On the Bloomberg point – Mike’s instinct is well-supported. Research showed that GPT-4, with zero financial training, outperformed BloombergGPT on nearly all financial benchmarks. And in late 2025, Bloomberg integrated Anthropic’s MCP protocol into the Bloomberg Terminal – building on general-purpose models rather than relying solely on their custom-trained one. The broader trend holds: the frontier moved fast enough that domain-specific pre-training lost its edge, and the AI Engineer’s toolkit shifted from training models to orchestrating them.