Stop Designing AI for Humans. The Future is Agents Talking to Agents.

I’ve spent eight years in the AI trenches. From building rule-based financial bots at Baidu to wrestling with hallucinating LLMs at ByteDance, I’ve watched this industry swing from ‘not smart enough’ to ‘too smart to control.’ I’ve made every mistake in the book so you don’t have to.

If you’re a product manager building AI today, you’re probably doing it wrong. We are treating AI like a fancy SaaS feature, forcing users into endless chat loops, and pricing our products like we’re selling electricity. It’s time to wake up.

Here are the five brutal truths I’ve learned the hard way.

1. Stop Charging for Tokens. Start Charging for Value.

Right now, the industry treats token costs like a premium feature. It’s not. Token costs are the new electricity bill—nobody pays a premium just to keep the lights on. Your AI’s value isn’t in the compute it burns; it’s in the revenue it generates or the costs it saves. Stop making users choose between a cheap, dumb model and an expensive, smart one. The product should invisibly route tasks to the right model. If the user has to think about model selection, you’ve already failed.

2. The Chatbot Interface is a Tax on User Patience.

Everyone thinks the future is a single, unified chat window. It’s not. While a unified entry point is inevitable, pure conversation is incredibly inefficient. If an AI compares prices across five vendors, the user shouldn’t have to chat with the AI to click ‘buy.’ Multi-turn dialogue is a tax on user patience. Sometimes, a single button click is worth a thousand tokens. Keep the UI. Let AI do the heavy lifting in the background, but let humans push the final button.

3. PRDs are Dead. End-to-End Execution is King.

In traditional tech, a PM writes a PRD, draws a wireframe, and begs engineering to build it. In the AI era, that workflow is obsolete. In the AI era, writing a PRD is like bringing a typewriter to a coding battle. You need to build, test, and deploy the prototype yourself. You don’t need to be a senior engineer, but you must be able to string together prompts, APIs, and workflows to validate an idea in 24 hours. If you rely on dev resources to test a hypothesis, you’re moving too slow.

4. If You Don’t Build Eval Loops on Day One, You’re Flying Blind.

AI models are fuzzy. Their outputs shift with every prompt tweak or data update. If you don’t bake an evaluation system into your product from day one, you’re shipping a black box. You need offline metrics before launch, user feedback mechanisms after launch, and an automated data loop that feeds corrections back into the model. If you don’t build the evaluation loop on day one, your AI product isn’t a software release—it’s a blindfolded gamble.

5. The Twist: Stop Designing for Humans. Build for Agents.

Here is the biggest misconception in AI today: we think the ultimate interface is human-to-AI conversation. It’s not. Humans communicate with words; machines communicate with data. You can’t explain a 100-page contract through a chat window. The real future is agents talking to agents.

Soon, your AI agent will negotiate with another company’s AI agent, handle the payment, sign the contract, and execute the delivery—all without a human playing middleman. The ultimate AI interface isn’t a conversation; it’s a machine-to-machine transaction. Stop designing for humans and start designing callable service units.

Are you building for the agent ecosystem, or are you just AI-washing a legacy SaaS product? The clock is ticking.