Users can understand what an AI agent does and still choose to handle the work themselves. The higher the stakes, the harder it is for them to step away. Giving it a chance becomes worthwhile when the problem hurts enough to outweigh their concerns.

Illustrative model: users give AI a chance when the cost of continuing manually outweighs the perceived risk of delegating the task.

As PMs, we need to understand both what the problem costs users and what worries them about delegating it. They might fear a wrong answer reaching a customer, a change they can’t undo, or spending just as much time supervising the agent as doing the work themselves.

Three ways typical users handle a task

Mode	How the user handles the task
Human-only	They do it themselves. The task may be novel or complex, they may not know how to use AI for it, or they may have tried and been disappointed.
Human-in-the-loop	AI does part of the job, while they review, correct, or steer it.
Agent-only	They trust the agent to complete the task without routine supervision. This is usually repeatable work where it has demonstrated consistently satisfactory results.

The mode can change with experience

Over time, as users get more exposure to agents, a task can transition from one mode to another. Users from all sorts of backgrounds and trades learn how to configure agents, see where they succeed or fail, and get used to relying on them.

Illustrative model for one user and task, drawn like the Kano model: willingness to delegate rises with how well the agent does the task, slowly for human-only tasks, steadily for human-in-the-loop tasks and early for agent-only tasks. Over time, as novelty and risk fall and trust grows, a task moves from human-only toward agent-only.

With the exposure comes their changing expectations. When we are introducing AI agents to our users, we should remember to apply the Kano model. In AI, a feature can move from a delighter to a basic requirement in weeks.

The Kano model applied to AI: delighters are AI that surprises users, and nobody misses it yet; with performance features, the more accurate it is, the happier users are; basic expectations are AI users take for granted and notice only when it fails. In AI, delighters quickly become expectations.

What we can do

Make the boundaries visible

The scope of the task is part of the user’s decision to trust AI with it. If they can predict what the agent will handle and where it will stop, they have something concrete to evaluate. If it promises to handle everything, they have to imagine everything that could go wrong.

We should make clear what information it uses, what it can do (read, write, or delete), and what happens when something falls outside its scope.

When review before sending fits the task, it lets users experience the output while keeping responsibility for the final action. For other workflows, they should be able to test their chosen configuration, change the settings, and try again free of charge.

Nurture users for the mode they’re in

Our nurturing campaigns should reflect the user’s current trust in the agent for that task:

Human-only: remind them what the manual task costs, address their specific concerns, and offer a free sandbox. The goal is getting them to give the agent a chance.
Human-in-the-loop: help them evaluate the output, correct recurring mistakes, and identify which parts can run without their approval. The next step may be delegating one reliable part of the workflow.
Agent-only: keep them informed about completed work and celebrate milestones. Don’t make changes to a running agent’s behavior. Create new agents instead.

Finding opportunities with techniques that check each other

All articles