AI in Manufacturing: What Works on the Shop Floor

Predictive Maintenance

Service a machine on evidence, not on a calendar drawn up by a stranger.

Machines rarely fail without warning. First they behave a little differently: they run hotter, they vibrate at a frequency they did not use before, they draw more current for the same work. Your best operator hears some of this. That is why the man who has run a machine for fifteen years can tell you it is about to give trouble. He is matching a pattern he learned over years. He also retires one day, and the signal leaves with him.

A model does the same job from recorded data. It learns what each machine looks like when healthy, then watches for drift away from that baseline. It measures distance from normal rather than waiting for a fixed alarm threshold to break. A slow change that never crosses any single threshold is exactly the kind that catches you out. The output is a window, not a date. 'This bearing is likely to fail in two to five weeks' is something a planner can act on. False precision is not.

Machines do not stop breaking. What changes is that failures move from unplanned to planned, and a planned stop costs a fraction of an unplanned one. You service at the weekend with the right part in stores, instead of halting a shift while a driver hunts across town for one.

Visual Quality Inspection

A person inspects well in the first hour, and far less well by the eighth.

The defects that reach your customer are rarely the ones no one could see. They are the ones no one was looking at: the last hour of a night shift, or the day the line ran late and the visual check quietly got faster. If your defect rate differs by shift and no one has explained why, you know this already.

A camera system compares each part against what a good part looks like. It is not a simple image match. Parts vary for honest reasons: position, lighting, finish. A system that flagged all of that would be switched off in a week, and a high false alarm rate kills a project faster than a missed defect does. So the model learns which variation matters. It does not beat a person on eyesight. It beats a person by judging a borderline part the same way at 3am as at 9am.

People underrate the record it builds: each part inspected, and each defect classed by type and tagged to the machine, tool and shift that made it. In year one, the catching pays for the system. In year two, that Pareto lets engineering fix causes instead of sorting symptoms.

Demand Forecasting

Plan output on evidence, not on whoever argued hardest in the Monday meeting.

Most plants forecast by taking last year's number and adding a percentage that reflects the mood in the room. This works well enough for stable, high-volume lines, and badly for the rest. That is why safety stock exists, and why it grows quietly.

A statistical forecast reads the same records. It then splits out what is driving them: the trend, the season, the effect of a price change. It picks up that one customer who always orders heavily in the fortnight before their own year end. You get a range with a confidence band, not one number. That sounds like hedging, and it is the useful part. 'Between 8,000 and 11,000, most likely 9,400' tells your planner how much cover to hold. '9,400' on its own says nothing about risk.

This is often the cheapest of the eight to try, since the data already sits in your ERP in a tidy form. It is also the one where a bad build is hardest to spot, because a forecast is only proved wrong months later.

Production Planning & Scheduling

Sequence jobs across machines and shifts, without a spreadsheet only one person can read.

Scheduling looks like an operations job, but it is really a constraint problem. Jobs have due dates and machines have limited capacity. A changeover costs time, and how much depends on what ran before it. Some steps must come before others, and skilled operators are on some shifts and not others. A human planner holds most of this in their head. That is impressive, and it is a real business risk the week they take leave.

An optimiser searches the possible sequences for one that meets the hard constraints and scores well on what you care about. That may be throughput, on-time delivery, less changeover time, or a weighting across all three. The weighting is a commercial call, and you should make it, not the person writing the model.

The honest caveat: this works only if your routing and changeover data are correct. Suppose the system thinks a changeover takes twenty minutes and it really takes ninety. The schedule will be beautiful and useless. Cleaning that data is usually the bulk of the project, and a vendor who does not raise it plans to find it out at your cost.

Smart Inventory Management

Reorder points set five years ago are still running your working capital today.

Most plants set a reorder level once, on a lead time that was true at the time, and then no one revisits it. Demand shifts. A supplier gets slower. A new variant arrives. The number stays put. The result is easy to predict: some lines carry months of cover while others run out, and both stay hidden until the next stock count.

Modelling stock well means treating two things as variable rather than fixed: how much you will need, and how long the supplier will really take. A vendor with a steady twenty-day lead time needs far less cover than one who averages fifteen days but sometimes takes forty. A fixed reorder point cannot say that, so it over-stocks the steady vendor and under-stocks the erratic one.

Yield Optimisation

Find which process parameters really drive yield, not the ones people believe do.

Each plant has its folklore. Run it hotter on humid days. This operator gets better output on that machine. Some of it is real and hard-won, and some of it is chance that hardened into a rule years ago and was never tested.

With enough process data, logged often enough, you can tell the two apart. The model looks across temperature, pressure, speed, material lot, ambient conditions and operator, and shows which combinations move yield and which are noise. Often the finding is a link no one was hunting for: this material lot is fine, except when it runs above a certain speed.

The limit is sampling rate. A reading taken once per shift will not find much that a supervisor had not already guessed. If your process data is that thin, better logging is the project, and it is worth doing on its own.

Energy Consumption Analytics

The easiest business case to prove, since the saving lands on a bill you already get.

Energy is billed at plant level and managed at plant level, so no one can say what a machine, a shift or a product costs to run. Attribution fixes that: you break usage down by asset and by job. The cost of a product then includes the energy that went into it, not an averaged overhead.

The second half is forecasting load to manage peak-tariff exposure. Your tariff may penalise demand above a threshold, and it is worth real money to know in advance that three heavy loads are about to coincide. You stagger them. This is arithmetic more than intelligence, but it needs a forecast to work.

We often suggest this first for a plant that wants a provable win before it commits to more. Sub-metering is cheap and the data is clean. The result shows up on next quarter's bill, where the finance director can see it.

Worker Safety Monitoring

The simplest of the eight to build, and the one most likely to fail on people.

Spotting missing PPE, entry to a restricted zone, or a worker too close to moving machinery is well-solved computer vision. Most plants have the cameras already, so that part is not hard.

What decides whether it survives is how you frame it. Run it as surveillance of individuals, with named alerts and discipline to follow, and the floor will resent it. Staff will work around it, and one day it gets unplugged, often with a supervisor's help. Run it as monitoring of conditions instead. Alert a zone, not a person, and report in aggregate. Then the shop floor accepts it, and at times welcomes it.

We raise this before the technical scope, since this is the real risk. There is a legal side too: India's DPDP Act covers workplace monitoring and notice to staff. Take advice on it rather than assuming.

Downtime Response Agent

A machine stops at 2am, and the case file is written before the engineer arrives.

Today a stoppage starts a scramble. One person checks when it was last serviced. Another tries to recall whether this happened before. A third hunts for the right page of a manual that lives as a PDF on a laptop in the office. Twenty to forty minutes go by before real diagnosis starts, and on a bottleneck asset those are the costliest minutes of the day.

The agent triggers on the stoppage itself. It pulls the machine's recent sensor trace and its last three service records. It pulls each earlier stoppage with the same fault code, and what fixed them. It pulls the right section of the OEM manual, plus current stock of the parts those fixes used. It builds a briefing and sends it to whoever is on call, so it lands before they reach the machine.

Where you allow it, the agent goes further. It holds the likely spare, raises the work order in draft, and tells the planner the schedule needs resequencing. Each of those actions is a rule you set, and anything with a cost attached waits for a human yes.

Procurement & Reorder Agent

The chasing that eats a purchase officer's week, done daily and without a reminder.

Procurement in most plants is one person with a spreadsheet and a phone, working through a list of things that might be late. The work is not hard. It is relentless, and easy to drop. An order slips two weeks because no one chased it on the right day, and the first anyone hears of it is when the line stops.

The agent watches stock against usage and open orders. When an item crosses its reorder point, it drafts the requisition with the right vendor, quantity and date. It chases unconfirmed POs on a schedule and flags any promised date that has moved. It escalates when a late delivery threatens a booked job. Not late in the abstract. Late in a way that will cost you.

The judgement stays yours. Approval limits, which vendors it may use, whether it can split an order: all of these are rules you write. What it removes is the twenty small follow-ups a day that no one has time for and the plant pays for.

Quality Incident Agent

When rejects spike, the enquiry starts itself instead of waiting for a meeting.

A reject rate climbs on Tuesday. It gets noticed on Thursday, when the numbers are compiled, and discussed on Friday. By then the material lot has been used up, the tool has been changed, and the operator does not recall the shift. The evidence that would have found the cause has gone.

The agent watches reject rates by line, product and defect type, all day. When one moves outside its normal range, it acts at once. It pulls the process parameters for that window. It pulls the material lots in use, tool change records, operator and shift, and any maintenance done in the days before. It compares all this with the last time the defect appeared. Then it drafts the incident report in your own 8D or CAPA format, with the evidence attached.

What lands on the quality manager's desk is not an alert saying rejects are up. It is a filled-in report: rejects are up, here is what changed in that window, and here is what it looked like the last two times. The manager's job becomes judgement instead of assembly.

Shift Handover Agent

The handover note written in full each time, even on the shift when all hands are busy.

Handover is where plants lose the thread. On a calm shift the note is thorough. On the shift where three things went wrong, it gets written in four lines by a man who wants to go home. That is the shift where a good note matters most. The incoming supervisor starts without knowing which machine sounded odd at 4am.

The agent builds the handover from what happened: output against plan by line, each stoppage with its duration and cause, and quality exceptions. It adds jobs part-finished and where they stand, open maintenance items, and anything a previous shift flagged to watch. It writes this in the format your plant uses, and posts it where handovers already happen.

The supervisor edits it and adds what only a person knows: the noise that was not quite a fault, or the operator who is off tomorrow. That takes two minutes instead of fifteen, and it also happens on the bad shifts, since the draft is already written.