Learning the Work Before Building the Agent
In my first weeks of enterprise AI work, much of my time has gone into operating materials and presentations. It is changing how I think about where an agent could actually help.
I recently joined a company building intelligent hardware and embodied AI systems to work on enterprise AI transformation. So far, a lot of my day-to-day work has involved operating materials and presentations. I have spent hours revising slides, then looked back at the day and struggled to say which decision became clearer.
It is tempting to see that as time taken away from the AI work. I have begun to see it as part of learning what the work actually is. A slide can be made quickly. The judgment behind it depends on where a figure came from, why an exception matters, what is still uncertain, and who can decide what happens next. Those details rarely fit into the request to "make a presentation."
Around the same time, someone at work asked me how to write a prompt. I could give a useful template for a task, context, and desired output. I also know how much of my own use of AI happens after the first prompt. I add missing context, ask the model to challenge an assumption, and change the question when I realize I was asking for the wrong thing.
That conversation made me think about what happens when the person writing the prompt is only one of several people involved. The next reviewer may know why a number is disputed. Another person may own the decision. A model can produce a polished answer while neither of those facts is available to it.
I have also been studying product development processes. The part that stays with me is how information becomes a decision and then a commitment of time or resources. Real processes can become heavy and formal, of course. Still, the question is useful when thinking about agents. Where would the agent get evidence? How would a person correct it? Which step can it prepare, and which decision belongs to someone who will live with the result?
I am early in this work and have no deployed enterprise agent to claim. For now, I want to understand one recurring handoff well enough to answer those questions with the people who do the work. That would give me a way to test whether an agent can make the handoff easier, or whether people need a simpler change in how they share information.