Designing Human-Computer Interaction for AI Agents

As AI agents become part of everyday work, building capable models is only one part of the challenge. People also need to understand how to work with these systems, see what they are doing, and know when their input is needed. That makes human-computer interaction, or HCI, just as important as the underlying technology.

I recently shared some of our learnings from real-world enterprise deployments at an event co-hosted by the Stanford Institute for Human-Centered Artificial Intelligence and Wells Fargo. Three design principles stood out: make agents proactive, give people clear checkpoints, and keep the interface consistent as workflows grow.

Move from opt-in chatbots to opt-out agents

A conventional chatbot waits for a person to start a conversation. The user has to recognize a need, open the tool, and ask the right question. That can work well for some tasks, but it also puts the burden of initiating every interaction on the person.

An agent can take a different approach. Instead of waiting for a prompt, it can run upfront and bring the user in when their attention or decision is needed. The interaction shifts from “ask the system to do something” to “review what the system has prepared and respond where necessary.”

This is an important change in the relationship between people and software. The system takes more responsibility for moving work forward, while the person remains involved at the moments that matter.

In our deployments, shifting from opt-in chatbots to opt-out agents drove adoption from 15% to 100%.

That result highlights how much the interaction model can affect whether people use an AI system. A capable model is not enough if people must remember to seek it out at every step. Proactive assistance makes the system present in the workflow, rather than a separate destination users have to visit.

Replace sequential screens with meaningful checkpoints

When software guides a person through a process one screen at a time, the user often has to manage the sequence as well as the work itself. They click through steps, keep track of what has happened, and determine what remains.

An agent can handle more of that execution. The person does not need to supervise every individual action through a separate screen. Instead, they can oversee progress through higher-level checkpoints, such as checklists and activity logs.

This does not mean removing human oversight. It means designing oversight around progress and decisions, rather than making people follow every small operation.

• A checklist can make the current state and remaining work easier to review.
• An activity log can show what the agent has done.

Together, those views can provide visibility without requiring the person to navigate every step.

The distinction matters. Automation that hides its work can leave people unsure about what is happening. A checkpoint-based interaction offers another option: let the agent execute, while giving the human a clear way to monitor and guide the process.

Keep the interface stable as workflows expand

Enterprise workflows change. Requirements can grow, and teams may need to add new steps or capabilities over time. If every new requirement calls for a new screen, the interface can quickly become difficult to maintain and navigate.

A more stable approach is to extend the workflow through additional checklist items or agent capabilities, rather than building a bespoke screen for each change. The interface remains familiar even as the system takes on more work.

That consistency helps on both sides of the product. Users do not have to learn a new screen for every added requirement, and the product team can avoid multiplying interface components. The goal is not to prevent workflows from evolving. It is to make that evolution possible without turning each change into a new user experience.

In our experience, combining checkpoint-based oversight with a stable interface reduced our interface count from 103 screens to one.

That is a concrete illustration of how workflow design can shape the product itself. Fewer screens can mean less interface overhead, while still giving people the visibility they need.

Design the relationship, not just the model

These principles point to a broader lesson for AI products. The design question is not only what a model can do. It is also how the system begins work, how people stay informed, and how new capabilities fit into an established workflow.

• Proactive agents can reduce the burden of getting started.
• Checkpoints can help people oversee execution without managing every step.
• A stable interface can accommodate new requirements without adding a screen for each one.

Together, these choices help define a more useful relationship between people and AI systems. The agent handles more of the work, while the interface helps people understand progress and remain involved where needed.

As agentic systems become more common, that relationship deserves the same attention as the models themselves.

I also learned a great deal from Stanford Professor Michael Bernstein and his insights on the agentic frontier.