Ask a company to design an AI agent and it will usually design an employee. The digital coworker. The AI analyst. The tireless assistant. We take a brand new kind of intelligence and turn it into a person with extra speed.
Structured imagination
There is a reason we do this. The psychologist Thomas Ward asked people to invent animals that might live on another planet. They came back with creatures that had symmetric bodies, two eyes, four limbs, and familiar senses. Free to imagine anything, people drew Earth animals with a twist. Ward called this structured imagination. Even our wildest ideas are built from things we already know. And when we design AI agents, the thing we know is people.
Copying human ability was AI’s original goal. But when it came time to put AI to work, the goal quietly became a limit. An agent designed as a stand-in for a person inherits the shape of a person’s work: one task at a time, one point of view, handoffs and queues. The ceiling is set at what a person, or a team of people, could have done.
What would truly AI-native design look like?
Honestly, nobody knows yet. That is what makes the question worth asking. Why should an agent read contracts like a lawyer, one at a time, when it could hold every contract the company ever signed in view at once? Why should it be one employee with a role, instead of a thousand small specialists created for a single question and then discarded? Why should work pass from colleague to colleague, losing context at every handoff, when memory can be shared instantly across every copy? None of these is always the right answer. Each is a tradeoff. But systems built by asking these questions will not look like a team of people. That is the point.
Familiar on the surface, AI-native underneath
That does not mean the tool has to feel strange to use. The surface, where people interact with the system, should feel familiar. Good AI meets users with familiar formats and workflows. That lesson still holds here. A contract-review agent can send its findings as a memo to the general counsel’s inbox while nothing underneath works like a lawyer reading files. Familiar on the surface. AI-native underneath.
Ward’s participants were not wrong that eyes and limbs are useful. They just never found out what else was out there. The companies that get the most from AI agents will be the ones that notice the human template in their designs and set it down. Not asking what a person would do here, but what this technology can do. The rest will build aliens with extra eyes: familiar, competent, and a fraction of what is possible.
Reference: Ward, T. B. (1994). Structured imagination: The role of category structure in exemplar generation. Cognitive Psychology, 27(1), 1–40.