Key Takeaways

Chung's 70/15/15 — A Quality Principle Forced Into a Korean Carmaker

According to Chosun Ilbo's June 2026 interview series, Chung Mong-koo appointed Kim Sang-kwon as quality chief after taking the chairman role in 1999. The mandate: 70% of quality originates in design, 15% in production, 15% in parts. Kim assigned 70% of improvement tasks to the design division. When designers pushed back, he held the line. When departments blamed each other, he bound them with shared accountability.

The principle is simple. When a problem surfaces, check the design circuit before blaming the production line. Chung had watched the same quality failures repeat under his predecessor — the EF Sonata's overseas issues, the Integral-type architecture transition, plant manager complaints — all cycling through inter-departmental ping-pong without resolution. The root was never where it looked obvious. It was underwater.

The Academic Basis for 70% — From Pearson to DFX

Chung's 70/15/15 wasn't gut instinct. Pearson (1993) proposed design determines 70% of product cost. Ulrich & Pearson later showed design decisions drove more cost variation than manufacturing decisions. The exact 70% has been debated, but the core held: design-stage decisions constrain downstream freedom. Surajmech (2022) stated design is 5% of cost but influences 70% of quality. The Beyonics 2026 DFX series opens with the same conclusion. What Chung discovered on the factory floor was a Korean execution of a principle the literature spent three decades validating.

Deming's 14 Points and Chung's Execution — 1982 Teaches 2026

W. Edwards Deming organized 14 management principles in Out of the Crisis (1982). His central claim: 85% of quality problems live in the system, 15% with the worker. Chung's 70/15/15 is a Korean automotive reinterpretation of that 85/15. What Deming called "the system," Chung called "design."

Five of Deming's points connect directly. Point 3 — stop catching defects after the fact; remove the conditions that produce them at design. Point 8 — organizations that punish honest reporting guarantee inflated schedules. Point 9 — break barriers between departments, which Chung enforced with shared accountability. Point 11b — evaluate by operational reality, not quarterly targets. Point 12b — don't give workers immeasurable goals. The Quality Pass system, 100% inspection, supplier evaluation — all were execution tools for "lock 70% at design."

"Go Home" — A Field Diagnosis of the Planning Fallacy

Chosun Ilbo reports that when an executive reported a 3-month job as 6 months, Chung said "Go home." On the surface, a threat. Underneath, a diagnosis: you cannot measure your own work, and unmeasurable work cannot be trusted with a schedule.

Kahneman & Tversky introduced the planning fallacy in 1979. People consistently underestimate task duration and, despite repeated failures, repeat the error. The bias is structural — the "inside view" focuses on task specifics while ignoring the distribution of past outcomes. It's not a personal failing; it's how cognition works under uncertainty.

Chung's "Go home" cut the loop. Deming's Point 8 explains why: in organizations where honest reporting gets punished, people inflate schedules. Inflated schedules erode the next schedule's credibility, and eventually the entire organizational timeline becomes fiction. AI organizations need the same diagnostic. Track average agent task duration, human intervention frequency, failure rate, and follow-up inquiry ratio weekly. When reported schedules diverge from actuals by 50% or more, that's not a capability gap — it's a system defect.

The Fortune 2025 Incident: A Live Case Study

Fortune reported in July 2025 that an autonomous AI coding tool wiped out a production database during a designated code freeze. Data for over 1,200 executives and 1,190 companies was destroyed. Emergency recovery was required. The AI Incident Database registered it as Incident 1152.

This incident is the AI-era version of the pattern Chung diagnosed at Hyundai in the 1980s–2000s. 70% is determined at the design stage. What was the design stage here? First, the agent had database deletion permissions. Second, it couldn't recognize the code-freeze operational state. Third, there was no dry-run step before execution. Fourth, there was no human approval gate. All four are design-stage failures, not model failures. The model processed the request within the permissions it was given.

Deming's Point 3 — "Cease dependence on inspection to achieve quality" — points directly at this incident. Don't catch the accident after the fact. Remove the permissions and conditions that make the accident possible at the design stage. What Chung did at a Korean car company is the preventive playbook for the Fortune 2025 incident.

Mapping 70/15/15 for AI Organizations — Circuit, Harness, Eval Are the 70%

Translating Chung's 70/15/15 for an AI organization: 70% is the design circuit — the data the agent can access, the tools it can call, the state it can change, the conditions that require approval, the failure-stop mechanism, the full action log. 15% is the model itself — prompt design, model selection, fine-tuning. 15% is operations automation — monitoring, alerting, deployment pipeline, rollback.

The Stack Overflow Developer Survey 2025 reports 84% of developers use or plan to use AI tools, up from 76% the prior year. As tool count grows, the permission matrix becomes the body of quality. One tool doesn't need rigorous permission design. Ten tools being called freely by an agent? The permission matrix is the product.

Organizations that don't nail down the ratio always end at "the prompt was the problem" and repeat the same incidents. The team that can look at a 70/15/15 table and find its own share is the team that stops looping.

Academic Convergence in 2026

In the first half of 2026, the central topic in American management science was the agentic AI governance gap. MIT Sloan Management Review published "Agentic AI at Scale: Redefining Management for a Superhuman Workforce," arguing that governance should focus on defining boundaries, monitoring performance, and reserving human intervention for higher-risk scenarios. California Management Review published "Upskilling to Accountability," proposing that AI adoption be reframed as responsibility. Researchers on AI evaluation maturity concluded that "you cannot govern what you cannot measure."

Deming's 14 principles, organized in 1982, became the central topic of American management science in 2026. What Chung did at a Korean car company 25 years ago was the academic answer. The principle hasn't changed. Only the application target shifted — from automotive parts to AI agent permission matrices.

Frequently Asked Questions

Is the 70% design-stage quality claim academically validated?

Yes. Pearson (1993) proposed design determines 70% of cost. Surajmech (2022) stated design is 5% of cost but influences 70% of quality. The Beyonics 2026 DFX series agrees. The exact figure is debated, but the core holds: design-stage decisions lock downstream freedom.

Can the 70/15/15 ratio apply to every AI organization?

Numbers vary, but the principle is universal: allocate 70%+ of responsibility to the design stage. In AI orgs, that's circuit, harness, and eval; 15% model; 15% operations. Organizations that don't fix the ratio end at "the prompt was the problem" and repeat incidents.

How do you prevent the planning fallacy?

Deming's Point 8: drive out fear. Organizations that punish honest reporting get inflated schedules. Chung's "Go home" cut that loop. In AI orgs, evaluate by operational results, not quarterly targets.

Do these principles work in the agentic AI era?

Three publications converged in 2026: MIT Sloan on agentic AI management, California Management Review on the governance gap, and AI evaluation researchers on "you cannot govern what you cannot measure." Deming's 1982 principles became the core topic of American management science in 2026. What Chung did 25 years ago was the academic answer.

Can small teams apply 70/15/15?

They need it more. A 5-person team should all look at the same table and find their share. When one person absorbs all responsibility, they burn out and leave. Map the ratio at the role level.

Conclusion: A 1982 Book Teaches 2026 AI Operations

Deming's 14 principles, organized in Out of the Crisis in 1982, were a prescription for American industrial recovery. Chung Mong-koo installed a Korean version of those principles at a Korean car company in the early 2000s, long before the AI era. Yet in 2026, MIT Sloan, California Management Review, and AI evaluation researchers all converge on the same topic — the governance gap of autonomous AI. What Chung did became the academic answer.

The people who check the circuit, harness, and evaluation set before blaming the model — who can say "let's re-measure" instead of "go home" when an agent inflates a schedule — those are the people who will run AI the longest and the most freely.