Before You Build an AI Agent, Ask What's Actually Broken

One of the problems we have with AI use right now is that, instead of asking what is broken in our domain, we keep asking ourselves how to fit an AI agent into our workflows. Is that actually a smart thing to do, and a learning process about something that will be helpful in the future, or is it just chasing hype? After all, it's probably a fair estimate that 70% of what we do during our workdays is deterministic, meaning it doesn't require AI or an agent and can be solved through a simple automation. My own path through this has had two stages so far.

• The first stage was pure inspiration at how fast AI works and what it could do for me.
• The second arrived when the apps and automations I'd built bounced back full of bugs, and fixing them ate so much time that I wondered whether I was doing anything at all, or just playing with bots that tell me I'm correct about everything.
• That stage passed too. Now I want to ask first what problem I'm solving, and only then decide where my experience with AI fits into the picture.
• The important thing to realise is that using AI is a work skill, like design, research or product management; it takes time to learn, and the more you learn it, the better your view becomes of what it can actually get you.
• Coming back to the 70% of work we can probably automate and forget about: automating that share might already be a good use case for learning to build those automations with AI.

And then there is the remaining 30%. Why is that part so hard to automate? Is it because the inputs are too messy, or rather because the outputs cannot be achieved without judgement calls somewhere along the way?

Yours, in plain text,
Zhenia