The more boring AI gets, the more it's actually working

While my social feeds this week are full of yet more AI hype, my actual week in AI has been kinda boring.

🥱 Tuning a document ingestion pipeline to correctly interpret unrealised capital gains across different investment platforms.

😴 Adjusting weightings in a tax classification model to better handle edge case transactions.

💤 Refining a conversational agent to understand the hierarchy of financial data when modelling retirement scenarios.

See? Pretty dull. I’m nearly nodding off just writing about it.

It’s detailed work that needs tight feedback loops, explicit acceptance criteria, and a lot of boundary condition testing. And none of it involves whatever model or tool was released yesterday.

In fact, for these apps, I avoid change rather than chase it. Once an AI process is working reliably in a high-trust workflow, stability trumps novelty.

It becomes mundane.

But I’m enjoying this work a lot because it’s where AI starts being seriously useful. Just another component in the architecture, alongside compute, storage, networking. Something you depend on without really thinking about it.

And this is still so recent. A few years ago, these architectures didn’t include LLMs at all. Now they’re just a part of the foundation.

So while my feed screams at me about what’s new, I’m finding that the more boring AI gets, the more it’s actually working. 🙂