When you build an AI application, I think there are two questions you have to answer up front:
1. Why am I using AI for this at all? In other words, what is the value add?
2. What am I optimizing for?
With very few exceptions, most AI apps I see don’t seriously answer either question.
What I find especially strange is the obsession with making LLMs behave like conventional computing systems. If your goal is determinism, rigid workflows, and fully codified logic, you already have workable tools for that.
They’re cheaper, faster, and easier to reason about.
AI gives you something else entirely.
Agents shine in dirty systems. Noisy data. Incomplete inputs with ambiguous goals. Situations where the process itself has to evolve because reality refuses to cooperate.
One question people ask all the time is, "Why can't we use ai to automate the boring stuff?" We can! Developers just haven't caught up yet.
Thing is, when used correctly (and yes, there is a correct methodology), agents should be able to fill in gaps, reason under uncertainty, and adapt the process as conditions change.
That’s where the economics start to get interesting.
The more adaptive your system is, and the better prepared your agents are for their actual environment, the fewer edge cases you have to hard-code.
You'll have fewer brittle rules, less scenario explosion.
Less human babysitting.
That’s the real cost savings: automating things that couldn't possibly have been automated in the pre-AI world.
The current wave feels constrained by old mental models that are no longer correct about computing.
Some of it is a reasonable fear of losing control.
Some of it is simple resource misallocation.
Point is, if all you're doing with AI is trying to shoehorn an agent into a conventional application model, you're missing the point... and leaving money on the table.
It's time to get creative.