AI Comes Last. Most Factories Put It First.

Insights from the Scaling Business Summit 2026, Ho Chi Minh City.

Every manufacturing company is somewhere on the AI journey right now. Most are making the same mistake: starting with the technology and working backward to the problem. Nguyen Hong Phuc, known as Anthony, Director of Digitalization Strategy and Business Development in Operations at Schaeffler Asia Pacific, walked onto the SBS stage with a different sequence.

1. Ask Why Before You Ask What

The first mistake most manufacturers make when approaching AI is skipping the three questions that should precede any investment. Anthony framed them simply:

• Why do we need this?
• How will we implement it?
• What do we expect to achieve?

The why is not "because our competitors are doing it" or "because the CEO asked about AI at the last board meeting." A serious why connects to a specific operational problem, a bottleneck, a quality failure rate, a labor cost trajectory that AI or automation is structurally capable of solving. Without it, the investment is theater.

The how requires a comprehensive vision and strategy before a single line item is approved. That means understanding the current state of your data infrastructure, your workforce readiness, and your integration architecture. "How many people know how the future will look?" Anthony asked the room. The honest answer is nobody. But a good strategy does not require certainty about the destination. It requires clarity about the next step and the capability to adapt.

The what defines success before deployment, not after. OEE improvement, reduction in unplanned downtime, headcount redeployment, these need to be quantified in advance. Without a target, there is no basis for evaluating whether the investment worked.

Lesson 1: Before any AI investment, answer three questions in order: why does the business need this, how will we implement it well, and what specifically will we achieve. Skipping to the tool is how companies waste capital.

2. You Cannot Automate a Non-Standard Process

This was the sharpest insight of the session, and the one most likely to be ignored.

Anthony was direct: "Before you buy the new machine, you need to ask yourself: do you have the standard process or not?"

The tension he named is real. Industrial engineering teams push for standardization. Lean teams push for continuous optimization. These two objectives are in genuine conflict: optimization means changing the standard, which means the standard is never stable enough to serve as the foundation for automation. Most factories are caught between them, optimizing processes that are not yet standardized, then trying to automate processes that are still changing.

The correct sequence is not complicated, but it requires discipline. First, establish the standard process. Then identify which steps within that standard can be replaced by automation. Then install the connectivity layer to integrate those solutions. Then train the people who will operate and maintain the new system. Only at the end, when the standard is stable, the automation is connected, and the team is capable, does it make sense to ask whether AI can further optimize the system.

Skipping any step in that sequence is what produces expensive installations that underperform. A cobot on an unstandardized line does not improve efficiency. It automates chaos.

Lesson 2: Automation applied to a non-standard process makes the mess faster. Standardize the process first. Identify what the automation replaces. Build the foundation before the AI layer.

3. Data Quality Is the Bottleneck. Not the Algorithm.

Anthony put this plainly: "Garbage in, garbage out. If you don't have good data, you cannot have anything."

This is consistently underestimated because the AI conversation tends to focus on the intelligence layer, the model, the algorithm, the platform. But the intelligence layer is only as useful as the data it runs on. Accurate, consistent, uniform data takes time to build and trained people to maintain. It does not arrive automatically when you install a new system.

Before evaluating any AI-powered tool demand forecasting, predictive maintenance, quality inspection, a manufacturer needs to audit the quality of the data those tools will consume. If the underlying operational data is incomplete, inconsistent across systems, or simply not being captured at the point of action, no algorithm will compensate for it.

There is also a verification responsibility that never goes away. A large language model can generate outputs that feel reasonable, but a senior expert still needs to review and confirm them. "You are the investor. You are the person who verifies if GPT works or not." The human role does not disappear when AI is introduced. It shifts from doing to verifying. That shift requires a different kind of expertise, not less of it.

Lesson 3: Audit your data before you evaluate any AI tool. The algorithm is not the constraint. The quality and consistency of what you feed it is.

4. The Cobot Math Nobody Tells You

When a manufacturer installs a cobot to replace an operator position, the immediate financial logic looks clean: lower labor cost, higher throughput, better OEE. Anthony walked through why the real math is more complicated.

Every automated system added to a line creates a corresponding maintenance requirement. Somebody has to service the hardware, manage software updates, and troubleshoot failures. That person usually does not already exist in the workforce. You either hire a maintenance technician, a new ongoing cost or you retrain an existing operator, which is the better workforce outcome but still requires investment and time.

The net financial benefit of automation is therefore not the cost of the replaced operator. It is that saving minus the cost of the new or retrained maintenance role, minus transition downtime, minus system integration work. Anthony's framing was honest: "That is another financial aspect you need to think about before you spend the money."

The ROI timeline for automation is long and non-linear. Manufacturers who expect a quick payback are working from the wrong model. The compounding benefits come later and only if the earlier steps were done correctly.

Lesson 4: Model the full cost of automation before approving the investment. Factor in the maintenance roles created, transition downtime, and integration work. The real ROI timeline is longer than the vendor will tell you.

The CEO Execution Playbook: What to Do Tomorrow

1. 1. Run the why-how-what test on every AI proposal you receive. Before approving any new AI or automation investment, require the team to answer all three questions clearly: why is this needed, how will it be implemented well, and what specifically will be achieved and measured?
2. 2.Audit your process standardization before your next automation purchase. Map one production line end to end. Identify which steps have a clear, documented standard. Those are candidates for automation. The ones without a standard need the standard first.
3. 3. Commission a data quality audit before evaluating any AI platform. Map where your operational data is generated, how it is captured, and where it becomes inconsistent or incomplete. Fix those gaps before signing a software contract.
4. 4. Build the maintenance cost into every automation business case. Require any proposal to include the cost of the maintenance role it creates, not just the operator cost it replaces. Approve only the cases where the net math holds.

→ Watch the full session on YouTube