NTT Data reported this week that smart factories are failing to deliver smart outcomes. I have stood on the floor of the exception, and the difference was not the sensors – the boring layer had come first. Somewhere between the control plans, the standard work and the first QRQC board, that plant had earned the right to a dashboard. Most plants skip the earning and go straight to the buying.
The trade press this month reads like a vendor catalogue: machine learning on plastics lines, AI vision tending CNC cells, inspection algorithms migrating from aerospace additive into coated wood and pharmaceutical validation. Every one of those stories assumes the process underneath the sensor is stable. Across two decades in automotive and aerospace I have watched that assumption fail four times out of five.
Digitised chaos is still chaos
I once watched a 55-inch dashboard refresh an out-of-control process every thirty seconds. Impressive hardware. The chart was a monument, not a control – a live-streamed confession that nobody on that floor could state what the process was supposed to do when it drifted. Visibility is not control.
Here is the mechanism that never makes the case study. When special-cause variation is everywhere, every alert is noise with a timestamp. Operators learn to acknowledge screens the way city dwellers learn to ignore car alarms. Within a quarter the dashboard is wallpaper, and the investment measures nothing except its own price tag.
The pattern followed me into consulting. At FOREAST I took lean engagements at ArcelorMittal and Meopta where clients arrived with dashboards already bought and flow still broken – an SPC system pristine to the third decimal while takt was a rumour, a steel line streaming parameters from every stand while work-in-progress piled up where the schedule said it should not exist. The data was beautiful. The plant was not.
Standard work is the operating system
You cannot detect deviation without a standard to deviate from. That is not a platitude; it is arithmetic. A sensor on an unstandardised process measures noise, and no algorithm turns noise into signal – it labels the noise faster.
A control plan tells the sensor stack three things it cannot work out for itself: which characteristic actually matters, what good looks like in numbers, and who moves when the numbers go wrong. Standard work gives deviation a baseline. The reaction plan gives the alert a destination. QRQC cadence – daily, at the line, containment inside 24 hours, root cause inside 72 – gives the loop a heartbeat.
I lived that sequence once, on purpose.
At SNOP I stood up the QA/QC function of a 900-plus employee greenfield plant from zero, and for the first year nobody priced a dashboard because we were busy writing control plans, standard work and reaction plans, and running QRQC at a fixed cadence. The 70% defect-cost reduction we later banked came from that discipline, not from data volume. When sensors finally arrived, every reading landed on a named reaction: this characteristic, this limit, this owner, this containment step. The instrumentation accelerated a control that already existed. It did not create one.
At WITTE Automotive the proof was the Q-Wall. Alerts are the cheap part; every plant has alerts. What made escalation work was that each alert arrived attached to a reaction plan with a name on it, and the wall – physical, public, embarrassing by design – made skipping the reaction more expensive than running it.
An alert without a reaction plan is a fire alarm in an empty building.
Buy the boring layer first
If you run a plant, weld this rule onto the capital approval process: stabilise, instrument, then automate the decision. In that order. Every time. Stabilising is cheap – a process engineer's time, printed standard work, a whiteboard, the discipline to hold a daily QRQC at the line. Instrumentation is the next order of magnitude up. Automating the decision is where the real money goes, and it only pays if the two layers beneath it are solid.
The wrong order is measurable in euros. I have watched a plant spend €300,000 on vision inspection to classify defects whose top three causes could have been removed for €8,000 of engineering time. I have seen a €120,000 OEE dashboard installed on a line whose changeovers varied by forty minutes depending on the shift running them. The dashboard reported those forty minutes beautifully. For years.
So interrogate vendors accordingly. Ask what happens after the alert fires: who receives it, what they are authorised to do, how you verify the loop closed. Ask how the system behaves when the process is unstable – a good answer involves aggregation, rate limits and a human reaction plan; a bad answer contains the word real-time three times and ends with a demo. Demand the boring integrations: the alert must write into the QRQC or 8D workflow, not die in a notification queue nobody owns. No vendor has lost a deal by promising real-time visibility. Plants lose money by buying it before they have something worth seeing.
Key takeaways
- Write the control plan before you price the instrumentation – characteristic, limits, reaction, owner. An unwritten reaction plan turns any sensor into decoration.
- Sequence capital as stabilise, instrument, automate the decision. The reverse order buys visibility of problems nobody is authorised or trained to remove.
- Put the reaction plan in the vendor specification. If the supplier cannot show who acts on an alert and how the loop closes, you are buying a notification, not a control.
- Judge digital quality systems by time-to-reaction, not data volume. Plants that win with AI inspection measure hours from deviation to containment, not dashboards deployed.
Connectivity is a multiplier, and multipliers are indifferent to sign. A plant with standard work, control plans and a living QRQC cadence compounds the return on every sensor; a plant without them compounds the noise, faster and in higher resolution. The NTT Data finding is not a verdict on the technology – the machine-learning and vision work in this week's press is real. It is a verdict on sequence. The plant is connected; the standard is not. Until the standard work is engineered as carefully as the network, the smart factory will keep streaming variation nobody can answer for. Build the operating system first. The apps are easy.