AI terms for everyday operations, without the jargon.

Thirty-nine terms as we use them in conversation, in three sentences each: what it is, why it matters for your company and how we deal with it.

Sketch of a thick reference book on a lectern with thumb tabs, a bookmark ribbon and a magnifying glass beside it.

Agent: An AI program that does not just answer but carries out steps itself: looking something up, filling in a field, drafting an email and then deciding what comes next. An agent can do more than a chat window, but it can also get more wrong, because it acts instead of merely suggesting. We only use agents within a workflow with a fixed rulebook: what it may do is defined in advance, what it does is in the log, and anything unclear goes to a person.
AI (artificial intelligence): The umbrella term for software that solves tasks that used to require human judgement; in day-to-day operations we almost always mean language models that read, classify and write texts. What counts for your company is not how clever a model is, but whether the work is then done, checked and in your system. That is why we only use AI as part of a workflow with rules, a check and a log, never as a chat window on its own. More on this: Why ChatGPT has changed nothing in the business.
AI literacy: The obligation under Article 4 of the EU AI Act, in force since February 2025, for employees who use AI systems to know enough about how they work and where their limits lie. It applies to every company that uses AI, and therefore also to Copilot at the desk and not only to high-risk systems. For our workflows we contribute to it through the handover: your team learns what the workflow does, what it puts forward and when a person has to step in. More on this: Checklist for introducing AI.
Allowance: A fixed number of hours per month included in ongoing operation. It covers small extensions, new rules or questions from your team, without requesting a quote each time. We state the scope in our proposal for ongoing operation, and anyone who wants more handover or training uses hours from the allowance for it. More on this: Ongoing operation.
Approval: A person's decision that an output of the workflow may be used: a text published, a record overwritten, a reply sent. It is the point at which responsibility stays within your company, and it corresponds to what the EU AI Act means by human oversight. We define with you what has to be approved and what the workflow may decide on its own.
Chat window: The interface in which a person asks an AI a question and gets an answer, as with ChatGPT or Copilot. That is useful for individual work at the desk, but not for work at volume, because someone has to copy, check and enter every answer into the system. We do not build chat windows for their own sake, but workflows that work in your systems without a window; where customers ask questions, there is a knowledge base behind it. More on this: Why ChatGPT has changed nothing in the business.
Check: An automatic test of every output before it moves on, such as whether all mandatory fields are filled, whether figures match the source or whether a text contains an impermissible statement. Checks are the reason a workflow can write a thousand texts without anyone having to read each one. We build several of them into every workflow, and anything that does not clearly pass goes to a person. More on this: Thousands of product texts with AI.
Cloud under contract: The operating mode in which models from providers such as Anthropic, OpenAI or Google are used directly through their interface, with a data processing agreement and an undertaking that inputs are not used for training. Today it delivers the strongest results for language and is the quickest to set up. We use it for product texts, translations and data without personal information, and for personal data only after a written assessment on your side. More on this: Data sovereignty.
CRM (customer relationship management): Customer relationship management, the system in which your sales team manages customers, contacts, enquiries and sales opportunities, such as Salesforce, HubSpot or Microsoft Dynamics. For AI workflows it is usually source and destination at once: enquiries are created there, data is read from it, and its data quality decides the result. We build into your existing CRM, respect its request limits and put the data in order beforehand where necessary. More on this: Putting data in order.
Data flow: The route data takes in a workflow: from which system, through which model, back into which system, and what never leaves your company along the way. It is the basis for your record of processing activities and for the question of whether your data protection officer and the works council can give their consent. We sketch it for each workflow in the analysis and disclose it before you sign. More on this: Data flow per service.
Data processing agreement (DPA): The contract under Article 28 GDPR by which a service provider may process personal data only on your instructions. It is mandatory as soon as a service provider or model provider gets to see personal data from your company, and it is the first thing your data protection officer will want to see. We conclude it before the first access to data and list the sub-processors with their data centres for each workflow. More on this: Data sovereignty.
Data sovereignty: The ability to decide for yourself which of your data flows where, who sees it and what it is used for. For a mid-sized company with customer lists, pricing terms and engineering know-how, this is not a formality but the basis of the business. We put it in writing in every contract: No customer data leaves your systems without your written decision, and none of your data ever trains a third-party model. More on this: Data sovereignty.
Duplicate: A record that describes the same company, person or product a second time, often with small differences in spelling or address. Duplicates distort evaluations, double up mailings and make any AI working on the data work with contradictory information. We merge them according to rules that you help decide, and put borderline cases before a person instead of guessing. More on this: Before AI is let into the CRM.
Embedding: A sequence of numbers that represents the meaning of a passage of text in such a way that similar content lies close together. This lets a system find matching passages even when different words are used, such as "delivery time" and "when will the goods arrive". We use embeddings for knowledge bases and for matching product data, and compute them on your own servers for sensitive data. More on this: Vector database.
ERP (enterprise resource planning): Enterprise resource planning, the system for items, stock, orders, purchasing and invoices, such as SAP, NetSuite, Microsoft Dynamics 365 or proALPHA. It holds the master data that product texts, quotations and enquiries draw on, and often the sprawl of many years too. We read and write through the ERP's documented interface, never through the user interface, and log every transfer. More on this: Systems.
EU AI Act: The EU AI Act, Regulation (EU) 2024/1689, in force since August 2024, which classifies AI systems by risk and attaches obligations to that classification. For mid-sized companies, what matters most are the transparency obligations under Article 50, AI literacy under Article 4 and the question of whether a workflow is a high-risk system. We classify every workflow in the analysis and fulfil the obligations through rules, a log and approval; this is not legal advice. More on this: The EU AI Act for mid-sized companies.
Fact anchor: A checked source to which a model is bound when writing, such as the actual product range from the CRM or the manufacturer's data sheet. Without an anchor, a model invents plausible details; with an anchor, it may only say what the source provides. We use fact anchors for all texts that contain products, figures or promises, and the check verifies that the text sticks to them. More on this: Thousands of product texts with AI.
Fixed price: A price that is set before the work begins and does not depend on hours worked. For you this means: the risk of a misjudgement lies with us and not in your budget. We charge for the analysis and the build at a fixed price, and we state the price for the build as a binding figure in the analysis, after we have seen the workflow. More on this: Approach and prices.
Glossary (specialist glossary): A list of your technical terms with the binding translation for each language, in other words what a product characteristic is called in French and what it is not called. Without a glossary, a model translates the same term one way one time and another way the next, and your customers read three names for one component. We build the glossary with your business department at the start of a translation workflow and check every translation against it. More on this: Texts and content at scale.
Hallucination: A statement by a language model that sounds convincing but is wrong, such as an invented standard, a wrong dimension or a product that does not exist. In a chat window the reader may notice; in a workflow with a thousand texts nobody does, which is why it is the biggest risk when working at volume. We counter it with fact anchors, rules and checks that test every output against the source, and with a log that makes every passage traceable. More on this: Thousands of product texts with AI.
High-risk system: An AI system that the EU AI Act classifies as particularly risky because of its purpose, for example under Annex III when filtering and assessing job applications or in decisions on promotion and dismissal. For such systems, strict obligations on risk management, documentation, logging and human oversight apply from December 2027 (after the postponement by the Digital Omnibus in 2026). We build HR workflows so that they record applications in a structured way but do not filter or assess them, and clarify the classification with you before building; this is not legal advice. More on this: The EU AI Act for mid-sized companies.
Interface (API): The technical access through which programs exchange data, for example a workflow with Salesforce, NetSuite or Personio. It decides whether AI can work inside a system or only alongside it, and its limits determine how much a workflow can get through in a day. We get to grips with any system that has an interface, and build in caching and throttling where requests are limited. More on this: Systems.
Knowledge base: A collection of your documents, manuals and answers from which an AI answers questions from customers or employees. It makes expertise available around the clock and in several languages, without every question landing with sales. We build it with source references and with handover to people when things get serious, as at a technology distributor with advice in five languages. More on this: case study on enquiries in five languages.
Log: The record of what a workflow has done: which record, which rule, which result, who approved it. It makes every passage traceable, for your business department, for data protection and for the documentation the EU AI Act may require. What is logged is defined in advance, with the works council if you wish, and the performance of individual employees is not part of it. More on this: Introducing AI with the works council.
Model (language model): The AI system that reads and writes text, such as Claude from Anthropic, GPT from OpenAI, Gemini from Google or open models such as Llama and Mistral. Models differ in capability, price, speed and where they may run, and they are constantly being replaced by new ones. We choose the model for each task and level of sensitivity, do not tie ourselves to any provider and build workflows so that the model remains an interchangeable component. More on this: AI operation and monitoring.
On your own servers (on-premise): The operating mode in which an open model runs on your hardware or on a dedicated server in Germany that only you and we can reach. No data leaves your network, but there are set-up and computing costs, and for difficult individual cases the models are weaker than the large cloud models. We recommend this operating mode for personal data, contracts and anything confidential. More on this: On your own servers, in Europe or in the cloud.
Open model: A language model whose weights are freely available, so that it can be run on your own hardware, such as Llama or Mistral. It is the way to use AI without data leaving your company, and there are no fees per request. We use open models for sensitive data and language at volume, and say openly where a large cloud model delivers the better result. More on this: Three operating modes.
Operating mode: The answer to the question of where a workflow's AI model does its computing: on your own servers, in an EU data centre or in the cloud under contract. It determines which data leaves your company, what the workflow costs and how capable the model is. We recommend the operating mode for each workflow in the analysis, with reasons, and you take the decision in writing. More on this: On your own servers, in Europe or in the cloud.
Parallel operation: The phase in which a new workflow runs alongside the existing work and its results are compared with those of the people doing it. It shows, with figures from your business, whether the workflow delivers what the analysis expected before anyone relies on it. In the first build it usually lasts two weeks, and the workflow only takes over on its own when the figures are right. More on this: The schedule for the first workflow.
PIM: A system for product information such as Akeneo or Pimcore, in which product data, characteristics and texts are maintained centrally and distributed to the website, online shop and catalogues. For manufacturers and dealers with thousands of items, it is where multilingual content and data quality are decided. We fill PIM systems with data from manufacturers' documents, with a fact anchor and specialist glossary, and write back only checked values. More on this: Systems.
Prompt: The instruction that a person or a program gives a language model. A good prompt improves an answer, but it does not turn a model into a reliable workflow, because it knows neither your systems nor your rules nor the volume. In our workflows, prompts are one component alongside rules, checks and a log, and they belong to you after payment, like the code. More on this: Why ChatGPT has changed nothing in the business.
RAG: Short for retrieval-augmented generation: before every answer, the model looks up the relevant passages in your documents and answers only on that basis. This lets an AI answer questions about your products, contracts or manuals without them being trained into the model, and it can name the source. We build knowledge bases on this principle, with short answers, source references and handover to a person when the documents do not provide an answer. More on this: Sales and communication.
Rollback: Undoing a run so that the previous state is fully restored. Anyone who has AI write into a CRM, a PIM or a shop needs this way back, so that an error does not cause lasting damage. We build such an undo function into every workflow that writes back into your systems and test it before the first real run. More on this: Case study: spare parts dealer.
Rulebook: The written rules of a workflow: what it may do, what it never does, which cases it puts before a person and how to recognise a good output. It translates the knowledge of your business department into something a workflow reliably follows, and it remains in place even when the model changes. We define the rulebook with you in week 2 of the build and refine it with every correction from the trial run. More on this: The schedule for the first workflow.
Token: The unit into which language models break down text and by which they bill, roughly a word or part of a word; a German sentence usually has a few dozen of them. Model costs and the amount of text a model can read at once are measured in tokens, which is why the price of a run depends on volume and language. We estimate the consumption for each workflow in the analysis and pass on the model costs without a mark-up. More on this: What determines the price.
Vector database: A database that stores embeddings and finds, in fractions of a second, the passages closest in content to a question. It is the memory of a knowledge base and helps decide whether the AI finds the right passage in your documents. We run it on your server or on a server in Germany, as you choose, so that your documents are not held by a third-party provider. More on this: RAG.
Workflow: A recurring piece of work with a fixed starting point, fixed rules and a result that can be checked, for example: read an enquiry, classify it, create it in the CRM. For you, the workflow is the unit in which effort and saving can be measured, not the project and not the tool. We analyse, price, build and run per workflow, instead of selling use cases from a catalogue. More on this: Services.
Workflow analysis: Three days in which we watch, on your premises, where work arises, who does it and which systems are involved. Afterwards you know which workflows pay off, which do not and what each build costs. We deliver it at a fixed price as a list with effort, saving, data flow sketch and a "does not pay off" column. More on this: Which workflows pay off with AI.
Zweite Schicht: Our name, German for "the second shift", with two readings: the shift that works while your office is empty, and the layer that sits on top of your existing systems and connects them. Both describe what AI in day-to-day operations should achieve: getting work done, not answering questions, and doing it between the systems you already have. That is why we build workflows that run at night and during the day, in your systems, with a report for the early shift. More on this: About us.

Handover

The first step is a 30-minute call.

You tell us about the workflow that costs you the most time. We tell you honestly whether AI pays off there and what the next step would be. Whether a workflow analysis follows is up to you.

Book a call Approach and prices