AI glossary for printing companies

A

7 terms

AI Act

The AI Act is the European law that imposes rules on whoever builds or deploys AI. The higher the risk of an application, the stricter the rules. Whoever uses AI at work must also make sure the staff know enough about it.

At the printing company: if you let a chatbot answer customer questions, the customer must know they are talking to AI.

AI agent

An AI agent is software that decides for itself which steps are needed to complete a task. For that it has connections to other systems such as your mail or your MIS. An assistant waits for your question. An agent gets to work on its own.

At the printing company: every morning an agent reads the new quote requests and prepares a proposal. You approve it before it goes out.

AI assistant

An AI assistant is a chatbot with a fixed role and some knowledge of its own. It answers when you ask something and does nothing on its own initiative. A Custom GPT is a well-known example.

At the printing company: an assistant knows the paper types and the delivery times of your company and helps to draw up a quote.

AI policy

An AI policy states on paper which AI tools you use and which data may never go into them. It is part of governance. For an SME a single page that everyone knows is often enough.

At the printing company: customer files and personnel data never go to a public chatbot. For quotes you only use the company account.

AI governance

AI governance is the way your company steers AI. Who decides what AI may do and who owns an application? How do you measure whether it pays off?

At the printing company: the business owner lays down that an AI assistant may prepare quotes. A price only goes to the customer after a check by the estimator.

Read-only connection

With a read-only connection the AI may look at data but change nothing. That way nothing can go wrong in your production systems while you still get answers.

At the printing company: in the trial at Drukkerij Moderna the AI looked into the ERP and the machine data. It could not write anywhere.

Anonymisation

Anonymisation is removing names and other identifiable details from your data before you share it. What remains still says something about the work but nothing about the person or the customer.

At the printing company: a local model replaces the customer names in an order list with codes. Only then does the list go to a model in the cloud.

API

API stands for Application Programming Interface. It is an agreed way for two programs to request data from each other without anyone having to retype anything.

At the printing company: through the API of your MIS a dashboard fetches today’s orders every five minutes.

Read more: What is an API?

Builder

API key

An API key is a secret code with which a program signs in to an API. Treat it like a password: never put it in an email or in a shared document.

At the printing company: you give your dashboard its own key with read-only rights on the MIS. In case of misuse you revoke that one key.

Read more: What is an API?

Builder

Atomic agent

An atomic agent is an AI agent that does one thing. Because its task is small you can test it well and you know exactly which rights it needs.

At the printing company: an agent that only reads order emails and fills in the order fields. It plans nothing and sends nothing.

Atomic task

An atomic task is a small task with a clear beginning and end. You first get it working and understand it before you expand.

At the printing company: a quick price check for a standard flyer instead of automating the whole calculation straight away.

C

8 terms

Chatbot

A chatbot is a program you talk to in plain language. The new generation runs on a language model and can therefore also handle questions nobody foresaw.

At the printing company: a chatbot at the machine knows the manual and tells the operator what an error message means.

Chunking

Chunking is splitting a document into pieces so that AI can later search it in a targeted way. You can cut it into equal pieces or follow the subheadings of the document. The second way keeps together what belongs together.

At the printing company: the manual of your cutting machine is split up per chapter. A question about changing the blade only retrieves that chapter.

Builder

Claude Code

Claude Code is a tool from Anthropic with which you build software by giving instructions in plain language. The AI writes the code and also runs it on your computer.

At the printing company: at Drukkerij Moderna the in-house team used it to build small applications that read along in the ERP and in the machine data.

Cloud AI

Cloud AI runs in the data centre of a supplier and not on your own devices. Your question and your data therefore travel outside. That is where the smartest models are.

At the printing company: you let a model in the cloud analyse last year’s figures, after you have removed the customer names.

Cobot

A cobot is a robot built to work next to people, without a cage around it. Within the project Sirris investigates how you control a cobot with plain language instead of code.

At the printing company: a cobot stacks boxes on a pallet while the operator keeps working next to it.

Computer vision

Computer vision is AI that looks at camera images and recognises something in them. It sees deviations that a human misses at high speed.

At the printing company: a camera above the web spots a smear or a register error while the press is running.

Context window

The context window is the amount of text a language model can take in at once. Your attachments and the conversation so far count too. Whatever does not fit, the model does not take into account.

At the printing company: a machine manual of eight hundred pages sometimes does not fit. Then RAG helps.

User

Custom GPT

A Custom GPT is your own version of ChatGPT with fixed instructions and your own documents. You set it up once and then use it again and again. Google Gemini calls the same idea a Gem.

At the printing company: an OfferBot knows your price agreements and writes a first version of a quote.

D

3 terms

Dashboard

A dashboard is a screen that shows the most important figures at a glance. It fetches those figures itself so that nobody has to retype them.

At the printing company: a screen in the pressroom shows the run length per order and today’s stoppages.

Builder

Digital work instruction

A digital work instruction shows step by step on a screen how to carry out a task, often with photos or short videos. That way craft knowledge is preserved when an experienced colleague leaves.

At the printing company: a tablet at the folding machine shows how to change over to another format.

Docling

Docling is an open-source tool from IBM that converts documents into a format AI can work with. It reads PDFs and Office files, among others, and it recognises the subheadings.

At the printing company: you convert the manuals of your machines with Docling before you load them into a RAG system.

Builder

E

3 terms

Edge AI

Edge AI is AI that runs on a device right next to the machine and not in the cloud. The answer therefore comes immediately and it also works when the internet drops out.

At the printing company: a camera with built-in AI inspects printed matter at the press without the images leaving the building.

Embedding

An embedding converts a piece of text into a series of numbers that preserves the meaning. Texts about the same subject get numbers that are close together. That is how AI also finds a passage that uses different words from your question.

At the printing company: you search for “paper tears” and also find the notes in which an operator typed “web break”.

Builder

ERP

ERP stands for Enterprise Resource Planning. It is the business software in which orders, stock, planning and invoices come together. In the printing industry that software is often called a MIS.

At the printing company: the ERP knows how many sheets were calculated for an order. The press knows how many were actually printed.

Evaluation gate

An evaluation gate is a test an AI application has to pass before it may go further. A good evaluation has three of them, from a test of the logic to a comparison of costs and response time.

At the printing company: the assistant gets ten known order emails. Only when it fills them all in correctly may it read real emails.

F

1 term

Fine-tuning

Fine-tuning is training an existing model further with your own examples so that it masters a task or a style better. It takes many good examples and computing power. For most companies a good prompt or RAG is enough.

At the printing company: you do not need to train a model to answer questions about your manuals. RAG does that.

G

2 terms

Generative AI

Generative AI creates something new by itself based on your instruction, such as a text or an image. ChatGPT and Claude are well-known examples.

At the printing company: you have a first version of a customer email written and finish it yourself.

User

GPU

A GPU is a graphics card. It computes AI models much faster than an ordinary processor. It is expensive though and it has limited memory.

At the printing company: an AI server with a solid GPU answers as smoothly as a service in the cloud, as long as the model fits in the memory.

Builder

H

4 terms

Hallucination

A hallucination is an answer that sounds convincing but is wrong. A language model predicts words and does not always know whether something is true. So ask for the source and check the figures.

At the printing company: a model invents a paper weight that is not in your stock list. Whoever asks for the source sees that immediately.

User

Hotfolder

A hotfolder is a folder that is watched by software. Put a file in it and a process starts automatically.

At the printing company: a PDF that lands in the prepress hotfolder is checked and passed on to imposition.

Human-in-the-loop

Human-in-the-loop means that a human looks at every proposal from the AI before anything is done with it. The AI prepares and the human decides.

At the printing company: the AI proposes a setting based on comparable orders. The operator judges the proposal and applies it or not.

Hybrid working

With hybrid working you combine local AI with AI in the cloud. Sensitive data stays inside. Only what may leave goes to the smarter model outside.

At the printing company: a local model anonymises your order data. A model in the cloud then does the analysis.

Builder

I

1 term

Make-ready waste and spoilage

Make-ready waste is the extra paper you allow for setting up the press. Spoilage is the printed matter you cannot sell. Together they determine how much paper an order really costs.

At the printing company: a shift leader asks what percentage of actual waste there was compared to the calculated waste. The AI works that out on the production data.

J

2 terms

JDF and XJDF

JDF stands for Job Definition Format. It is the digital job ticket of the printing industry, managed by the organisation CIP4. XJDF is the simpler successor.

At the printing company: the MIS sends the run length and the paper type of an order as JDF to the press. Nobody retypes them any more.

Read more: What is an API?

JSON

JSON is a simple text format for storing and passing on data. Each item gets a name and a value. People can read it and so can software.

At the printing company: an order in JSON contains lines such as “run length: 5000” and “paper: 135 gsm”.

Builder

L

3 terms

LLM

LLM stands for Large Language Model. It is trained on an enormous amount of text and predicts the next piece of an answer each time. ChatGPT and Claude each run on such a model.

At the printing company: a language model reads a hundred thousand operator notes and groups them per problem.

LM Studio

LM Studio is a user-friendly app with which you download AI models and run them on your own computer. It is a good way to try out local AI.

At the printing company: you install LM Studio on your laptop and test a small model of a few gigabytes with your own questions.

Local AI

Local AI runs on hardware at your own company. So your data does not go outside. What runs on affordable hardware reaches roughly the level of the top models of a year ago.

At the printing company: an AI server in your own building answers questions about orders without customer data leaving the company.

Builder

M

5 terms

Machine learning

With machine learning software learns patterns from examples instead of from fixed rules. The more good examples, the better the prediction.

At the printing company: a model learns from sensor data how a bearing vibrates when it is almost worn out.

Machine data

Machine data is what a machine keeps track of itself: speeds, counters, stoppages and fault codes. Modern presses log that down to the minute.

At the printing company: the web presses at Drukkerij Moderna record every stoppage and every web break.

MCP

MCP stands for Model Context Protocol. It is an open standard that connects an AI assistant to your programs and your data. An MCP server is the piece of software that opens up a source, for example your calendar or your MIS.

At the printing company: through an MCP server on your MIS your assistant requests this week’s orders itself.

Feedback loop

The feedback loop is the path from a note on the floor to an answer coming back. Whoever reports something and never hears anything stops reporting.

At the printing company: at Drukkerij Moderna the logging stopped in 2025. A weekly triage with an answer to the operator should close the loop again.

MIS

MIS stands for Management Information System. In a printing company it is the software for estimating, orders, planning and invoicing. Almost every AI application needs data from it.

At the printing company: an agent that prepares quotes takes the paper prices and the machine rates from the MIS.

Read more: What is an API?

N

1 term

n8n

n8n is a tool with which you build workflows by connecting blocks to each other. Each block does one step, such as reading an email or sending a message. You can run it on your own server.

At the printing company: if a press stands still for more than ten minutes, a workflow in n8n sends a message to the shift leader.

O

4 terms

OEE

OEE stands for Overall Equipment Effectiveness. The figure shows how much of the planned time a machine really delivers good work. It drops through stoppages and through running slower. Rejected work counts too.

At the printing company: a press with an OEE of sixty percent delivers only six hours of good work in ten planned hours.

Changeover time

Changeover time is the time between the last good sheet of one order and the first good sheet of the next. With short runs it weighs heavily on the cost price.

At the printing company: one of the proposed follow-up trials lets AI suggest settings based on comparable orders from the past.

Open-weight model

With an open-weight model the parameters are free to download. So you can run it yourself on your own hardware. On the Hugging Face platform there are more than three million of them, including from Europe.

At the printing company: you download a model from the French company Mistral and run it on your own AI server.

Builder

Operator note

An operator note is a short note an operator types in at the machine about what went wrong. The machine logs a fault code. The operator describes what he sees and smells.

At the printing company: at Drukkerij Moderna printers have typed more than a hundred thousand notes since 2001. The machine data never contradicted them.

P

4 terms

Parameters

Parameters are the numbers an AI model consists of. The more parameters, the more the model can do and the more memory it needs. A model of nine billion parameters takes up about ten gigabytes.

At the printing company: a laptop with sixteen gigabytes of memory runs a small model. For a larger model you need an AI server.

Builder

Pilot phase

The pilot phase is the period in which an AI application is tested on a small scale. Many projects get stuck there. A step-by-step plan with fixed evaluation helps to grow into everyday use.

At the printing company: for a month the assistant only processes the order emails of two regular customers.

Predictive maintenance

Predictive maintenance uses data from sensors to see wear coming. You plan the maintenance before something breaks instead of after.

At the printing company: the vibrations in a bearing of the press increase. The system warns the maintenance department before the press stops.

Prompt

A prompt is the question or instruction you give to AI. The clearer you say what you want and who it is for, the better the answer.

At the printing company: “Write a reply to this complaint about colour difference. Stay businesslike and propose a reprint.”

Proof-of-concept

A proof-of-concept or POC is a small trial that shows whether an idea works in practice. You test with real data first before you invest.

At the printing company: the first POC of PRINTELLIGENCE investigated what AI gets out of twenty years of operator notes from Drukkerij Moderna.

Q

1 term

Query

A query is a question to a database, written in a fixed language such as SQL. AI can translate an ordinary question into a query itself.

At the printing company: you ask how many web breaks there were this week. The AI writes the query and shows the answer with the source alongside.

Builder

R

2 terms

RAG

RAG stands for Retrieval Augmented Generation. The system first looks for the matching passages in your documents and only then writes an answer. That way the answer rests on your sources and not on the memory of the model.

At the printing company: an operator asks how to clean the washing unit. The answer comes from the manual of the company’s own press.

Builder

RAM

RAM is the working memory of a computer. With local AI the amount of RAM determines how large the model may be. The speed of that memory determines how smoothly it answers.

At the printing company: a model of thirty gigabytes does not fit on a laptop with sixteen gigabytes of RAM.

Builder

S

3 terms

Speech to text

Speech to text converts spoken words into written text. There are open models for it that you can run locally.

At the printing company: an experienced printer explains how he solves a problem. The recording becomes text that colleagues can search later.

Downtime code

A downtime code is the reason a machine or an operator attaches to a stoppage. Without a code you know how long a machine stood still but not why.

At the printing company: at Drukkerij Moderna the screen asks what is wrong when the press stands still for five minutes without a note.

System prompt

The system prompt is the fixed instruction an assistant is given beforehand. It determines its role and what it may and may not do.

At the printing company: “You are the quoting assistant of our printing company. Answer briefly and never quote a price without a source.”

T

1 term

Token

A token is a piece of a word. Language models read and write in tokens. Whoever uses a model through an API pays per token. The speed of a model is expressed in tokens per second.

At the printing company: the well-known services in the cloud write forty to eighty tokens per second. An ordinary laptop with a local model often manages only about ten.

User

V

1 term

Vibe coding

With vibe coding you describe in plain language what you want and AI writes the code. You test the result and adjust until it is right. That way you go from an idea to a working example in hours.

At the printing company: an estimator builds a small tool himself that sorts incoming emails per customer.

W

2 terms

Webhook

A webhook is a message one system automatically sends to another system as soon as something happens. So the other system does not have to go and check each time.

At the printing company: as soon as an order in the MIS switches to “printed”, finishing gets a signal.

Read more: What is an API?

Builder

Workflow

A workflow is a fixed path of steps that runs the same way every time. That is reliable but it remains automation: the software decides nothing itself.

At the printing company: every new order in the webshop automatically creates a job and sends a confirmation to the customer.

AI terms in plain language

A

AI Act

AI agent

AI assistant

AI policy

AI governance

Read-only connection

Anonymisation

API

API key

Atomic agent

Atomic task

C

Chatbot

Chunking

Claude Code

Cloud AI

Cobot

Computer vision

Context window

Custom GPT

D

Dashboard

Digital work instruction

Docling

E

Edge AI

Embedding

ERP

Evaluation gate

F

Fine-tuning

G

Generative AI

GPU

H

Hallucination

Hotfolder

Human-in-the-loop

Hybrid working

I

Make-ready waste and spoilage

J

JDF and XJDF

JSON

L

LLM

LM Studio

Local AI

M

Machine learning

Machine data

MCP

Feedback loop

MIS

N

n8n

O

OEE

Changeover time

Open-weight model

Operator note

P

Parameters

Pilot phase

Predictive maintenance

Prompt

Proof-of-concept

Q

Query

R

RAG

RAM

S

Speech to text

Downtime code

System prompt

T

Token