AI should earn the right to act

AI needs to crawl before it runs

We've spent decades shaping how humans interact with software. But software, for the most part, still makes people do all the work. Click here. Type that. Navigate to settings. Flter to find. Dig through menus.

AI (agents in particular) promises to change that.1

1. Anthropic defines agents→ as "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. Unlike workflows with predefined paths, agents can plan and operate independently, potentially returning to humans for guidance when needed."

But before AI can act on its own (or replace effort), it has to earn the right to do so. Most apps today aren't broken because they lack automation. They're broken because they make users think like machines. And AI should change that. It should first remove the dumb parts of using software.

AI as an interaction layer

Instead of forcing users to adapt to rigid interfaces, AI should make software feel fluid. The way we interact with apps—through buttons, dropdowns, and menus—hasn't really changed in decades. AI should bridge that gap.

• If I need to mute Slack notifications at night, I shouldn't have to go three menus deep—I should just be able to type "Mute Slack after 6 PM" and it gets done.
• If I want to schedule a meeting, I shouldn't have to manually check my calendar—I should be able to say "Find a 30-minute slot with Alex next week", and it figures it out.
• If I need to pull up a past invoice, I shouldn't have to remember where it's stored—I should be able to ask, "Show me my last three invoices", and they appear.
• If I want to know how much I spent last month, I shouldn't have to open a dashboard—I should just ask, "How much did I spend last month?", and get an answer.

None of this is agency. It's just a better way to interact with software. But if AI gets this right, it earns the right to go deeper.

When AI becomes an agent

AI becomes truly powerful when it learns to spot patterns on its own, without being told what to look for. So an agent in this case doesn't just respond to commands—it acts independently, within limits. It doesn't wait for instructions every time. It recognizes intent (and the context around it), anticipates needs, figures out what to do, and then acts.

• After learning my email habits, it automatically filters low-priority messages.
• If it sees an upcoming trip on my calendar, it researches, plans, and drafts personalized recommendations.
• If I mute Slack at the same time every evening, it automatically sets up quiet hours.
• If it notices I frequently schedule meetings with a colleague, it blocks recurring time.

That's where AI stops being just an interaction layer and starts becoming an agent. But that transition has to be earned.

AI needs to earn the right to act

If AI jumps too fast, it fails.

It starts acting before it understands. Making decisions before proving it's useful. No one wants an AI that moves things around unpredictably or sends emails without knowing why. The moment AI feels intrusive, people pull back.

A good AI doesn't start by doing. It starts by removing friction.

AI doesn't need to replace human intent. It needs to amplify it. First, make interacting with software feel effortless. Then, once it actually understands how you work, it earns the right to take action.