Quality isn't automated: Ford's model with 350 experts and AI

It all started with a disciplinary issue, not a lack of software

Ford’s vice president of engineering, Charles Poon, described the previous situation as a set of issues stemming from inconsistent discipline in problem-solving. He also cited a loss of expertise that had occurred three or four years earlier. This diagnosis is significant: the company did not present the problem simply as a lack of data or a flawed algorithm, but as an organizational weakness in the process by which failures are identified, examined, and resolved.

The solution was to bring the accumulated expertise back to the decision-making level. The 350 senior engineers—including some who had retired—took on the role of expert troubleshooters and consultants. They were not brought in merely as a token committee. Their mission was to help younger engineers review designs, identify weaknesses, and conduct the necessary technical analysis.

The red teams incorporated skepticism into the process

Ford compares these experts to red teams. They look at a component and ask how it might fail. They examine the choice of materials, the dimensions, and the reasoning behind minor design decisions. Poon gave the example of asking why a radius is two millimeters instead of two and a half. This detail shows that the review isn’t limited to a general «it looks right,» but requires documentation of the choice.

Five or six specialists may be involved in fuel systems, while the design is presented by a team of two or three design engineers. The reviewers are knowledgeable about tanks, fuel vapors, material interactions, injectors, pumps, software, and control strategies. This is critical for any business: meaningful reviews require people who understand the interdependencies, not just a checklist.

AI acts as a sensor for human judgment

The article makes it clear that tools complement engineers rather than replacing them. An iPhone-based system at the Kentucky Truck Plant checks wire connections and the installation of critical bolts while these points are still accessible for correction. If a bolt is found on the floor, the technician can photograph it, and the system compares it to the vehicle’s parts list to determine where it belongs.

The value here isn't the impressive interface. The same principle applies when a team designs workflow automation using artificial intelligence: The alert must reach the person who can take action in a timely manner. It involves shifting control to the moment and location where a correction can be made at the lowest cost. In e-commerce, the equivalent would be a tool that flags an unusual drop in checkout completion rates or incorrect product mapping before the problem escalates. The tool does not determine the business cause on its own; it simply provides the person in charge with a specific alert to investigate.

Hundreds of thousands of traces require both machines and people

According to Poon, hundreds of thousands of time series of pressure, force, velocity, and torque are collected in a single day during transmission functional tests. It is difficult for a person to reliably examine such a large volume of data. Machine learning looks for small deviations and alerts engineers to which specific component requires attention.

The example given is particularly instructive: a minor anomaly led to a disassembly, during which a contamination was found that had blocked a hydraulic passage and was causing slightly higher pressure. The AI did not «diagnose» the entire cause as if by magic. It flagged the outlier; the technical team followed up on the finding, opened up the system, and confirmed the physical cause.

From design to service, the same brand must be consistent

This approach is not limited to the production line. Poon discusses the application of tools during the design phase, the manufacturing phase, and the service phase. This continuity allows an organization to view quality as a cycle: feedback from the service phase can inform the design phase, a finding from the production phase can change quality control procedures, and a design choice can lead to a specific test.

For a digital business, this means linking data from ads, the website, checkout, customer support, and returns. The source does not claim that Ford uses this marketing model; it is a business analogy. The safe conclusion is that value increases when data points are not trapped in separate silos.

In-house technology meets specific needs

Ford uses both commercially available tools and its own solutions. Poon mentions in-house neural-network machine-learning algorithms for pattern recognition and predictability, generative AI for creating geometries and optimization, as well as agentic AI that supports peer reviews. He emphasizes that many applications require specialized training and development due to their unique nature.

This does not mean that every company must build its own model. It does, however, point to a selection criterion: the more closely an application is tied to proprietary data, critical processes, and domain knowledge, the more carefully its customization, evaluation, and human oversight must be designed. A general-purpose chatbot is not equivalent to a control system trained to detect specific anomalies.

The cultural shift was part of the technical system

One of the source’s strongest observations concerns how the company handles failures. Poon says that the company now celebrates the discovery of failures and the early detection of anomalies, whereas in the past such a finding might have been perceived as a weakness. Without this change, even the best monitoring can fail: people will be motivated to hide a signal or downplay it.

Psychological and organizational safety are not referred to here as abstract values. They are linked to specific behavior: identifying the problem early, reporting it, reviewing it with experts, and correcting it. For marketers and business owners, the corresponding practice is not to label every negative outcome as «bad luck,» but to use it as a starting point for reviewing the campaign, the message, the funnel, or the customer experience.

The authority of reviewers is just as important as their knowledge

Technical specialists hold senior leadership positions and, according to the article, have the authority to take the necessary actions. This detail distinguishes a true governance model from a consultative workshop. If a reviewer can identify a risk but cannot require an analysis or a change, their insight has no practical impact.

As the Failures of Operational AI Automation Systems, an effective AI workflow requires clearly defined roles: who reviews the alert, who verifies the cause, who approves the change, and who monitors whether the correction was effective. The source does not provide a general organizational chart for other companies, but its model highlights the relationship between expertise, authority, and execution.

What we can—and cannot—conclude from the ranking

Design News notes that one might disagree about exactly what the Initial Quality Study measures. At the same time, it describes Ford’s rise from the bottom of the rankings to the top within just a few years as a genuine achievement and notes that the company itself attributes the improvement to its experienced engineers and tools. The distinction is crucial: we have a documented ranking and a corporate interpretation, not experimental proof that every improvement was caused by a specific factor.

That is why business analysis must remain cautious. We do not simply copy a 350-person structure just because it worked on a different scale. We stick to verifiable principles: drawing on experience, independent scrutiny, tools for large volumes of data, human verification, and a culture that rewards early discovery.

A Practical Framework for Marketing and E-Commerce Teams

A team can start by identifying three things: the points where a mistake becomes costly, the people who recognize the early warning signs, and the data that is currently too voluminous to check manually. Then, it can conduct a small red-team review before critical launches, involving people from performance, content, analytics, customer service, and operations where relevant.

AI can be used to identify outliers, inconsistencies, and recurring patterns, but every alert requires an owner and a verification process. This is the same criterion required by the Evaluation of AI agents based on actual deliverables: Speed without documented quality is not enough. The team must also document what was found, what the actual cause was, and what changes were made as a result. This way, the experience isn’t confined to one person’s mind, and the evaluation model or rule can be improved based on actual events.

Finally, the KPI shouldn’t just be «how many alerts the AI generated.» More meaningful questions are whether problems were detected earlier, whether their recurrence has been reduced, and whether reviewers can explain why an intervention was made. These are recommended implementation principles derived from the case study, not metrics published by Ford.

The competitive advantage lies in the composition

The Ford case does not support the «human or AI» dilemma. It supports a complex system: experienced people formulate difficult questions, junior engineers document decisions, sensors and tests generate data, machine learning highlights anomalies, and teams confirm the physical cause. Generative and agentic AI are added as additional layers of support, not as a substitute for accountability.

For any business seeking results from AI, the message is straightforward: don’t start with the tool. Start with the point of failure, the knowledge needed to identify it, and the process that enables you to take action. Then technology can reveal what humans can’t see in time—and humans can understand what it means.