LLMs might be better coders, but humans are better communicators

AI doesn’t have an output problem. It has an input problem.

For most coding and analytics tasks, LLMs perform as well as a human. Sometimes better. The models are good enough. And yet the results often aren’t.

Recently I saw this play out: someone filed a request for “development of revenue.” Routine ticket, handed to an AI agent. It executed flawlessly. Pulled the data, built the chart, delivered exactly what was asked. The problem: “development of revenue” wasn’t what the requester actually wanted. Anyone who’d worked with this person for years would have known. What they meant was weekly YoY development of net revenue after returns. The agent didn’t know that. And honestly, neither would a new hire.

That’s a context gap and no model fixes it. Human intuition built up over months and years. A human analyst reads between the lines because they know who filed the ticket, which business function cares, what projects are running this quarter. No system prompt captures that.

We can throw more context at AI agents through documentation and retrieval systems. It helps your AI understand your business better. But there’s a ceiling. If your AI keeps delivering technically correct but practically wrong results, the problem is your ability to specify what you actually want. Specification is a skill and most teams are still treating it as a tooling problem.

So before you blame the model, ask yourself: could a brand-new hire, with no context about your business, deliver what you wanted from that same brief?

If not, your AI never stood a chance.

Half the work I’m doing with teams right now isn’t about the AI. It’s about how they specify what they want.

#AI #AgenticWorkflows #AIStrategy

First comment: One habit that’s helped me: before I hand a task to an agent, I write a two-sentence “context header” at the top of the brief. Who is this for, and what decision will it inform. It takes 30 seconds and cuts the back-and-forth in half. Simple, but it changed my hit rate more than any prompt template.