Blueprint for an AI-First Organization: The Practical Solution of 'Flows x Agents' Demonstrated by Zapier
AI-first companies are not simply companies that add "AI features." They are companies that rebuild everything—from process design and authority structures to talent requirements—with an "AI-first" premise.Wade Foster, co-founder and CEO of Zapier, provides statements and real-world examples that clarify this vision. This article explains the hybrid design of agents and workflows, the automation of renewal negotiations, the hiring and internal culture of AI-fluent talent, and the realities of funding and business operations, incorporating quotes and case studies.
1. Zapier's Evolution into an "AI Orchestration Platform"
1-1. The Three-Stage Evolution: Integration → Automation → Agents
Zapier has evolved from simple "trigger-to-action" connections to multi-step deterministic workflows, and now to agentic workflows. Foster emphasizes a design that integrates everything from context collection—such as CRM, Zendesk, and conversation logs—to executing multiple prompts, and even generating sales preparation materials and email drafts.
1-2. Explosion of Adoption Use Cases
The company has seen a surge in AI task usage. They have even announced a record of 50 million tasks in 20 days. This is evidence that practical, "real-world" use cases are expanding rapidly.
2. Case Study: Enterprise Renewals Agent
2-1. What is Being Automated?
The renewal negotiation agent built by the RevOps team performs cross-analysis of usage trends from the past year, Gong sales call transcripts, and Zendesk tickets. Through a refined internal judgment prompt, it generates recommendations for upsells, flat renewals, or price reductions, automatically registers them in HubSpot, and even creates email drafts for the account manager. The human representative handles the "last mile" by reviewing and making the final decision.
2-2. The Essence of the Results
Foster highlights the philosophy that "AI does 90%, and humans handle the final 10%." Deterministic (accurate, low-cost) parts are handled by flows, while parts requiring ambiguity or interpretation are handled by agents. This "hybrid" approach is the most practical solution currently available, as it is easy to implement and minimizes the risk of errors.
3. Design Principles for Workflows x Agents
3-1. Identifying "What Stays the Same Every Time"
Input normalization, data entry, and auditing should be fixed as deterministic flows. Conversely, tasks without a single correct answer—such as summarization, insight generation, and proposal drafting—should be delegated to agents.
3-2. Feedback Loops
Humans read the automatically generated briefs or scripts, verify them against actual call results, and then re-train the prompts or judgment criteria. Humans are positioned as the final quality assurance (QA) layer and providers of learning data.
4. Building an AI-First Organization: Hiring, Development, and Culture
4-1. Making "AI-Fluent" a Hiring Standard
Zapier has institutionalized hackathons (every 4–6 months) and "show-and-tell" sessions at All-Hands meetings to continuously share internal success stories. In hiring, they assess skills through practical tasks and screen-sharing problem-solving exercises. They have clearly defined the "expected level of AI proficiency" for each role to serve as a benchmark for hiring decisions.
4-2. Start Small and Iterate Quickly
"AI-first cannot be achieved with notes alone." Build a loop of process definition, embedding into daily operations, and sharing in regular meetings to avoid reliance on specific individuals.
5. Ecosystem Expansion: MCP is a "New Protocol"
MCP (Model Context Protocol) is a "new pipeline" that enables conversational access and actions for tools and data. It requires an ambidextrous approach: retaining the stable, low-cost, deterministic processing that traditional APIs excel at, while incorporating variable use cases that expand through agent-to-agent and human-to-agent interaction via MCP.
6. Summary: If AI is the Premise, Then "Decomposition -> Wiring -> The Last Mile"
AI-first is not "magical full automation," but rather a grounded implementation involving the optimal division of deterministic and agent processing, the human "last mile," and the institutionalization of internal learning.
Start small, and rapidly cycle through real data, feedback, and improvements to prompts/judgments. Zapier's journey to a $5B valuation through capital efficiency proves that you can win through operational design and customer value without relying on excessive funding. First, take inventory of your company's "always the same" tasks and "always require deliberation" tasks; assign the former to flows and the latter to agents. That wiring diagram becomes your blueprint for an AI-first organization.