"We want to use AI on the shop floor" — asked with rising frequency across 2024-2025.

But many enterprises conflate industrial AI and office AI, and only discover mid-project that shop-floor constraints are an entirely different game.

This piece lays out industrial AI's real boundaries — which scenarios fit, which don't, what baseline conditions deployment needs, and the real pilot-to-production rhythm.

1. Industrial AI vs office AI: five essential differences

Difference 1: Data source

  • Office AI: text, tables, PDFs, images — stable, easy to parse
  • Industrial AI: sensor signals (vibration, temperature, current, pressure, sound, vision) — unstable sources, broken sensors, loose wires, signal drift

Industrial AI must assume 30-50% of data is dirty or missing from day one. Not a model-accuracy issue — a physical fact of the floor.

Difference 2: Deployment environment

  • Office AI: servers or cloud — controlled, stable network
  • Industrial AI: production-line edge devices — harsh (temperature -10 to 60°C, dust, vibration, EMI), frequent network drops

Industrial AI devices must continue operating without network (even in degraded mode). Cloud inference is infeasible for many industrial scenarios.

Difference 3: Response time requirements

  • Office AI: users tolerate a few seconds
  • Industrial AI: many scenarios need millisecond response (visual QC, control loops, safety interlocks)

Millisecond response rules out "signal → cloud → inference → return". Must infer at the edge.

Difference 4: Reliability requirements

  • Office AI: 1 hour of downtime usually isn't catastrophic
  • Industrial AI: 1 minute of downtime may mean production halt, millions in loss, safety incident

Industrial AI's reliability requirement is 2-3 orders of magnitude higher. SLAs typically start at 99.9%.

Difference 5: Compliance and safety

  • Office AI: mostly data privacy compliance
  • Industrial AI: plus Functional Safety. AI decisions cannot cause injury or equipment damage

In regulated industries (pharma, chemicals, nuclear), AI entering production decisions requires functional safety certification — a specialty domain with its own standards (IEC 61508 / 61511).

2. Four scenarios that fit industrial AI

Not all scenarios are ready today. Ordered by ROI and technical maturity:

Scenario 1: Visual quality inspection (most mature, top Phase-1 recommendation)

Description: cameras + image AI replace human visual inspection of product appearance / dimensions / defects.

Fits:

  • Electronics (PCB solder joints, mounts, surface)
  • Auto parts (weld seams, castings, stampings)
  • Packaging (label position, seal, integrity)
  • Agriculture (grading, foreign-object removal)
  • Textiles (defects, color variance)

Typical outcomes:

  • Miss rate drops from 0.5-1% down to under 0.1%
  • Inspection speed 3-10× faster
  • QC headcount reduced 50-70%

Difficulties:

  • Sample collection — thousands to tens of thousands of defect samples needed (including edge cases)
  • Environmental control — lighting, angle, background must be relatively stable
  • Miss vs false-positive trade-off — industries tolerate these differently (pharma accepts more false positives, cannot accept misses)

Typical budget: ¥200-800k per line (camera + edge compute + custom model + deployment).

Scenario 2: Predictive maintenance (fits rotating-equipment sites)

Description: predict equipment failures from vibration, temperature, current — schedule maintenance proactively.

Fits:

  • Motors, pumps, fans (classic)
  • Gearboxes
  • Compressors
  • Large machine tool spindles

Typical outcomes:

  • Unplanned downtime reduced 30-50%
  • Maintenance cost reduced 20-40%
  • 72-hour advance warning (under good data conditions)

Difficulties:

  • Rare failure samples — a good piece of equipment may not fail all year, sample accumulation is slow
  • Ongoing sensor maintenance (vibration sensors degrade with contamination)
  • Models need continuous recalibration with equipment aging

Typical budget: ¥500k-1.5M per workshop.

Scenario 3: Energy optimization (for chemicals, metallurgy, other high-energy industries)

Description: AI adjusts energy usage in real time based on production parameters, reducing per-unit energy consumption.

Fits:

  • Cement (kiln, grinder consumption)
  • Steel (furnace, rolling mill)
  • Chemicals (reactors, distillation columns)
  • Data centers (cooling, PUE optimization)

Typical outcomes:

  • 3-8% reduction in per-unit energy consumption
  • In energy-heavy industries, absolute savings are large (millions to tens of millions annually)

Difficulties:

  • Large historical process datasets required
  • Must develop with process engineers (AI cannot violate chemistry/physics)
  • Online-adjustment risk — wrong adjustment can affect product quality

Typical budget: ¥1-3M.

Scenario 4: Production scheduling optimization (discrete / flexible manufacturing)

Description: AI generates optimal schedules from orders, inventory, equipment state, and delivery constraints.

Fits:

  • Multi-SKU small-batch
  • Frequent order changes
  • Complex equipment / process / material constraints

Typical outcomes:

  • Equipment utilization up 10-15%
  • On-time delivery 70-80% → 90-95%
  • Scheduling headcount reduced 50-70%

Difficulties:

  • Many constraints, complex modeling
  • Business judgment required (what products can delay, which cannot)
  • Edge cases (rush orders, equipment failure)

Typical budget: ¥800k-2M.

3. Three scenarios that don't fit

Scenario 1: Pure process industry with stable processes (e.g., mature refining)

Decades of DCS / APC (Advanced Process Control) systems make AI's precision gains marginal. "AI + traditional control" may gain 1-3% — high investment, slow return.

Scenario 2: No production data digitization

If key data is still on paper or only in PLC without export, do data acquisition and digitization first, then AI. Industrial AI's data foundation demands are an order of magnitude higher than office AI's.

Scenario 3: Very few SKUs, very large batch (bulk raw materials)

Processes have been optimized near physical limits. AI finds little meaningful improvement.

4. Six hard constraints for industrial AI deployment

Before launching a project, evaluate these six clearly:

Constraint 1: Network

  • Stability of plant internal network?
  • OT network (industrial) and IT network (office) physically isolated?
  • Bandwidth sufficient from edge to central server?
  • Can AI continue working when network drops?

Constraint 2: Hardware

  • Environmental adaptability of edge devices (temperature, dust, vibration)?
  • GPU / NPU / FPGA selection (affects inference speed and power)?
  • Supply lead time (domestic cards 1-2 months, imported 3-6 months)
  • Redundancy and failover? (Mandatory in industrial)

Constraint 3: Data sources

  • Are key sensors already installed?
  • Is sampling rate enough (millisecond scenarios need 1kHz+)?
  • Historical data time span and quality?
  • Data formats, protocols, interfaces unified?

Constraint 4: Safety certification

  • Does this scenario need functional safety certification?
  • For pharma / nuclear / chemicals, which industry-specific certifications (GMP, GAMP, NRC)?
  • Network security (IEC 62443) requirements?

Constraint 5: Process knowledge

  • Are process engineers willing to engage deeply? (AI cannot operate without process understanding)
  • Documented process materials? (Or all in veteran operators' heads)
  • Abnormal-condition handling process?

Constraint 6: Operations capability

  • Factory has people who understand AI system ops? (At least 1)
  • Retraining mechanism as models age?
  • Sensor replacement / calibration process?

If 2-3 of these are seriously unmet, project risk is very high. Close gaps first, then launch.

5. The real pilot-to-production path

Standard industrial AI project path:

Stage 1: Feasibility (2-4 weeks)

  • On-site survey; evaluate the six constraints
  • Collect sample data; build a quick prototype
  • Evaluate expected technical precision / speed

Output: "feasible / infeasible / needs precursor work". Kill infeasible projects; don't force them.

Stage 2: Pilot deployment (2-4 months)

  • Deploy on one line (or one representative piece of equipment)
  • End-to-end wiring: sensors → edge inference → result → feedback
  • Run 30-90 days to gather real-world data
  • Continuous tuning

Pilot conditions must be close to production — no "curated pilot".

Stage 3: Replication and scaling (3-12 months)

  • Extend to other lines / workshops / sites
  • Each new site gets calibrated separately (industrial equipment has large individual variation)
  • Establish centralized monitoring + distributed ops

Stage 4: Ongoing operations (permanent)

  • Model aging monitoring + periodic retraining
  • Sensor health monitoring
  • Adaptation to business changes (new products, new processes)

Overall timeline: from kickoff to first-line production typically 6-9 months; full-plant scale-up with stable operation typically 1.5-2 years. Don't expect "3 months for an industrial AI project" — unrealistic.

6. Closing

Industrial AI is the highest-barrier, longest-cycle, highest-value direction in enterprise AI.

The difficulty isn't the algorithms — vision AI, time-series AI, optimization algorithms are all mature today. The difficulty is shop-floor data, network, hardware, safety, and process constraints. These only reveal themselves on-site.

Our PLC + AI Industrial Intelligence service line targets this direction — engineers on-site, hands-on commissioning, sustained operations. If your plant is evaluating industrial AI, start with a feasibility diagnosis — we'll tell you which scenarios your current conditions support, which need precursor work, and which we don't recommend yet.

Industrial AI isn't about installing a model. It's a complete on-site capability build.

Think it through. Then build.