Essay · July 2026

Designing the Human Layer

AI delivers the signal. People still have to turn it into shared understanding.

AI can show you more than any organization has ever been able to see: the pattern across an entire data set, the anomaly forming in some corner of the business, the signal that used to take a quarter to surface. What it cannot do is turn that signal into shared understanding fast enough to act on. Only people do that, and most organizations have left it to chance.

That is the part the technology does not solve. A model can flag a risk in seconds. Whether the three people who see it read it the same way, agree on what it means, and move together is a separate question, and it sits above the tools. I call it the human layer, and in most companies no one designed it. It runs on whoever happens to be in the room, the relationships that happen to exist, the context that happens to be shared. When it works, people credit the team. When it fails, they blame the data. Since the layer was never built, no one thinks to look at it.

Watch it break. The machine surfaces a pattern. It reaches a risk lead, a finance partner, and an operator, and each of them, reading through a different frame, sees a slightly different thing. They agree on the words, say "improved insight" or "emerging exposure," and nod. Then they walk out and act on three different meanings, because the words were shared and the mechanism underneath was not.

I have watched sharp teams agree on a sentence and disagree on everything the sentence stood for, with no one noticing until the work came back wrong. Agreement on language only looks like alignment, and that is more dangerous than open disagreement, because it never announces itself. Derek and Laura Cabrera, who study systems thinking at Cornell, give the underlying move a shape: people build meaning by drawing distinctions, seeing parts and wholes, mapping relationships, and taking perspectives, and two people can hold the same information and land somewhere different if any one of those differs. Shared words paper over the gap. Designing the human layer means closing it on purpose.

The design is concrete. It is forums where weak signals get interpreted together before they harden into decisions; a vocabulary precise enough that a word means the same thing across functions; escalation paths that move at the speed of the signal rather than the speed of the calendar; and one small discipline I run at the end of any conversation that lands on a real decision, where each person says, in their own words, what they think was decided and what they will do first. If the versions match, you have alignment. If they diverge, the meeting is not over, whatever the agenda says.

This matters more now because of speed. When information moved slowly, the human layer had time to self-correct; someone caught the divergence in a hallway a week later, before it cost much. AI took away the week. The signal and the misreading travel at the same pace now, and the misreading compounds before anyone bumps into it. The faster the tools get, the more deliberately you have to build the human layer, because the slack that used to absorb the gaps is gone.

This is the work my team is on now, and it is harder than the technology was. Building the agents was a known problem with known steps. Designing the layer above them, so that insight becomes shared meaning by default rather than by luck, comes with no manual. We know it is possible, because we have seen it work in the moments we designed for. The open question is how to make it run without depending on the few people who happen to be good at it.

That is the real test of the human layer: whether the system produces shared understanding even when your best people are not in the room.