The Trust Ladder: Why AI Adoption Should Be Earned

Trust grows when people can understand what AI is actually doing before they choose to ask it to do more. That progression rarely fits neatly into a yes-or-no decision tree. Instead, trusting AI to flag an unclear sentence, prepare a reply, and send a message without review grants three different levels of authority, each requiring its own evidence.

Previous Superhuman research has distinguished between familiarity and trust when it comes to AI. Repeated exposure strengthens trust when the experiences themselves are reliable. People need lower-stakes opportunities to see how AI performs before relying on it in consequential situations.

As AI takes on more of the process, visibility becomes even more important. “There’s this notion of automation and ‘take it away from me,’ and it starts to feel like a little black box,” says Ailian Gan, Head of Agents at Superhuman.

The three levels of the trust ladder help teams decide how much AI gets to do, based on what it’s already proven it can do:

1. Suggestion: AI proposes a change.
2. Instruction: AI takes initiative within clear human-set boundaries.
3. Delegation: AI handles more of the process within established limits.

Each rung of the ladder preserves human control while increasing how much initiative AI can take. Teams can advance along the ladder when AI performance supports it and stop at the level appropriate for their workflows.

Suggestion: Let people see and judge the help

Grammarly’s underline is an easy-to-understand example of the first rung of the trust ladder. The product identifies opportunities to improve writing and highlights them without replacing the original text. Its visible, reversible suggestions let people evaluate AI as they work. The writer can inspect the suggestion, revise it, and choose to accept it or dismiss it.

“If you want help from Grammarly, you can just click on it,” Gan says. That small interaction creates several conditions for trust:

• AI’s role is clear.
• Its recommendation is visible.
• The writer retains authority over the final words.

Every suggestion shows you a bit more about what the system does well and where you’ll want to rely on your own judgment.

The same type of AI interaction can provide help beyond grammar. Superhuman’s Knowledge Checker agent can flag a questionable claim and surface relevant information from connected sources. This allows the writer to focus on expressing the idea while the agent checks the supporting details. “I don’t have to remember everything,” Gan says. “I just need to get my words and my thoughts out.”

AI earns the next rung of the trust ladder when its suggestions are consistently useful, understandable, and easy to correct when needed. Frequency alone doesn’t add value if people regularly dismiss the suggestion or rewrite the output. “Having those very subtle layers of control help build the user’s confidence that I can use AI in the way I want,” says Gan.

Instruction: Give AI a defined job

At the next level, you give AI one specific job and say what a good result looks like without telling it how to get there. For example, you could ask Superhuman Go to turn a pile of notes into a first draft, or to find an hour of focus time on a Thursday.

Triggers and configuration can also add another layer of direction and control. A team can determine when an agent should run, which context it can use, and which situations call for its help. “You can define triggers for when they should push for you,” Gan says. “You can set triggers for when you want these agents to run.”

This structure allows AI to become proactive within boundaries a person or team has set. The agent may recognize that a defined condition has been met and offer help without waiting for another prompt, but its initiative remains tied to a known task, trigger, and set of information.

The evidence needed to advance along the trust ladder depends on task completion. The system should follow the directions, use the appropriate context, stay within scope, and make its output easy to review. In the instruction rung, missing or conflicting information should remain visible rather than disappearing into the process.

Delegation: Allow AI to handle the process, but check its work

At this level, AI determines how to get work done, and a person signs off on the final product. In practice, this means people spend less time doing the work and more time setting limits, checking how the work is going, and approving high-stakes outputs.

Superhuman Mail’s Auto Drafts shows how this can work. To draft an email reply, Auto Drafts can search older emails for context, look things up online, check your calendar for availability, and tailor the language to your voice. Rahul Vohra, Founder and Head of Superhuman Mail, described an Auto Draft that surfaced an internal resource the user hadn’t considered. “The kernel of the perfect response was actually in another email,” he says. “The AI took that and did a bunch of research online, and then it wrote a reply using the user’s own voice and tone.”

While AI handles much of the preparation autonomously, the person still decides whether the message is ready to send. Auto Drafts gets a lot of freedom up front, but nothing goes out to another person until a human has looked at it.

Usage provides evidence of how people react to this mix of autonomy and control. Vohra says 40% of Auto Drafts are sent within one day, and 60% of those are sent without edits. Superhuman also found that users save about nine minutes per Auto Draft. At four drafts per workday, that adds up to three hours each week.

Each rung requires evidence

The appropriate level of autonomy depends on the stakes, the action's reversibility, and the consequences of an error.

Advancement is optional and doesn’t have to happen across all tasks all at once. A team may delegate meeting summaries while keeping financial analysis at the instruction rung. The same organization may allow AI to prepare customer emails while requiring a person to approve every send.

Product capability and organizational authorization also develop at different rates. An AI system may technically be capable of completing an entire workflow before a team has enough evidence to grant that authority.

This distinction helps organizations move quickly without building trust on a weak foundation. Teams can begin experimenting at the suggestion rung, where outputs remain visible and errors are easy to catch. Each successful interaction provides evidence for expanding AI’s role.

Moving faster than the evidence allows creates a larger failure surface. A visible error after premature delegation can weaken confidence in the system and discourage people from using it in lower-risk situations. Repeated reliable performance produces trust that can withstand an occasional correction or exception.

How to earn the next rung

Organizations can apply the trust ladder to one workflow at a time:

1. Choose a specific task or workflow
2. Begin at the lowest useful level of autonomy
3. Keep AI’s work visible and correctable
4. Define the evidence required for a larger role
5. Preserve human approval for consequential actions
6. Expand AI’s authority when repeated performance supports it

This approach gives people time to understand how the system behaves while allowing the organization to capture value early. Trust grows alongside capability, grounded in work people can inspect and outcomes they can evaluate.

“AI should enhance the way you work,” Gan says. “It should give you control and give you a sense of agency.” That agency is the foundation of durable adoption. As AI takes on more responsibility, people gain the confidence to use it for bigger tasks, new skills, and more ambitious work.

Learn more about how Superhuman applies these principles across writing, email, agents, and connected work.