---
title: "Add AI to old factory machines without replacing them | Quantlix"
description: "You can add AI to an old factory machine without replacing it: read its controller, fit a few sensors and let a small computer learn what normal looks like."
canonical: "https://quantlix.com/en/blog/add-ai-to-old-factory-machines/"
last-updated: 2026-11-21
---

Source: https://quantlix.com/en/blog/add-ai-to-old-factory-machines/
Language: en
Description: You can add AI to an old factory machine without replacing it: read its controller, fit a few sensors and let a small computer learn what normal looks like.

Playbook Industrial AI

# Can you add AI to an old factory machine without replacing it?

You can add AI to an old factory machine without replacing it: read its controller, fit a few sensors and let a small computer learn what normal looks like.

[Ali Muhammad](https://quantlix.com/en/about/leadership/ali-muhammad/)

Founder and Chief Executive Officer

Nov 21, 2026 (2026-11-21) 16 min read

![An old blue industrial machine with pressure gauges, with the headline “Your old machine can learn”.](https://quantlix.com/_astro/insight-add-ai-to-old-factory-machines.DvtzpGmW_4324r.webp)

01 / Engineering notes

Field notes: Blog

**On this page**

The short answer

Yes. An old machine can use AI without being replaced. You read the data its controller already holds, fit low-cost sensors where it has none, and put a small computer beside it that learns how the machine normally behaves. The machine runs as before while it starts reporting on itself and flagging changes early.

Walk through almost any plant and you will meet it: the machine that is older than the person running it. A loom from the eighties, a lathe that came second-hand with a manual nobody on the floor can read, a press that has outlived three owners and a few confident predictions of its retirement. It runs six days a week, it pays salaries, and everyone is fond of it, including the people who curse it on Monday mornings.

**Adding AI to an old machine** means leaving the machine as it is and building the intelligence around it: reading what its controller already knows, fitting sensors where it knows nothing, and letting a small computer beside it learn what normal looks like. Nobody has to scrap anything. Here is how that works on a real floor, what the published evidence shows, and where it stops.

## What does an old machine already know?

Many old machines already hold useful data in their controller, and a small adapter can read it without touching the machine’s program. The controller is usually a PLC, short for programmable logic controller: a small, tough computer whose job is to run the machine’s sequence. It starts the motor, closes the guard, counts the cycle and stops everything when something is wrong.

To do that job, the controller knows the speed, how many parts were made this shift, when the machine stopped and which sensor tripped. In many plants none of that leaves the cabinet. The operator sees a red light or a code, clears it, presses start, and the knowledge is gone. Repeat that for every stop on every shift for years and the plant has been throwing away its own history.

The first step is to listen. A small adapter called a **gateway** connects to the controller’s communication port and reads the values it already produces. Set up to read only, it leaves the machine’s program unchanged, and from that day the machine has a memory.

Controllers speak a handful of common industrial protocols. [Modbus](https://modbus.org/faq), introduced by Modicon in 1979, is still widely used, and [OPC UA](https://quantlix.com/en/standards/opc-ua/), released by the OPC Foundation in 2008, is a platform-independent newer option. Gateways exist for many of them.

If the controller has no usable port, or its maker has closed down, you can read the machine from the outside. Researchers at Seoul National University pointed a webcam at a milling machine’s control screen and read the numbers off it with open-source software, as described below.

## How do you add sensors to a machine with no electronics?

You fit sensors on the outside: a clamp, a probe, a beam or a camera, none of which has to touch the machine’s controls. A mechanical press, an old mixer or a belt-driven gearbox does its work beautifully and tells nobody anything.

| What you want to know | What you fit | What it can tell you |
| --- | --- | --- |
| Effort | A split-core clamp around one of the motor’s power wires | When the machine runs, idles or strains |
| Vibration | A small sensor fixed to a bearing housing | Changes in vibration, a common early sign of bearing wear |
| Heat | A temperature probe on a motor, bearing or heater | Parts running hotter than they used to |
| Count | A light-beam sensor across the outfeed | How many pieces came out, and how evenly |
| Sight | A camera at the end of the line | Whether each piece came out right |

The clamp is the quiet hero of that list. It closes around a wire like a bracelet, with no cutting or rewiring of the conductor, and the current through it says a lot. A machine that idles draws a little, a machine that works draws more, and a machine that draws more for the same output may have a dragging bearing or a blunt blade. From that one reading you can work out running hours and idle time.

If one of these sensors fails, the machine carries on as before. Fitting is still not a job for just anyone. Anything inside an electrical panel belongs to a qualified electrician with the power isolated and locked out under your site’s procedure, ideally during a stop you had planned anyway. Leave safety circuits (guards, emergency stops, interlocks) untouched, and in a food plant ask your food safety team before anything goes in or near a product zone.

## Where does the AI actually live?

The AI lives in a small, rugged computer fixed near the machine, which engineers call the **edge** because it sits at the edge of the network, beside the work, rather than in a distant data center. It does two jobs, and the order matters.

Its first job is to learn what normal looks like for this machine in particular. Old machines have personalities. One always runs a few degrees warm. One draws a little more power on the night shift because the building is cooler and the oil is thicker.

A good system spends its first stretch simply watching how current, heat and vibration behave at each speed and on each product, and only then starts to judge.

Its second job is to notice when something leaves that picture: current creeping up each week, a vibration that was not there in October, a run of pieces all a fraction too short. None of these would trip an alarm alone, and a person walking past would not spot them. A patient computer that never blinks will.

Because the computer sits on site, it keeps working when the internet does not. Only the report to the office waits.

## What does the first month look like?

In its first month of listening, an old machine starts to show how many hours it really runs, why it stops and when something begins to drift. The plant below is an illustration: it is invented, and so are its numbers, so it shows the order of events, not a real result.

Picture a plastics plant with an injection molding machine bought in 1998. It makes bottle caps, its controller has never been connected to anything, and the owner is sure it runs about twenty hours a day.

Day one: An electrician fits a clamp on the main motor wire and a temperature probe on each heater zone, and a gateway reads the controller’s cycle count and alarms. The machine stops during a mold change that was happening anyway.

Day four: The first surprise arrives. Over three days the machine ran about fourteen and a half hours in every twenty-four. The rest went on gaps at shift change, waits for raw material and a string of short stops for the same small jam nobody had written down. Nothing has been predicted yet, but the owner knows the truth for the first time.

Week three: The system has learned the machine’s habits: how much current each heater zone draws, how long a normal cycle takes, and how both change when the cap color changes.

Week six: One heater zone starts drawing a little less current each day. To the operator the caps still look fine, but to the system it looks like a heater band on its way out. The maintenance lead gets a message with a simple chart and fits a new band on Sunday. Left alone, that zone would have run cold, and the first sign would have been a pallet of misshapen caps.

## What does a monitored old machine give you, and in what order?

A monitored old machine gives you three things, in this order: the truth about what it does, early warnings, and answers.

First, the truth: How many hours the machine really ran, how often it stopped, for how long and at what time of day. Owners are often surprised by their own numbers, and many problems become obvious the moment they can be seen. This stage can pay for itself before anything clever has happened.

Then, warnings: Once the system has seen enough normal running to know what normal looks like, it can flag drift before a failure. The repair happens on your schedule, with the part already on the shelf, instead of in the middle of an urgent order.

After that, answers: A reason attached to every stop, so the same fault is not rediscovered by a different shift each month. A check on every piece, so a bad run is caught after ten pieces and not ten thousand. And a record of what your process did, which you can show a customer or auditor who asks how you know it is under control.

## Has anyone added AI to very old machines?

Yes: a packaging plant in India has used machine learning on machines about sixty years old, and a university team has watched a milling machine with a webcam. Both examples have limits, stated below so you can weigh them.

In its January 2026 white paper on the [Global Lighthouse Network](https://reports.weforum.org/docs/WEF_Global_Lighthouse_Network_2026.pdf), the World Economic Forum describes ACG Packaging Materials’ plant in Shirwal, India. Operators there set up more than 70 parameters by hand on 60-year-old machines, across more than 5,000 product variants, and the report says this hurt yield, lead time and defect rate. ACG built machine learning models in-house, trained on more than two years of “golden batch” data (records of runs that came out right).

The models prescribe settings, operators transfer them to the machine’s PLC, and the models keep retraining on outcomes and operator feedback. The report lists a **37 percent improvement in first pass yield**, which is the share of output that is right the first time.

Read that figure with care. The report gives no baseline or period for it, and the plant is a Lighthouse member, not a typical one. It is also a different kind of AI from the monitoring described above: it advises on settings, and people stay in the loop. The same report’s technology diagram makes the broader point that best-practice IoT architectures can be deployed in existing plants by retrofitting smart sensors, IoT gateways and/or industrial PCs to get data from legacy machines.

The second example is much smaller. In October 2019, Hyungjung Kim and colleagues at Seoul National University published [a low-cost monitoring system](https://pmc.ncbi.nlm.nih.gov/articles/PMC6832709) in the journal *Sensors*. A webcam watched the control screen of a 3-axis CNC milling machine, and open-source software (Tesseract and OpenCV) read the program line number, feed rate and spindle speed once a second.

The paper’s hardware table totals 393 US dollars, under the 500 dollars its abstract cites, and the camera only watched the screen. It is a proof of concept: one machine, one part program, about 40 minutes of running.

We looked for independent published results on fitting sensors to old machines and found few, so this section rests on these two sources and does not stretch them.

## What will adding AI not do?

Adding AI to an old machine will not repair the machine, fix a badly set-up process or replace the people who know it. It gives you better information, and four limits are worth stating before you start.

- It will not make a worn machine precise again. It tells you sooner and more clearly that the machine is worn, and roughly how fast it is getting worse, which is what you need when choosing between a rebuild and a replacement.
- It will not fix a badly set-up process. You will simply have very good information about the bad settings, which is the first step to changing them.
- It will not replace the people who know the machine. The best results come when the system and the old hands check each other: the data shows something odd, and the operator who has run that machine for fifteen years says, “That’s the cooling water. It does that every summer.”
- A warning changes nothing on its own. It helps only if someone is responsible for acting on it, so decide who that person is before the first sensor goes on.

## How do you connect an old machine safely?

Let readings leave the machine through one guarded route, and let nothing from outside send it an instruction. Older controllers were often designed for a time when the only thing talking to them was an engineer’s laptop on a cable, so assume one may accept commands from whatever can reach it. A little care at connection time matters.

Let the machine report out, and never let anything send an instruction back in. In practice the gateway only reads from the controller and never writes to it. The small computer sits on its own part of the network, kept apart from the office and the internet. Any connection to the outside world goes through a single protected route, with proper passwords and nothing left on the factory default.

Public guidance points the same way. In December 2025, the US Cybersecurity and Infrastructure Security Agency (CISA) and the Australian Signals Directorate’s Australian Cyber Security Centre published [Principles for the Secure Integration of Artificial Intelligence in Operational Technology](https://www.cisa.gov/sites/default/files/2026-01/joint-guidance-principles-for-the-secure-integration-of-artificial-intelligence-in-operational-technology-508cV2.pdf), together with agencies from the United States, Canada, Germany, the Netherlands, New Zealand and the United Kingdom. Among its recommendations are push-based designs, where data is pushed out of the operational network for AI to use without persistent access back in, failsafe mechanisms that revert to traditional automation or manual operation, and limits on AI controlling equipment without a human in the loop. It is written for critical infrastructure owners and is a useful yardstick for any plant.

Two more references help when you question a supplier. The US National Institute of Standards and Technology published its [Guide to Operational Technology (OT) Security](https://csrc.nist.gov/pubs/sp/800/82/r3/final) in September 2023, covering how to secure industrial systems while respecting their performance, reliability and safety needs. The [IEC 62443](https://quantlix.com/en/standards/iec-62443/) series describes dividing a system into zones and the paths between them.

Naming a guide here is not a claim that any product or plant meets it. Our note on [keeping connected machines safe as AI tools get cheaper](https://quantlix.com/en/blog/open-weight-ai-models-factory-security/) goes further.

Whoever you work with, these six questions tell you a great deal in a short conversation:

- Does anything you fit change the machine’s program, or touch the wiring to its controls?
- If your box fails or loses power, what happens to the machine? (The only good answer is “nothing”.)
- Does it keep working when the internet or a vendor’s cloud is out of reach?
- Can anyone send the machine a command from outside the plant?
- Who owns the data, can we take it with us if we stop working together, and will anyone train models on it?
- What will we know in a few weeks that we do not know today?

A good supplier will enjoy these questions. This is general guidance, not a safety or legal assessment: electrical, machine-safety and food-safety rules apply on your site, so involve the people responsible for them.

## Where to start

Start with one machine, not the whole plant: the one whose stoppages cost you the most, the one the night shift dreads. You can do these steps this month without buying anything.

1. Name the machine and write down what it offers. Note the controller’s make and model, its communication ports and the nameplate details. This is a small version of the asset inventory (an organized, regularly updated list of systems, hardware and software) that [CISA’s August 2025 guidance](https://www.cisa.gov/resources-tools/resources/foundations-ot-cybersecurity-asset-inventory-guidance-owners-and-operators) recommends for operational technology.
2. Keep a simple stop log for a few weeks. Record the time, the length and the reason for each stop on paper. It is the baseline you will judge any sensor against, and it often shows the biggest gap before any technology arrives.
3. Decide who receives a warning and what they will do. Name a person and an action now, so the first alert has somewhere to go.
4. Agree the safety terms. A qualified electrician, a planned stop, read-only connections and a separate network segment. Write them down before anyone quotes.
5. Pick the one extra sense. Usually vibration on a bearing or temperature on a heater, wherever that machine is most likely to fail.

Our guide to [starting with AI in a factory](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/) covers the wider picture, and [predictive maintenance for food plants](https://quantlix.com/en/blog/predictive-maintenance-food-plants/) shows where warnings lead. For the engineering side, see our [IoT and edge computing](https://quantlix.com/en/services/iot-edge-computing/) service.

If you have a machine like this in mind, tell us about one machine or one line and we will reply with a short note on what we would measure first. [Tell us about your machine](https://quantlix.com/en/contact/).

## Key takeaways

1. Many machines already hold useful data in their controller, and a read-only gateway can collect it without changing the machine's program.
2. A machine with no electronics can still be monitored from outside, with sensors for power, vibration, heat, count and sight.
3. The AI runs on a small computer beside the machine. It learns what normal looks like first, then flags drift, and it keeps working if the internet drops.
4. Start with the one machine whose stoppages cost you most, and decide who acts on a warning before anything is fitted.
5. Sensors will not make a worn machine precise again. They show sooner and more clearly that it is worn.

## Questions people ask

### Can a purely mechanical machine use AI?

Yes. A machine with no electronics can be fitted with sensors on the outside: a clamp on a motor power wire, a vibration sensor on a bearing, a light-beam counter on the outfeed, or a camera watching what it makes. A small computer beside it learns the machine's normal behavior from those readings and flags changes that point to wear, a jam or a drift in quality.

### Will adding sensors change how my machine runs?

It should not. The sensors listen: they are fixed to the outside of the machine or read signals its controller already produces, and nothing in the machine's program is altered. If every sensor were removed tomorrow, the machine would run as it did before. Work inside an electrical panel belongs to a qualified electrician with the power isolated, and safety circuits should be left alone.

### What if my machine's controller is too old to connect?

There is often a way around it. Adapters exist for many older controllers. Where none fits, you can read the machine from the outside instead: a camera pointed at its display, as researchers at Seoul National University did on a milling machine in 2019, or sensors that measure what the controller would have reported, such as power, cycles and temperature.

### Does it need an internet connection?

Not to do its job. The small computer beside the machine does its thinking on site, so it carries on through an outage, and only the report to the office waits. A connection is useful for sending alerts to a phone or an office, and it should pass through one protected route that cannot be used to send the machine commands.

### How long before it is useful?

The first useful picture can come as soon as data flows: real running hours, stops and idle time, often with a surprise or two. Warnings about wear take longer, because the system first has to see enough normal running to know what normal looks like for that machine, on your products and across your shifts. The time it takes depends on how varied your work is.

### Is it cheaper than buying a new machine?

Usually, because nothing is replaced, but the honest comparison is against what that machine's stoppages and scrap cost you. The Seoul National University team priced the hardware for its 2019 prototype at 393 US dollars. A fuller setup with several sensors and an edge computer costs more, and we do not quote prices here because they depend on the machine and the site.

## Sources

1. [World Economic Forum, Global Lighthouse Network: Rewiring Operations for Resilience and Impact at Scale (white paper, January 2026)](https://reports.weforum.org/docs/WEF_Global_Lighthouse_Network_2026.pdf)
2. [Kim, Jung, Choi and Ahn, Seoul National University, A Low-Cost Vision-Based Monitoring of Computer Numerical Control (CNC) Machine Tools for Small and Medium-Sized Enterprises (SMEs), Sensors 19(20):4506 (October 2019)](https://pmc.ncbi.nlm.nih.gov/articles/PMC6832709)
3. [CISA, Australian Signals Directorate's ACSC and partner agencies, Principles for the Secure Integration of Artificial Intelligence in Operational Technology (December 3, 2025)](https://www.cisa.gov/sites/default/files/2026-01/joint-guidance-principles-for-the-secure-integration-of-artificial-intelligence-in-operational-technology-508cV2.pdf)
4. [CISA, Foundations for OT Cybersecurity: Asset Inventory Guidance for Owners and Operators (August 13, 2025)](https://www.cisa.gov/resources-tools/resources/foundations-ot-cybersecurity-asset-inventory-guidance-owners-and-operators)
5. [NIST, Guide to Operational Technology (OT) Security, SP 800-82 Rev. 3 (September 2023)](https://csrc.nist.gov/pubs/sp/800/82/r3/final)
6. [Modbus Organization, Modbus FAQ (Modbus introduced by Modicon in 1979)](https://modbus.org/faq)
7. [OPC Foundation, OPC Unified Architecture (released 2008)](https://opcfoundation.org/about/opc-technologies/opc-ua/)

Sources checked on Nov 21, 2026.

- Playbook
- Industrial AI

Written by

## Ali Muhammad

Founder and Chief Executive Officer · Quantlix

[See full profile](https://quantlix.com/en/about/leadership/ali-muhammad/)

[Talk to Ali](https://quantlix.com/en/contact/)

[More blog](https://quantlix.com/en/blog/)

## Keep reading

[More blog](https://quantlix.com/en/blog/)

- ![A circular conveyor in a large, bright factory hall, with the headline “The model is the easy part”.](https://quantlix.com/_astro/insight-start-ai-in-a-factory.BsmMIis5_Z1XFbFd.webp)

  Playbook

  Industrial AI Nov 21, 2026 (2026-11-21)

  ### [How should a factory start with AI?](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/)

  Start AI in a factory with one line and one costly problem. Before you buy, check five things: machine data, wrong answers, an owner, internet loss and access.

  15 min read Read
- ![Close-up of old machinery with brass gears, with the headline “Machines that warn you before they stop”.](https://quantlix.com/_astro/insight-predictive-maintenance-food-plants.BXZKidCI_ZquUlF.webp)

  Playbook

  Food manufacturing Nov 21, 2026 (2026-11-21)

  ### [How can a food plant know a machine will fail before it does?](https://quantlix.com/en/blog/predictive-maintenance-food-plants/)

  How a food plant can tell a machine will fail before it does: the readings to watch, who already does it, where projects go wrong and how to start small.

  16 min read Read
- ![A long control room lined with instrument panels, with the headline “One screen for the whole plant”.](https://quantlix.com/_astro/insight-central-monitoring-system-food-plant.CmxcUk1I_ZIbB0B.webp)

  Playbook

  Food manufacturing Nov 21, 2026 (2026-11-21)

  ### [What is a central monitoring system, and how does a food plant build one?](https://quantlix.com/en/blog/central-monitoring-system-food-plant/)

  A central monitoring system shows every line, cold room and machine stop on one screen. See what belongs on it and how a food plant builds one line at a time.

  14 min read Read
- ![An open electrical control cabinet with wired modules, with the headline “Open AI models, safe machines”.](https://quantlix.com/_astro/insight-open-weight-ai-models-factory-security.DRyiqiUu_1rrXUK.webp)

  Perspective

  Industrial security Nov 21, 2026 (2026-11-21)

  ### [Do open AI models make attacks on factories easier?](https://quantlix.com/en/blog/open-weight-ai-models-factory-security/)

  Do open-weight AI models make attacks on factories easier? Slightly, and the defense is familiar: list what is reachable online, close it, guard the rest.

  16 min read Read

## Need the version tailored to your platform, risk model and roadmap?

Bring us the context. We will turn the patterns in this library into a pragmatic delivery path your team can use.

[Plan a workshop](https://quantlix.com/en/contact/)
