---
title: "How to start with AI in a factory: five questions | Quantlix"
description: "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."
canonical: "https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/"
last-updated: 2026-11-21
---

Source: https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/
Language: en
Description: 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.

Playbook Industrial AI

# How should a factory start with AI?

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.

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

Founder and Chief Executive Officer

Nov 21, 2026 (2026-11-21) 15 min read

![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_26TRHJ.webp)

01 / Engineering notes

Field notes: Blog

**On this page**

The short answer

Start with one production line and one problem that costs money every week, and measure it before changing anything. Then check five things before you buy: whether your machines can report what they do, what happens when the system is wrong, who owns it, whether it works without the internet, and who can reach it.

Picture a camera above a packing line that has learned to spot dented cartons. In the supplier’s demonstration it is right almost every time, and the order is signed that afternoon. By the second Friday on the floor, the supervisor is quietly checking by eye again. (This is an illustration, not a real plant.)

**Starting with AI in a factory** means choosing one line and one problem that costs money every week, measuring it, and checking five practical things around the model before you buy anything. Those five things are the subject of this article, and almost all of them are in your hands.

## Why does factory AI fade after a good demo?

Factory AI usually fades because the conditions around the model change, not because the model gets worse. A demo runs once, with clean inputs, good light and everyone watching. A floor does none of that.

Go back to the camera. Around three o’clock the sun comes through a side door and lays a bright stripe across the belt, something the camera has never seen. A new supplier’s cartons arrive in a slightly different shade of brown. Someone nudges the mounting while cleaning on the night shift.

The model still does exactly what it learned to do. What changed was everything around it.

The model has also become the easier part to get. On April 2, 2026, Google DeepMind launched Gemma 4 under the Apache 2.0 license, and Google’s [developer blog](https://developers.googleblog.com/bring-state-of-the-art-agentic-skills-to-the-edge-with-gemma-4/) described bringing it to the edge on a Raspberry Pi 5, a small single-board computer. That shows how far capable models have moved onto small hardware.

When the model is this easy to get, what decides success is everything around it: the machine data, the plan for wrong answers, an owner, the network, and who can get in. Each of those can be checked before you spend anything.

## What are the five questions to answer before you buy?

Five questions cover most of what goes wrong after a demo: machine data, wrong answers, ownership, connectivity and access. The table gives a good answer for each and one question to put to any supplier.

| Ask this | What a good answer looks like | What to ask a supplier |
| --- | --- | --- |
| Can the machine tell you what it is doing? | Running time, stops, faults, speed and rejects are recorded where people can see them | What will the system collect from our machines, and can we see and keep all of it? |
| What happens when the system is wrong? | It sets doubtful items aside and asks a person | What does it do when it is not sure, and how do our operators correct it? |
| Who looks after it on a Tuesday night? | One named owner with a short daily check | Who on our side looks after it, and what does the daily check involve? |
| Does it work without the internet? | Fast decisions are made on a computer beside the line | What keeps working during an internet outage or a power cut, and what stops? |
| Who else can reach it? | Only the people you chose, through one guarded entrance | Who can reach the system from outside the plant, and how is that access controlled and switched off? |

Add a sixth question for every supplier: how will we know it worked, and what are we measuring against? The first-project section below shows one way to answer it.

## Can your machines tell you what they are doing?

Most machines already know a great deal about themselves, but in many plants that knowledge never leaves the controller. The controller is the small, tough industrial computer that runs the machine’s sequence. It holds speed, temperature, counts and a fault code for every stop, and the operator often sees only a light. They clear the fault, and the information is gone.

Getting it out and onto a screen where people can see it is, in our view, some of the best money a plant can spend on a first line, because it pays off before any AI is involved. Once the stops are visible, a surprising number turn out to have plain fixes: a sensor that needs cleaning, a changeover that always runs long on nights, one supplier’s material that jams more than another’s.

The stakes are real at large scale. Siemens’ 2024 report [The True Cost of Downtime](https://assets.new.siemens.com/siemens/assets/api/uuid:1b43afb5-2d07-47f7-9eb7-893fe7d0bc59/TCOD-2024_original.pdf), which draws mainly on large manufacturers, found that an average large plant loses 27 hours a month to unplanned downtime, down from 39 in 2019. It put the cost of an hour at about 36,000 dollars in fast-moving consumer goods, the lowest of the sectors it covered. It also found that nine out of ten manufacturers surveyed collect at least some data on machine health.

A smaller plant will have smaller figures, but the pattern holds: the stops you cannot see are the ones you keep paying for.

What should you capture first? For most lines, four things are plenty: when the machine is running, when it is stopped and why, how fast it is going, and how many pieces it rejects. Teams often roll these into overall equipment effectiveness (OEE), a measure of availability, speed and quality.

Older machines can be given senses from the outside: a clamp around the power cable, a small vibration sensor on a bearing, a camera pointed at the screen. In a [2019 paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC6832709) in the journal *Sensors*, Hyungjung Kim and colleagues at Seoul National University described monitoring a CNC milling machine with a webcam reading its control screen and open-source software, for a cost the abstract puts under 500 dollars. They tested one prototype on a three-axis mill and say the approach suits legacy machine tools.

At larger scale, the World Economic Forum’s January 2026 [Global Lighthouse Network white paper](https://reports.weforum.org/docs/WEF_Global_Lighthouse_Network_2026.pdf) describes ACG Packaging’s site in Shirwal, India. Operators had set more than 70 parameters by hand on 60-year-old machines across more than 5,000 products. Machine learning models trained on more than two years of records from batches that went well now suggest the settings, and the report lists a 37 percent gain in first pass yield, which we would treat as the site’s own reported result. For the method in more detail, see [how to add AI to old factory machines](https://quantlix.com/en/blog/add-ai-to-old-factory-machines/).

## What should happen when the system is wrong?

A system built for the floor should know when it is unsure, set the item aside and ask a person instead of guessing. Every model is wrong some of the time. In a chat window a wrong answer costs a moment, but on a line it can cost a rejected batch or a good batch thrown away.

On a packing line, asking might be a small screen beside the camera that shows a photograph of the carton and two buttons: “dented” and “fine”. The operator taps one in a second and carries on. When those answers are fed back into training, the system can learn from the people who know the product best. The same WEF white paper describes the Shirwal site’s pipeline as refining its models from outcomes and operator feedback.

That keeps mistakes small and visible. It also makes operators the system’s teachers rather than its audience, which we think is usually the difference between a tool people trust and one they quietly work around. A rising count of “unsure” items is also an early signal that conditions have changed. For how a camera system learns what good looks like in a food plant, see [edge AI vision in food inspection](https://quantlix.com/en/blog/edge-ai-vision-food-inspection/).

When a supplier shows you a system, ask to see what it does with an item it is not sure about. If the answer is that it is always sure, be skeptical.

## Who looks after it on a Tuesday night?

One named person in the plant should own the system and run a short daily check, because conditions on a floor never stop moving. The light changes with the seasons, a new product goes on the line, a lens collects a film of flour or oil, and a supplier changes the shade of their packaging. None of it is dramatic, and all of it pulls the system away from the conditions it learned in.

Engineers call this drift. [Google Cloud’s documentation on model monitoring](https://docs.cloud.google.com/vertex-ai/docs/model-monitoring/overview) describes it as production data changing over time, and notes that after a model is deployed the input data can deviate from the data used to train it. Its advice when that happens is to re-evaluate or retrain the model.

Somebody has to notice, and that somebody needs a name. Not “the IT team” or “the supplier”, but one person who thinks of the system as theirs. Their job does not have to be big:

- **Daily, a few minutes:** wipe the lens, look through the last twenty rejects to confirm they really were faults, and check that the count on the screen matches the packing area.
- **Monthly:** look at the numbers for anything creeping in one direction.
- **Any shift, any time:** one easy way to say “this does not look right, please have a look”.

Without a name attached, drift tends to go unnoticed until someone stops trusting the screen.

## Does it still work when the internet stops?

Fast decisions should run on a small, rugged computer beside the line, so production never waits on a connection. If a decision has to be made in the time a carton passes a camera, often a fraction of a second, the computer making it should be standing right there.

Engineers call this running AI at the edge. A plant can afford to lose its weekly report during an outage, but it should not lose its line. Our [IoT and edge computing](https://quantlix.com/en/services/iot-edge-computing/) page describes how we separate local processing from central storage.

The cloud still matters. It is the right home for slow, heavy thinking: comparing this month with last, learning from a long run of data, teaching the model a new product. Fast decisions near the machine, slow thinking far away.

So ask any supplier one friendly question: what happens when the internet stops? A good team will have a clear answer, because it will have thought about it long before you asked.

## Who else can reach your machines?

Only the people you chose should be able to reach your machines, and the first step is finding out what can be reached today. Industrial equipment is often more reachable than its owners realize.

In a September 2026 post, the internet-scanning firm [Censys](https://censys.com/blog/2026-state-of-the-internet-ics) said it observed an average of about 134,000 hosts running industrial control system services exposed on the internet in early 2026, with roughly 70 percent on consumer and mobile networks. Censys points out that cellular links are common at remote sites and quotes CISA’s warning that modems installed by operators, vendors or integrators may not be documented. That is one firm’s count from its own scans.

The US Cybersecurity and Infrastructure Security Agency (CISA) gives a plain sequence in its [Internet Exposure Reduction Guidance](https://www.cisa.gov/resources-tools/resources/exposure-reduction), revised in August 2026:

1. **Assess** what is reachable from the internet, including remote access held by integrators and vendors.
2. **Evaluate** which assets genuinely need to be reachable.
3. **Mitigate** the risk for those that stay: change default passwords, apply patches, replace devices that no longer receive security support, use a monitored jump host, and use multifactor authentication where possible.
4. **Repeat** the review routinely as your systems change.

CISA is explicit that reducing exposure does not mean switching off remote access you need. We would add one habit of our own: let a new AI system read from a machine and recommend before you let it write to one. The international [IEC 62443](https://quantlix.com/en/standards/iec-62443/) series describes dividing a control system into zones and conduits, each with a target security level, and our piece on [open AI models and factory security](https://quantlix.com/en/blog/open-weight-ai-models-factory-security/) covers what changes when model weights can be downloaded openly.

A safeguard worth stating plainly: emergency stops, guard interlocks and other machine safety functions belong in the machine’s own safety system, designed and checked under the rules that apply where you operate. In a food plant, your food safety plan decides what counts as a critical control. An AI system should flag or recommend, and a person or the safety system keeps the last word. Check the specifics with your machine builder and safety officer; this is general guidance, not legal advice.

## What could a first project look like, week by week?

A sensible first project measures before it builds, tries the simple fix first, and compares against its own baseline. The shape below is an illustration built around a common problem, not a description of a real plant, and the timings are a suggestion rather than a standard.

Picture a biscuit factory with three lines. Line two has a packing machine that jams several times a shift. Every jam means lost minutes, a small pile of broken packets and an exasperated supervisor. Nobody knows quite how many jams there are, because each one is cleared and forgotten.

- **Weeks one and two: measure.** Before changing anything, the plant reads the packer’s controller and records every stop with its time, its length and its fault code. Where the code is unclear, the operator adds a one-word reason on a tablet. For the first time the plant has a real count, and it can see when the jams cluster, perhaps in the hour after a product changeover.
- **Week three: build the smallest useful thing.** A pattern like that can point to a fix that needs no AI at all, such as a tighter changeover checklist. Alongside it, a camera goes in just before the packer to spot misaligned biscuits and push them aside.
- **Weeks four to six: compare.** The plant keeps measuring exactly as before and sets the new numbers beside the first two weeks. If jams have fallen clearly, the case for lines one and three makes itself, with the plant’s own figures. If not, the plant has learned precisely where the problem is not, which is still progress.

Notice what made it work: a clear problem, a measurement taken before anything changed, a named owner, and the good sense to try the simple fix first.

## Where to start

You can do all of this month’s steps without buying anything.

1. **Choose one line and one problem.** Pick a problem that costs money every week: a stop that keeps coming back, a defect caught too late, a morning report that takes someone two hours to piece together from paper. Write it as one sentence with a number you can count.
2. **Record it before changing anything.** Use a clipboard or a tablet: when the machine stopped, why, how fast it ran, how many pieces were rejected. A few weeks of this gives you a baseline.
3. **Name an owner and write their daily check.** One person, a few minutes a day, and an easy way for any shift to raise a concern.
4. **Write down your answers to the five questions for that line.** Make a list of what is connected to your network, as CISA’s first step suggests. Ask any supplier the same questions from the table.
5. **Decide what worked means.** Agree the number, the review date and what you will do if it has not moved.

If it works, you extend it with your own evidence. If it does not, you have spent a short time and learned something true about your plant, which is a very good trade.

If there is one machine or one line you have been wondering about, [tell us about it](https://quantlix.com/en/contact/) and we will reply with a short note on what we would measure first. Sometimes that note will say AI is not the first thing to try.

## Key takeaways

1. Pick one line and one problem that costs money every week, and measure it for a few weeks before anyone changes anything.
2. Check five things around the model: machine data, wrong answers, a named owner, loss of internet, and who can reach the system.
3. Siemens' 2024 report puts unplanned downtime at 27 hours a month for an average large plant. Stops you cannot see are stops you keep paying for.
4. CISA's August 2026 guidance: find what is reachable from the internet, remove what does not need to be, protect what remains, and check again regularly.
5. Name one owner with a short daily check, because light, packaging and wear slowly change what the system sees.

## Questions people ask

### What is the first step to using AI in a factory?

Pick one line and one problem that costs money every week, then record that problem before changing anything. Capture when the machine runs, when and why it stops, its speed and its rejects. That baseline tells you what better means in numbers, and it often shows fixes that need no AI at all, such as a slow changeover or a sensor that needs cleaning.

### Why do AI projects that work in a demo struggle on the factory floor?

A demo runs once, with clean inputs, good light and everyone watching. A real floor has changing light, new packaging, dust and night shifts. The model itself is rarely the weak point. The usual gaps are missing machine data, no plan for wrong answers, no named owner to catch drift, and a dependence on a connection that is not always there.

### Does a factory need new machines to use AI?

Usually not. Many machines already record speed, stops and fault codes in their controller, and that information can often be read without changing the machine's program. Older machines can be given low-cost sensors on the outside, such as a current clamp or a camera. The intelligence is added around the machine, which carries on running as before.

### Should factory AI run in the cloud or on site?

Both, split by speed. Decisions that must happen in a fraction of a second, such as pushing a faulty item off a moving belt, belong on a small computer beside the line so production never waits for the internet. Slow, heavy work such as monthly analysis, comparing plants or retraining a model belongs in the cloud.

### How do you know whether a first AI project worked?

Decide the measure before you start, record it for a baseline period, then record it the same way afterward and compare. Use one number tied to the original problem, such as stops per week or rejects per shift, and agree a date to review it. If the number has not moved, you have learned where the problem is not.

### Is it safe to connect factory machines to the internet?

It can be, if the connection is deliberate. CISA's guidance says to find out what is reachable, remove access that is not needed, and protect what remains with changed default passwords, patches, a monitored jump host and multifactor authentication where possible. Reducing exposure does not mean switching off remote access you genuinely need.

## 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. [Siemens (Senseye), The True Cost of Downtime (2024)](https://assets.new.siemens.com/siemens/assets/api/uuid:1b43afb5-2d07-47f7-9eb7-893fe7d0bc59/TCOD-2024_original.pdf)
3. [Google Developers Blog, Bring state-of-the-art agentic skills to the edge with Gemma 4 (April 2, 2026)](https://developers.googleblog.com/bring-state-of-the-art-agentic-skills-to-the-edge-with-gemma-4/)
4. [Censys, 2026 State of the Internet: industrial control systems (September 8, 2026)](https://censys.com/blog/2026-state-of-the-internet-ics)
5. [CISA, Internet Exposure Reduction Guidance (revised August 21, 2026)](https://www.cisa.gov/resources-tools/resources/exposure-reduction)
6. [Kim, Jung, Choi and Ahn, A Low-Cost Vision-Based Monitoring of CNC Machine Tools for SMEs, Sensors 19(20):4506 (October 17, 2019)](https://pmc.ncbi.nlm.nih.gov/articles/PMC6832709)
7. [Google Cloud, Vertex AI Model Monitoring overview (documentation, read November 2026)](https://docs.cloud.google.com/vertex-ai/docs/model-monitoring/overview)

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/)

- ![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_2uP2mp.webp)

  Playbook

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

  ### [Can you add AI to an old factory machine without replacing it?](https://quantlix.com/en/blog/add-ai-to-old-factory-machines/)

  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.

  16 min read Read
- ![A row of automated machine cells with cable looms in a factory, with the headline “A camera that checks every item”.](https://quantlix.com/_astro/insight-edge-ai-vision-food-inspection.C95-UJkG_Z1OhuA9.webp)

  Playbook

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

  ### [How can a camera check every item on a food line?](https://quantlix.com/en/blog/edge-ai-vision-food-inspection/)

  A camera and a small computer beside a food line can check every item all shift, even offline. Here is how edge AI vision works and how to start.

  15 min read Read
- ![A grey control panel with buttons, dials and switches, with the headline “From the logbook to the dashboard”.](https://quantlix.com/_astro/insight-data-driven-decisions-food-manufacturing.SP8PWWLf_LESqq.webp)

  Playbook

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

  ### [How does a food plant that runs on paper start deciding with data?](https://quantlix.com/en/blog/data-driven-decisions-food-manufacturing/)

  A food plant that runs on paper can start with one costly question and five numbers: stops, yield, giveaway, energy and rejects. Here is how, step by step.

  16 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/)
