Ford Brings Back Workers, Oracle Cuts 20,000—What Is the True Cost of the AI Replacement Line?

Ford has announced that it will “bring engineering back to humans” after previously relying on AI. Around the same time, Oracle revealed plans to cut approximately 21,000 employees. The reason for both decisions is the same: “Is it cost-effective?”

While the judgments of these two companies may seem opposite, they are actually answering the same question: “Is it cheaper to let AI handle this work, or is it more expensive?”—that’s all there is to it.

This question is not limited to large corporations. It applies equally to a small factory with ten employees or a five-person accounting firm, presenting the same structure of decision-making.

—

The Real Reason Ford “Brought Back Humans”

In recent years, Ford utilized AI in parts of its vehicle development process, leading to a reduction in engineering staff. What was the outcome? Quality declined, recall incidents increased, and rework costs ballooned.

Ford executives have openly acknowledged this: “We were wrong to think that introducing AI would result in high-quality products.” They have since shifted their policy to rehire experienced engineers.

It is important to note that Ford is not saying, “AI is bad.” What they are indicating is that “using AI for this task resulted in quality loss costs that outweighed the savings from reduced labor costs.” In other words, it’s purely a result of cost calculations.

Automotive engineering is a mass of tacit knowledge that cannot be captured in data. “This part looks fine on paper, but it could crack due to vibrations after three years on the production line”—such judgments can only come from the experience of seasoned professionals. AI requires data to learn, but there is no data for “issues that have not yet occurred.”

The areas where AI struggles are clear: “Limited data,” “Ambiguous correct answers,” and “Massive failure costs”—jobs that overlap with these three factors are still cheaper for humans.

—

The Structure Behind Oracle’s 21,000 Cuts

On the other hand, Oracle announced plans to cut around 21,000 employees, representing about 12-14% of its total workforce.

The cuts primarily target customer support, internal administrative tasks, and routine development work. In short, these are jobs that are “procedurally defined,” “have clear correct answers,” and “would not result in catastrophic failures if mistakes occur.”

Such tasks are where AI excels. For customer inquiries, LLM-based chatbots can handle about 80% of the workload. Internal expense reimbursements and approval processes can be automated. Data suggests that productivity for generating routine code can increase two to three times with tools like Copilot.

If we assume an average annual salary of $100,000 (approximately 15 million yen) for the positions being cut, this results in an annual labor cost reduction of $2.1 billion (approximately 315 billion yen). Even if the costs for implementing and operating AI tools are one-tenth of that, it still leads to an annual cost saving of about 280 billion yen. From a management perspective, this appears to be a highly rational decision.

However, there are pitfalls here as well.

—

How Long Will “AI Is Cheaper” Last?

As of 2025, the costs of using major LLM APIs continue to decline. Even for models like GPT-4, the cost per inquiry is just a few yen to several dozen yen. This is why the decision to “cut human workers and replace them with AI” is currently viable.

However, one thing to consider is: Will the operational costs of AI continue to decrease?

Electricity costs are rising. Demand for GPUs is tight. The computational resources needed for model training and inference are increasing exponentially. Even OpenAI has not yet turned a profit. At some point, the prices for APIs could reverse and start to rise.

In other words, the decision to “cut human workers now because AI is cheap” is only valid if the current cost structure remains unchanged. For robust companies like Oracle, even if costs rise, they can absorb them with their own infrastructure. But small and medium-sized enterprises cannot do the same.

—

How Should Small and Medium-Sized Enterprises Make Decisions?

The decision-making framework that can be drawn from the cases of Ford and Oracle is actually quite simple.

1. Draw the Line with “Failure Costs”

When AI makes a mistake in that job, how much loss will occur?

• If AI makes a mistake in creating internal meeting minutes, the loss is almost zero. → It’s safe to delegate.
• If AI miscalculates an estimate for a customer, it becomes a credibility issue. → Human oversight is essential.
• If AI overlooks something in product safety design, it could lead to recalls costing millions. → A human should handle this.

“How much will it cost if AI makes a mistake?” This figure becomes the basis for judgment.

2. Compare Using “Hourly Rates”

Convert the current costs of tasks assigned to AI into hourly rates. Next, calculate the costs of performing the same tasks with AI (API costs + human oversight costs) in hourly rates.

For example, if accounting data entry takes 40 hours a month and the part-time employee’s hourly wage is 1,200 yen, that totals 48,000 yen per month. If the AI tool costs 5,000 yen per month plus oversight (5 hours × 1,200 yen = 6,000 yen), that totals 11,000 yen per month. This results in a difference of 37,000 yen, leading to an annual saving of about 440,000 yen.

Such calculations should be done diligently for each task. It may not be flashy, but it is the most reliable method.

3. Decide Based on “Whether It Can Be Reverted”

The essence of Ford’s failure lies in the point that “cutting human workers incurs time and costs to bring them back.” Experienced engineers do not easily return once they leave. Their know-how disappears along with them.

For small and medium-sized enterprises, this is even more critical. “If this person leaves, no one else can do this job”—replacing such work with AI is too risky.

Conversely, tasks that can be standardized and easily handed over can be delegated to AI, and if it doesn’t work out, they can be reverted back to humans. “Test with jobs that can be reverted.” This is the golden rule.

—

Start Small. Make Decisions Based on Numbers. Ensure Reversibility.

Large corporations can cut or bring back thousands of employees. Small and medium-sized enterprises lack that capacity. Therefore, the only option is to enhance the precision of decision-making.

There are just three things to do:

1. Decide on one task to delegate to AI on Monday. It can be something small, like meeting minutes, drafting emails, or organizing data.
2. Review the numbers on Friday. How many hours were saved? What was the quality like? Were there any mistakes?
3. If it works well, expand it. If not, revert it.

Ford learned the hard way, spending hundreds of billions of yen to find out that “AI didn’t work.” Oracle took a gamble by cutting 20,000 employees, believing, “This will work.” Small and medium-sized enterprises can learn both lessons for free.

The boundary line is not something to be debated in a conference room; it should be tested in the field and drawn based on numbers.

Next Monday, start with just one task. That’s all it takes.