Three Steps to Smart Factory: From Rule Checks to Autonomous Agent Loops

The article frames manufacturing AI adoption as three verifiable steps — rule-based inspection , visual inspection , and agent closed-loop — and argues most factories never move beyond the second layer.

1. The Data Disconnect

Manufacturing data lives in three isolated silos:

Inside equipment : PLCs, sensors, CNC systems generate signals but machines are often not networked; data stays on local controllers.

On paper forms : Work orders, quality records, equipment checklists remain on paper or scattered Excel files — unstructured, no timestamps, hard to trace.

In veterans' heads : Welding current settings, abnormal sounds that signal faults — tacit knowledge that leaves when people retire.

Result: companies have data but lack decision-ready data . Dashboards answer "how are we doing now?" but not "what do we do next?" — a dashboard is reporting, not intelligence.

2. Three Maturity Layers

Each layer uses a different tech stack and has a distinct capability ceiling.

Layer 1: Rule-Based Inspection

Core tech : Thresholds / rule engines — humans write rules, machines execute.

Boundary : Can "block" but cannot "see". Coverage is limited; explainable but high false-positive rate, relies on human fallback.

Typical data : High misjudgment, depends on manual re-check.

Layer 2: Visual Inspection (AI Vision)

Core tech : Computer vision, large models / industrial inspection models.

Boundary : Can "see" but cannot "judge". Detects defects but does not write back or adjust parameters.

Cases : Deli Group pen line — 1.4 seconds per 8 pens, 15 defect types, 42× efficiency gain , defect rate 3.2% → 0.5% , headcount 1000+ → 100+ . Huaxiang flange factory — 12× efficiency for 100% online inspection of hundreds of part types.

Layer 3: Agent Closed-Loop

Core tech : Multi-agent systems — detection → root-cause → parameter adjustment → re-inspection.

Boundary : Can "fix", self-executing. Humans intervene only on low-confidence actions.

Cases :

PetroChina Kunlun Digital + Alibaba Cloud: chemical CDU unit, 60 days zero human intervention , root-cause accuracy 90%+ , troubleshooting workload -70% .

Jinko Solar slicing workshop: breakage rate -3.3 percentage points , annual benefit ~¥11.48M .

Voyah Auto supply chain: 6000+ suppliers connected, risk warning lead time 30 → 52 days , false alarms -18% .

Foxconn MoMClaw: root-cause analysis 80% faster , productivity +15% , failure rate -10% .

Midea factory collaborative scheduling: 14 collaborative agents, changeover time -50% .

Industry signal: industrial agent deployment spans R&D design (32%), operations & maintenance (25%), supply chain (19%) — value extends far beyond the shop floor. Dingjie Digital's "Athena" runtime embeds 160+ agents covering 100 high-frequency scenarios, achieving "human sets goal, agents execute autonomously".

3. The "AI+Industry Five Questions" Framework

Applied to every case in the 20-part series; at least three answered per episode:

① Data source : Equipment sensors + MES + ERP. Build data pipes before models; otherwise agents are blind executors.

② Decision authority : Who approves a false-pass? Irreversible actions (line stop, parameter change) must keep human sign-off; agents only "suggest + execute reversible actions".

③ Accountability : If AI passes a defective part, who owns it — algorithm, ops, or the sign-off person? Must be written into SOP.

④ Metric unit : Defect rate, unplanned downtime hours, OEE — not isolated "model accuracy".

⑤ Boundary : Vision inspection auto-judged; parameter changes & line stops need human review. Agents governed as "suppliers" — permission boundaries, audit logs, kill switches.

4. Three Reality Checks

Separate "algorithm ceiling" from "steady-state production" . Deli's "99.9% space utilization" or "99.98% full-inspection accuracy" are peak lab numbers. Real lines face continuous operation, lighting shifts, material variance — steady-state data often discounts sharply. Always ask: lab or floor? peak or average?

"Cut 900 heads" rarely appears in press releases . Efficiency gains are trumpeted; headcount reduction is muted. The remaining 100 aren't "optimized away" — their role shifts from "watch" to "judge", from repetitive labor to final exception arbitration. Organizational change cost is seldom documented.

Data collection is the real bottleneck . Research shows <4% of industrial data is high-quality enough for AI. A 95% lab model can collapse under dust, vibration, EMI. Before chasing models, ask: is this machine even connected?

5. Self-Checklist for Manufacturing AI Projects

Can data be pulled at the decision moment? If not, no algorithm helps. Machine connectivity, structured work orders, codified experience are 0→1 prerequisites — higher priority than model selection.

Build a closed-loop, not a dashboard. Dashboards tell "now"; loops tell "next". Run one slice of "detect → root-cause → adjust → re-check" end-to-end — more valuable than a giant screen.

Calculate the value ledger, not the accuracy ledger. Deli's ¥11.48M, PetroChina's 60 zero-touch days are outcome metrics; accuracy is a process metric, not the goal.

Manufacturing is the hardest and most tangible "AI+Industry" track. It doesn't fly on hype — it lands on yield, uptime hours, and hard cash saved.