Podcast Atlas
concept Updated 2026-08-06 Topics: Technology

Vertical Workflow AI

Vertical workflow AI is the source’s nameable pattern for AI products that solve a concrete domain workflow rather than only expose model generation. In 263.Sora死了，Adobe跌了，美图何去何从？, the clearest examples are Meitu Design Studio / 美图设计室, Kaipai / 开拍, and Wink, where output has to fit ecommerce, advertising, beauty, video, teleprompter, or training-material needs.

The source argues that AI can often produce a rough 80-point result quickly, but users with existing 95-point workflows still need the final 20 points: taste, correction, consistency, delivery format, batch operations, and business-fit judgment. That last-mile work is where AI Application Layer Moat can remain.

一个 AI 创始人的虚荣心、装，和愚昧之巅｜对谈 invoko.ai 创始人梦琪 adds a boundary case. invoko.ai / Invoqo’s early Sourcing Agent and growth-Agent work looked vertical, but 梦琪 / Mengqi later concluded that automating a low-value slice can still produce a SaaS-like or agency-like product if the user needs process controls, communication, and service follow-through more than the initial AI-generated list.

Founder Mode: Jake Heller, Founder & CEO, Casetext adds a legal-workflow case through Casetext and Co-Counsel. Jake Heller says the breakthrough was not simply that GPT-4 wrote better text, but that the company could test it against legal research, document review, contract drafting, and other customer-requested workflows. That makes the product a Frontier Model Inflection Pivot example where domain context and customer memory turned a frontier model into a vertical product.

Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授 adds Scenario-Specific AI as a sharper unit of analysis. 张奇 says AI’s core is the scene rather than the broad industry: Cursor is strong because it is optimized for coding, Perplexity is strong because it is optimized for search, and even one domain may require different products for beginners and experts.

Key Claims

  • A vertical workflow is defined by the downstream job, not by the model modality.
  • The product must understand acceptance criteria, not only generate plausible media.
  • Vertical workflow products can use Model Container Strategy because the model is one component of the workflow.
  • To-Agent Distribution can expose vertical workflow capabilities to agents once the capability is stable enough to be called from outside the app.
  • A vertical Agent is not automatically a strong vertical workflow product; the product must cover the valuable job, not only the automatable step.
  • Expert-user controls can improve reliability while also pulling the product away from end-to-end automation.
  • A frontier model can create a vertical workflow opening only when the product team translates raw capability into domain tasks, reliability expectations, and buyer timing.
  • A vertical label is still too vague unless it names the concrete scenario, user type, input, output, review standard, and downstream action.

Connections