Why the rich don't click buttons, and how AI brings this to everyone
A normal person goes to the grocery store to buy groceries. A wealthy person decides what they want and has someone else get it. An even wealthier person gives the directive "I want to eat healthy" and has a team figure out what to buy and how to buy it.
This pattern repeats across domains. Tax filing. Managing schedules. Running errands. As people gain resources, they insert human intermediaries to handle progressively more of the decision-making.
AI enables this same shift in how we interact with software. The companies that recognize this first will win the coming wave of agentic use cases.
LLMs are improving rapidly in intelligence and the length of tasks they can complete.1 Yet, consumer AI adoption is mostly limited to ChatGPT and the occasional slop video on Instagram while agentic AI is still mostly absent from enterprises.2 Where are the personal assistants and remote co-workers we have been promised by movies like Her?
The bottleneck is not intelligence. Intelligence is still spiky, but with enough context it works wonders (example: Claude Code). The hold-up is the product interaction paradigm, which needs to shift to unlock most of AI's value.
AI enables three interaction models: Self-Serve (you do everything), Natural Language (you tell it what to do), and Proactive (it figures out what to do).3 Each model shifts cognitive load to a higher level of abstraction. At the bottom is execution. Above that is figuring out what to do. At the top is pure intent. Each step up increases your leverage.
In coding, this shift is already happening. Software engineers write code. Engineering managers make decisions while others write code. CTOs dictate core principles while engineering managers handle implementation.
For many workflows, Claude Code has abstracted software engineers to engineering managers. It's only a matter of time until it abstracts us even higher. Instead of "re-write my checkout form," you say "optimize the checkout" and it figures out the best strategies and implements them. Eventually, the agent pings you unprompted: "I optimized your checkout flow and ran A/B tests. Conversion is up 12%. Approve deployment?"
Self-Serve: The past decades featured software like TurboTax. You control it. You use the UI to fill in the details. The mental load of translating your income and deductions into checkboxes and input fields sits entirely with you.
Natural Language: Drop your tax files into an upload and the AI analyzes them. It figures out what deductions it can claim and might ask you follow up questions. This is akin to hiring a human accountant to do your taxes at the end of the year.
Proactive: You subscribe to a tax agent. Throughout the year, it monitors your email, calendar, and financial context. It knows what you buy and what can be deducted. It advises you to sell stocks for tax loss harvesting. By tax filing season, it has your return ready and optimized, maxing out your deductions. You submit with one click. The UI becomes superfluous. Everything happens in the background. This is akin to have a team of tax lawyers and accountants proactively managing your taxes for you.
Successful companies in the next wave will move toward the proactive model as quickly as possible. They will bias toward gathering context and fixing issues before users ask, rather than only when prompted.
Most AI personal assistants today remain stuck in the natural language phase. Users must know what to delegate and explicitly prompt each task. This limits adoption to power users who already think in terms of leverage and delegation. The general public won't adopt until someone figures out the proactive step.
The key distinction between natural language and proactive models is not just abstraction level. It's who initiates action.
In natural language interfaces, the user thinks of solutions and the model executes. In proactive interfaces, the model thinks of solutions, gets user approval, and then executes. The inertia shifts from the user's side to the agent's side.
This is not a gradual shift. It's a fundamental inversion of the interaction model.
1 Measuring AI Ability to Complete Long Tasks - METR (March 2025). The time horizon doubled every 7 months on average from 2019-2025, with acceleration to every 4 months in 2024-2025. Models progressed from completing tasks that take seconds (GPT-3 in 2020) to 4 minutes (early 2023) to 40 minutes (late 2024).
2 Enterprise adoption data from McKinsey State of AI 2025 (88% regular use, 80%+ report no meaningful EBIT impact) and Gartner (August 2025) (agentic AI in <1% of enterprise applications in 2024, predicted to reach 40% by 2026).
3 My co-founder views this as two steps rather than three: from UI to Natural Language. He sees Natural Language as one long continuum where, with increasing intelligence and context, you give agents higher levels of abstraction (example: "make the checkout button bigger" → "optimize the checkout page" → "build the product to generate the most money possible"). I split Natural Language from Proactive to highlight a key distinction: in Natural Language, the user thinks of solutions and the model executes; in Proactive, the model thinks of solutions, gets user approval, and executes. The inertia shifts from the user's side to the agent's side.