Mercari's AI-Native Transformation: Measurement, Platform, and Culture

Mercari's aggressive pivot to an AI-native operating model provides a definitive blueprint for enterprises navigating the chasm between AI adoption and tangible business value. Following a CEO mandate requiring 100% AI utilization across all roles, the Japanese marketplace platform—serving 23 million monthly active users—uncovered that tool proliferation without strategic alignment leads to fragmented productivity and hidden inefficiencies. The organization's experience highlights that AI acts as a force multiplier for both capability and dysfunction, necessitating a comprehensive overhaul of measurement frameworks, platform infrastructure, and development lifecycles.

The Measurement Imperative: Beyond Code Generation Metrics

Mercari's initial challenge was a critical visibility gap. Despite winning developer experience awards, existing dashboards could not quantify AI's impact on engineering velocity. Leadership demanded answers regarding ROI, but metrics focused solely on lines of code or tool usage failed to capture system-wide performance. In response, Mercari instituted a strict rule: "If you don't know how to measure, you don't start." The engineering team constructed a new DX visibility layer that ingests telemetry from AI tools like Cursor and DevIn, stitching this data with SDLC systems such as GitHub, Jira, and Datadog. This integration enabled tracking of nuanced metrics including Gen AI code acceptance rates, cross-platform PR cycle times, and deployment frequency. The data revealed a stark reality: local optimizations within teams did not translate to system-level gains. While individual engineers reported faster coding, end-to-end cycle times remained stagnant due to bottlenecks in review queues and CI pipelines. This measurement shift forced a transition from vanity metrics to value-chain analysis, ensuring that AI investments directly correlated with delivery throughput. The mandate to "rethink everything" permeated the organization, challenging assumptions about what to build, how to build it, and how teams operate. Leadership questioned whether processes requiring weeks for development were still necessary when AI could generate ten ideas in a month. This cultural shift was supported by data showing a 64% year-over-year increase in engineering output proxies, though leadership cautioned that output volume does not equate to value delivery. The 70% utilization rate of AI-generated code indicates strong tool adoption, yet the disparity between code volume and release velocity highlighted the need for deeper integration.

Infrastructure Stability as the AI Prerequisite

A profound insight from Mercari's journey is that AI cannot compensate for a brittle foundation. As AI agents accelerated code generation, they collided with legacy constraints, amplifying existing friction points. The "blast radius" of AI errors proved larger than human mistakes, and debugging AI outputs initially consumed more time than the savings generated. Mercari identified CI flakiness, review latency, and legacy monolith complexity as primary drag factors. The review queue emerged as the biggest drag on the company, not coding or CI. The organization responded with aggressive platform engineering investments, reducing iOS CI time from 16 minutes to 8 minutes and cutting deployment P95 from 48 minutes to 22 minutes. Furthermore, cluster right-sizing efforts yielded a 43% reduction in resource costs, demonstrating that operational efficiency is inextricably linked to AI scalability. The data confirmed that AI exposes systemic weaknesses; without stable pipelines and reliable infrastructure, AI velocity merely accelerates the accumulation of technical debt and maintenance overhead. The organization emphasized that brownfield modernization, rather than greenfield development, represents the true economic opportunity for AI. Agents are increasingly deployed to refactor, document, and test legacy services, addressing the "code bases nobody wanted to touch." However, this requires agents to navigate heterogeneous architectures where context is often missing. Mercari's approach includes rigorous guardrails, particularly for fintech components, ensuring that risk mitigation evolves alongside automation. The lesson that "gates slow you down" led to a systematic audit of approval steps, revealing tribal knowledge and unowned services that hindered progress. By manually walking through service creation and infrastructure changes, the team identified and eliminated hidden gates that would otherwise throttle agent performance.

Cultural Engineering and Frictionless Adoption

Achieving 100% adoption required dismantling cultural and procedural barriers. Mercari recognized that AI integration was a community problem, not merely a tooling issue. The company deployed a 40-engineer AI Task Force and identified 100 AI Champions across departments to drive enablement. To combat the awareness gap, Mercari launched AI Jamborees, pizza demo days, and role-specific workshops for non-engineering functions like finance and legal. Friction reduction was equally critical; a Slack bot was introduced to automate tool provisioning, handling security configurations instantly. These initiatives yielded immediate operational wins: help desk workloads dropped by 60%, platform support burdens decreased by 25%, and accounting tasks fell by 45.7%. Additionally, the internal BI agent "Socrates" democratized data access, driving 500 weekly users to ask business intelligence questions, thereby fostering a product-centric mindset across the organization. The task force mapped 33 domains in 100 days, identifying leverage points far beyond engineering. This holistic approach ensured that AI benefits rippled across the entire enterprise, reinforcing the cultural shift through visible, measurable outcomes.

The Second Dip and Spec-Driven Development

Mercari's adoption curve followed a "W-shape," characterized by productivity dips as new challenges emerged. The "Second Dip" occurred when code generation speed outpaced release throughput. While greenfield projects saw dramatic acceleration—such as the NACO project completed in 1.5 months versus a six-month estimate—teams stalled due to upstream planning bottlenecks and downstream review constraints. Maintenance PRs spiked as code quality suffered from rapid generation without adequate context. To address this, Mercari pioneered "Agent Spec-Driven Development," moving AI agents to the front of the lifecycle. Agents now ingest code, documentation, and Slack history to generate specifications before implementation, ensuring that execution aligns with product intent. This shift emphasizes that the bottleneck is often idea-to-code translation, not code synthesis itself. The W-curve serves as a critical framework for leadership expectations. Mercari experienced two distinct dips: the first driven by visibility gaps, and the second by the mismatch between coding speed and system throughput. Leadership anticipates a third dip as agents interact more deeply with legacy systems. This pattern underscores the necessity of the Stockdale Paradox: maintaining total faith in the AI destination while confronting the brutal reality of current constraints. Optimism without data leads to crushed expectations, whereas disciplined iteration allows organizations to navigate the dips.

Future State: Multi-Agent Loops and Behavioral Shifts

Looking ahead, Mercari is restructuring development around heterogeneous small teams and multi-agent workflows. The architecture is evolving into distinct loops: an inner loop for intent, generation, and validation; an ephemeral loop for isolated testing; and an outer loop for deployment and measurement. The strategic goal is to eliminate human approvals in inner loops, allowing agents to iterate autonomously while humans focus on guardrails and risk management. This transition requires significant behavioral change, as engineers must adapt to reduced accountability for execution and increased focus on constraint setting. The behavioral challenge remains the hardest hurdle; engineers often resist removing human approvals due to fears of accountability loss. Overcoming this requires a cultural pivot where humans set constraints and direction, while agents handle execution, fundamentally redefining the software engineer's role. Mercari embraces the Stockdale Paradox, maintaining faith in the multi-agent future while practicing brutal honesty about current realities. The organization acknowledges that legacy systems and heterogeneous architectures will continue to challenge agents, necessitating ongoing investment in brownfield modernization and robust observability. Ultimately, Mercari's strategy demonstrates that AI success depends on a holistic integration of measurement, platform reliability, cultural enablement, and lifecycle innovation.