Pinkoi CEO Peter: If Sellers Don't Know They're Using AI, the Product Is Done
Pinkoi co-founder and CEO Peter discussed PinZap's design philosophy in an episode of the Side Chat podcast: for AI products targeting non-technical micro-sellers, the ideal state is that users don't realize they're using AI. Around this goal, he described three design choices—0-to-1 generation handled by AI agents, subsequent refinement via a WYSIWYG editor, and human approval for high-risk operations. He also pointed out the tension between AI lowering generation costs while increasing the cognitive load of human judgment, and noted that balancing token costs against quality improvements remains an ongoing challenge.
Pinkoi co-founder and CEO Peter said one thing about PinZap's product goal during the E428 episode of the Side Chat podcast: "If we can launch an AI store-building service where people don't realize they're using AI, then we've accomplished it."
PinZap is an AI store-building service from the Pinkoi team, targeting micro-sellers who want to do business online but lack technical backgrounds. Sellers answer three questions in natural language—who are you, what do you sell, what does your potential customer base look like—then drag in product photos, and an AI agent generates a website with integrated payment and logistics within ten minutes.
But "opening a store in ten minutes" is only the surface. What Peter really spent time explaining was why this product looks the way it does, rather than some other shape.
One key design choice is the "hybrid mode." Peter said that going from 0 to 1 with an AI agent is simple, but if every subsequent tweak—enlarging or shrinking text, adjusting whitespace—requires back-and-forth prompt editing, "it's actually quite painful." So the team separated generation from refinement: AI handles the full site from 0 to 1, while subsequent fine adjustments go to a WYSIWYG editor that sellers operate themselves.
This choice relates to the goal Peter repeatedly emphasized. He noted that many sellers don't know how to use AI, let alone how to command it. By handing refinement to a WYSIWYG editor, what sellers feel is "this tool is really convenient," not "I'm having a conversation with an AI."
Peter revealed a set of usage data: AI onboarding (generating a site from 0 to 1) has a 100% adoption rate, while product-page and blog optimization agents see roughly 70%. His explanation: many sellers migrating from Shopee or Maihuobian haven't opened their own websites because they don't know what one should look like—AI shortened the distance to imagination.
The second technical choice is routing models by task. Peter said quality is absolutely tied to compute cost, and opening up to micro-sellers requires balancing the two. The team integrated multiple models, assigning high-tier, mid-tier, and low-tier models by role. For example, copywriting uses a cheaper model specialized in text—"a very expensive model isn't necessarily better."
The third choice is the boundary of human approval. MCP lets sellers list hundreds of products at once, but write permissions affect sellers' actual data, and operations with higher risk levels require human approval—"it absolutely must be human in the loop." Peter put it plainly: "I think humans are the masters of AI."
Looking at these three choices, the division of labor follows a consistent thread: AI handles generation, analysis, and routing; humans handle visual judgment and oversight of high-risk operations.
But Peter also flagged a tension this model hasn't yet resolved. "The AI era has lowered generation costs enormously," he said, "but honestly, it has increased the cognitive load of human judgment a great deal." AI can produce many versions quickly, but sellers still need to pick one and tweak it to satisfaction—the judgment process itself has become heavier.
Token cost is another ongoing challenge. Peter admitted the team is still learning: new models keep arriving, sellers who see a ten-minute site launch then demand better images, and quality improvements drive up compute costs. Finding the balance between value and price has no ready-made answer.
PinZap took over a year to develop. Peter described integrating AI into the legacy Pinkoi platform as "a bit like changing the engine while the plane is in flight." PinZap, by contrast, was built around AI from day one, with each module clearly describing its own specifications like LEGO bricks, letting AI invoke them through natural language. Even so, the trust and governance mechanisms inside took considerable time—"because we were very worried AI would overstep."
Peter called this product direction a "quasi-second startup." He believes many existing e-commerce products were born ten years ago: "Think of it this way—a building from ten years ago and a building of the future will definitely be completely different." At the same time, he thinks people sometimes overestimate technology's impact; real societal change may take five to ten years to brew.
Peter said that sellers don't actually realize they're using AI—they just feel like there's something incredibly convenient about the tool. When you talk to sellers about AI agents, they may not understand the concept, but they can feel that work which used to take a long time now takes minutes. With the barrier to opening a store dropped this low, what remains as competitive advantage? Peter thinks it comes back to product strength and brand strength, including how you serve customers.