智能工厂并非完全无人化The Smart Factory Isn’t Fully Automatic

Tom, the board wants a smart-factory roadmap. One proposal is to automate the entire final-assembly area within eighteen months. Before I support it, I want to know whether we are solving the right problem.

That is the right starting point. Automation level is not a useful target by itself. We should define the operational problems, process stability and business outcomes before selecting technology.

The area handles twelve product variants, frequent engineering changes and several manual adjustments. Labour availability is also becoming less predictable.

That combination suggests selective automation rather than a single unmanned-line concept. Stable transport and repetitive checks may be good automation targets, while variable fitting and exception recovery may still need human judgement.

Some managers argue that AI can learn the technicians' decisions and remove that dependency. Is that realistic?

AI can reveal patterns and support decisions, but tacit knowledge is often incomplete, contextual and difficult to label. Experienced technicians must explain cues, compare cases and approve which knowledge becomes a controlled rule.

Let us discuss ROI. The automation proposal uses labour savings as the main benefit. Is that sufficient?

No. Include quality risk, downtime, changeover effort, investigation time, maintenance capability and the cost of handling variants. We also need a baseline and time horizon, otherwise the ROI number will look precise but mean very little.

Would you recommend a POC before the investment decision?

Yes, but it must answer a narrow uncertainty. For example, can vision and workflow signals identify a missed fastening step across representative variants? It should not pretend to prove full production readiness.

What would production deployment add after a successful POC?

Approved interfaces, line-side response logic, access control, monitoring, fallback procedures, training, maintenance ownership and acceptance under live conditions. Deployment is an operating system, not just a working model.

The board also expects scalability. How can we show that without promising a simple copy-and-paste rollout?

Separate the reusable core from site-specific adaptation. Model components, workflow templates, reporting rules and test methods may be reusable. Camera position, lighting, product definitions, interfaces and ownership still require local confirmation.

Could we structure the roadmap in stages?

Stage one maps the process and baseline. Stage two runs a focused POC. Stage three deploys one production cell with monitoring and fallback. Stage four scales only the validated template and keeps a site-readiness gate for each new line.

Where should operators participate in that roadmap?

From the beginning. They expose real variation, usability constraints and recovery practice. Their participation also helps explain that the system supports the process rather than silently rating individuals.

How should we choose between a conventional automation cell, a specialised vision model and a larger AI model?

Choose from the process requirement, not the technology label. Deterministic controls suit stable signals and safety logic. Specialised models suit well-defined visual targets with controlled latency. Larger models may assist semantic review, but their role still needs evidence, cost limits and an approved response path.

What about data governance when we scale across plants in different countries?

Define what remains local, what may be aggregated and who can approve access. Real-time inference can stay at the edge, while authorised summaries support management. Retention, transfer and remote support must follow each customer's policy and applicable requirements rather than one universal cloud assumption.

Then the recommendation is not less ambitious; it is more disciplined. We automate stable work, augment variable work and build evidence before scaling.

Exactly. The future factory is not defined by how few people it has. It is defined by how well people, machines and AI coordinate around a measurable process.