I'm building an AI agent for a client to handle parts of a product manager's role. Three weeks in, I've barely touched the agent.
Nathaniel Whittemore draws a useful distinction between two modes of AI work. Efficiency AI speeds up things you already do: faster drafts, quicker summaries, cleaner meeting notes. Opportunity AI is different. It lets you do work you couldn't do before, or not at any reasonable cost.
Most of the AI conversation right now, especially around agents, is sold as opportunity AI. A lot of what's actually getting delivered is efficiency AI in nicer packaging.
There's a reason for the gap, and I'm in the middle of it.
The agent I'm building is meant to pull together two kinds of signal about how a product is being used. The numbers side (how people actually behave in the app) and the human side (what they say in forum posts, support conversations, and direct feedback). A product manager spends hours every week stitching those together. In theory, the agent should make it overnight work.
In practice, the data isn't in a state the agent can use. The numbers have gaps. The feedback is spread across three platforms that don't talk to each other. A question as simple as "what are users struggling with this month?" hits a wall the moment you try to answer it properly.
So the last three weeks haven't really been about the agent. They've been about tidying up the underlying data and, just as importantly, the process around it. Who owns which signal. How feedback gets captured in the first place. Who reviews what the agent produces before anyone acts on it.
For nonprofits thinking about agents: opportunity AI isn't waiting on a better model. It's waiting on the data and process work most organizations have been putting off. The agent can't rescue a data layer that isn't there.
Efficiency AI is a fine place to start. Just be honest about which mode you're in.
If you've tried to build an agent, where did you hit the wall?