The bottleneck in building agents is human communication

The whole market has decided the hard part of AI is the AI.

I build agents for real businesses, every day, and I’ve come to think that’s exactly backwards. The models are cheap now. Anyone can stand up an agent that holds a conversation or works a tool. The hard part — the part nobody has solved — is getting what’s in people’s heads into a shape an agent can actually use.

The bottleneck in building agents isn’t deployment skills. It’s human communication: how well one person, or a group inside an organisation, can explain and articulate not just their use case, but their procedures and their context.

Here’s a story I heard about a year ago. A business paid an AI agency to come in and make them agentic. They spent the best part of £50k on the tools. Good tools, too. But when the agency sent someone in, they couldn’t bridge the gap. They couldn’t pull out what the staff actually knew — the context they worked in, the intent behind what they did, the procedures they followed. The tooling was never the problem. The knowledge never made it across.

Give those three things names, because they’re the whole game: business context, business intent, business procedure. That’s what you’re trying to get out of people. And almost none of it is written down where an agent can reach it.

Some of it is half-buried in the codebase — a comment, a variable name, a flag that only makes sense if you already know the rule. The rest never touches the code at all: it’s in a head, an SOP, a meeting that happened once. Two lanes, and both of them free-form. Unstructured either way.

Take a button on a CRM. The code tells you exactly what the button does — it’s right there. But there are unwritten conditions that have to be met before someone’s even allowed to press it. That rule doesn’t live in the codebase. It lives in the head of the one person who’s been doing the job for six years. They know. Nobody wrote it down.

So what people do instead is record the meeting. Video the whiteboard. Dump the transcript in as context and hope. And it sort of works — but it lands unstructured. And unstructured context is a bucket, not a cupboard. You’ve got everything and you can find nothing. The agent drowns in it the same way a new hire would.

This is what’s on my mind at the moment. I don’t have it wrapped up — I’m looking for the answer as much as anyone. But there has to be a way to get business context, intent and procedure up out of people and into an agent-friendly format. Something like what Google’s Open Knowledge Format is reaching for: every fact in a known place, in a known shape. Less bucket, more cupboard.

Because the bottleneck is the humans. It’s the communication layer between them. How well an organisation and its people can say what’s in their heads, how open they’re willing to be, and how good the person building the agent is at going into those departments, sitting with people, and drawing it out — that’s what will decide whether any of it works.

The bottleneck is human communication.