From HMI to HAII: Designing the Interface Between Humans and AI

For most of the history of human-machine interface design, the problem was simple in framing if not in execution: a human wants to do something, a machine can help them do it, the interface connects intent to action. The human operates. The machine executes. HMI is the layer in between. That framing is chaging now.

The shift: from operating to monitoring

What's happening across autonomous vehicles, military AI, industrial robotics, and increasingly consumer software is a fundamental redistribution of roles. The AI is now doing the task. The human's job is to understand what it's doing, decide whether to trust it, and intervene when it's wrong. This isn't a reduction of the human role, it's a transformation of it. And it makes the interface problem substantially harder. Operating a machine gives you constant feedback. Every action produces a response. Monitoring an AI gives you nothing until something goes wrong, and then it gives you everything at once. The cognitive challenge of sustained, low-stimulus vigilance followed by high-stakes intervention is one of the hardest problems in human factors. The interface has to carry most of the burden of keeping the human in a state of useful readiness.

The CNC moment

Germany's manufacturing sector faced a version of this transition in the 1970s and 1980s as CNC machining displaced manual operation on the factory floor. A skilled machinist running a manual lathe had continuous, physical contact with the work, they could hear the cut, feel resistance through the handwheels, read the quality of the chip coming off the tool. CNC automation removed all of that. The machine now ran the program. The human's job became writing the program, setting up the tool paths, and watching for the moment the process drifted out of tolerance. The interface shifted from handwheels and dials to a monitor and a keyboard. The German industry's response was instructive: rather than simply replacing machinists with programmers, the most successful manufacturers invested heavily in training and in the interface layer itself. They understood that the operator's ability to interpret what the machine was doing and catch errors early depended entirely on how well the control system communicated the machine's state and intentions. The quality of that interface was the quality of the factory.

AI makes this problem exponentially harder

CNC machines are deterministic. Feed a program the same inputs and you get the same outputs. You can read the G-code, trace the tool path, and understand exactly what the machine intends to do before it does it. AI systems are not like this. A neural network navigating a complex intersection, an AI pilot managing a contested airspace, an autonomous agent scheduling a fleet of vehicles, these systems make decisions through processes that are not directly readable by humans. The interface can't just expose the instructions because there are no instructions, only weights, probabilities, and learned behaviours. This is the core challenge of what I'd call HAII: Human-AI Interface. The design problem isn't connecting human intent to machine action. It's communicating AI intent to human understanding.

Communicating what the AI means to do

If a self-driving vehicle is about to yield to a pedestrian, the operator or passenger should be able to understand that from the interface before it happens, not reconstruct it afterward. If an AI agent is declining to engage a target because confidence is below threshold, the operator needs to see that reasoning in a form they can evaluate and override. If an AI is requesting human assistance because it has reached the boundary of its competence, the human needs enough context to step in with authority. These are fundamentally communication design problems. The AI has intentions and behaviours, and the human's ability to remain meaningfully in the loop depends on how clearly those intentions are expressed. Good HAII design is, at its core, a translation problem: it takes machine reasoning and renders it in a form a human can act on under workload.

What I've been building

The thread running through most of my career is this exact problem, before it had a name. At Zoox, TeleDash was built to give teleoperators the situational awareness needed to unblock a vehicle that had reached the edge of its autonomy. The vehicle had a decision it couldn't make. The interface had to communicate why, and give the operator enough spatial and contextual information to make that decision on the vehicle's behalf within seconds. At Shield AI, the Agent Designer on the Forge platform is a tool through which engineers define and configure how autonomous agents behave. It's a HAII surface in the purest sense: humans shaping the intentions of an AI system, with the interface carrying the entire translation burden between human intent and machine behaviour. And across detection and avoidance systems for unmanned aircraft, the design challenge is precisely about communicating the Autonomy's probabilistic assessment of collision risk in a way that pilots can integrate into their decision-making in real time. In every case, the problem is the same: the AI is capable, but the human must remain connected to it, informed, trusted, and ready to intervene.

The interface is not a feature

One pattern I see repeatedly in AI-first companies is that the human interface is treated as a delivery mechanism. The AI is the product. The interface is how you show the AI working. This is backwards. In any system where a human is expected to supervise, correct, or collaborate with an AI, the quality of the interface is the quality of the system. A brilliant autonomy stack that operators can't interpret correctly is a dangerous system. An AI agent designer that engineers can't reason through clearly will produce agents that don't behave as intended. The interface isn't downstream of the product, it is the product, measured by whether the human in the loop is actually useful. As AI capability accelerates, this becomes more urgent, not less. The more an AI can do, the more consequential it is when it does something wrong and the more critical it is that the human monitoring it can catch the error before it propagates. The interface is the last line of defence. It deserves to be designed like one.