What to Check When Asked 'Can AI Do This?'
A note on evaluating whether a task can be handled by AI based on input variability, failure impact, and human review cost rather than model performance alone.
I get asked “Can we do this with AI?” more often these days.
Technically, AI can handle quite a lot: reading documents, classifying content, writing summaries, generating candidates, suggesting next steps. Prototyping any of these is easy.
But fitting it into a real workflow is a different story. The first thing to look at should not just be how capable the model is.
How much does the input vary?
What is the impact when it makes a mistake?
Will human review time still be needed?
Can the reasoning be explained later?
Can we revert to the original process on failure?From this perspective, you can tell which tasks suit AI and which ones are still better handled by humans. For example, rough classification of inquiries might be a good fit, but automatically deciding whether to approve a refund probably deserves more caution.
In AI adoption discussions, asking “can we still operate when it fails?” gets you further than asking “can it do this?” Small experiments are fine, but if you’re putting it into a production workflow, you need to design review, rejection, logging, and scope of responsibility together.
