Glossary

Of a system that does not just answer, but decides which tools to use and in what order to reach a goal. The difference with a prompt is that the path is not written by the person asking.

Where we discuss it: AI agent orchestration

Systems that do tasks that used to need a person, such as reading a document or answering a request. On this site it nearly always means language models (LLM) connected to a company's data and tools.

The way one piece of software offers its data and functions to another, with precise rules on what can be asked and what comes back. It is what lets the management system talk to the online shop without anyone retyping data by hand.

Selling to or serving companies rather than end consumers (B2C). It changes the product: more users per customer, different roles, contracts and connections to their systems.

The measure of how much the content moves while the page loads. Below 0.1 is considered good; you get there by reserving the space for images before they arrive.

The record of who is a customer and who might become one: contacts, deals, quotes, exchanges. It earns its place when the same information lives in three inboxes and nobody can say where a deal stands.

Every moment a customer deals with a company, from first contact to support after the purchase. UX is about one product, CX about the whole relationship.

Whoever uses an artificial intelligence system under their own authority, in the course of a professional activity. Not whoever builds it: whoever puts it into service. It is the role the AI Act places most of the practical obligations on.

Where we discuss it: EU AI Act, a guide for deployers

The system that holds a company's administrative and operational processes together: orders, stock, purchasing, invoicing, sometimes production and people. Custom software usually grows where the standard ERP stops, not in its place.

A ready-made code base that gives an application its structure and solves the common problems, so the work goes into the parts that are yours alone. Choose it with care: it is rarely changed, and changing it is expensive.

Regulation (EU) 2016/679, applied since 25 May 2018, which sets how a company may collect, use and keep personal data. For software it means deciding from the design stage which data is needed, where it lives and who can see it.

A constraint that limits what an AI system can do or answer, applied before the output reaches the person who asked for it. It is not a suggestion inside the prompt: it is a check that can refuse an answer.

Where we discuss it: AI workflows and agents

A point in the flow where a person checks or approves before an AI system acts, for example before a payment or an email to a customer. It belongs wherever a mistake costs more than the time a check takes.

The function that looks after the health and safety of workers and the protection of the environment, and the documents it produces, that is risk assessments, procedures, registers and reports.

A language model trained on large amounts of text. It knows a lot in general and little about a private domain: that is the reason RAG exists.

Where we discuss it: Custom RAG agents with Mastra

An open standard that lets an AI model use outside tools and data, such as a management system or a document archive, always in the same way. A service exposed over MCP once can be used by different assistants.

The system that runs production while it happens: work orders, progress, machine stops, scrap. It sits between the ERP, which says what to produce, and the machines that produce it.

The calculation of what is needed to produce and when to order it: materials, quantities, lead times. It is the part of the ERP that turns a production plan into a purchase list with a date on it.

An architecture where a single installation serves several customers while keeping their data separate. The alternative is one copy per customer, which costs less to design and far more to keep up to date.

Where we discuss it: Digital Factory

The agreement in which two parties undertake not to disclose what they tell each other. We sign it before talking about a project when the project itself is confidential.

The coordination of several agents, tools and steps in a flow that has a state, an order and checkpoints. You need it when a task does not fit into a single call to the model.

Where we discuss it: AI workflows and agents

The data that leads back to a person: name, email, phone number, device identifiers, sometimes an IP address. Recognising it is the first step in deciding where it may live and for how long.

The instructions and context given to a language model to get an answer. In a product it is treated as code: versioned and tested at every change.

An approach where the model, before answering, retrieves the relevant passages from a data source made available to it, and answers on those. It cuts down made-up answers because it ties the output to a text that exists.

Where we discuss it: Custom RAG agents with Mastra

Granting permissions by role rather than by person: what a role can do is described once, and people inherit the permissions of the role they hold. Moving somebody to another job becomes one line instead of a review of every permission.

The archive where a project's code lives with its whole history: every change, who made it and when. Having it in your name lets you change supplier without starting from scratch.

How much a project returns compared with what it cost. For software it is measured on the process that changes, in hours saved or new revenue, not on the technology.

The Amazon Web Services service for storing files in the cloud, often used as the starting archive for documents and data.

An architecture where each customer has a dedicated, standalone installation with their own data. It costs less to design but much more to update, in exchange for the highest level of customization and security.

Where we discuss it: Digital Factory

The written commitment on how the service runs: how fast a report is answered, how fast it is resolved, how much unavailability is acceptable. Without numbers it is not an agreement, it is a promise.

In the CMS, a typed content block that writers put together to build a page. Each slice matches a component in the design system: it is the point where editorial work and design meet.

Companies whose headcount and turnover stay below the thresholds set by the European Commission.

The language used to query databases made of tables. It answers counts and sums exactly, which a search over texts does not do.

A young company building a new product, often software, and looking for a way to grow fast. The product changes often, so it has to be built so that it can be changed without starting over.

The unit in which a language model reads and writes text, usually a piece of a word: models are billed per token. In security the same word means a temporary key that proves who you are, like the access token you get after logging in.

How easy and clear a product is for the people who actually use it, from first login to finished task. It is designed before the screens are drawn, not after.

The international guidelines for the accessibility of web content, organized in three conformance levels: A, AA and AAA. This site conforms to level AA and meets a good part of level AAA: the status criterion by criterion is in the accessibility statement.

Where we discuss it: Accessibility