AI Fundamentals

This chapter covers the basics of AI: what AI can do, where its limits lie, and how its results are checked. These basics help in the daily work with agents, workflows, and automations. Agents are AI programs with a fixed task, their own knowledge, and their own abilities (see Understanding Agents).

How AI Works

Artificial intelligence (AI) stands for computer programs that take over tasks that typically require human abilities: recognizing patterns, understanding and producing language, classifying documents, finding content similarities, or extracting information from images and documents.

Modern AI does not get these abilities from rules that developers program. Instead, an AI model learns patterns from very large amounts of data and applies what it has learned to new tasks.

For different tasks, there are different types of AI models:

• Language model – an AI model that understands and produces text: It answers questions, summarizes documents, and drafts texts.
• Multimodal model – an AI model that processes different types of input – for example, text and images. It describes, for example, the content of a photo or answers questions about an image.
• OCR model – an AI model that reads text from documents and images. OCR stands for Optical Character Recognition.
• Embedding model – an AI model that recognizes content similarities between texts. It enables semantic search in the company knowledge, which searches by meaning instead of exact words.

Four properties apply to every language model – and thus also to working with incoreon AI:

• Learned, not specified. A model's behavior stems from learned patterns – not from a specification that anyone could fully check and guarantee.
• Plausible, not verified. A language model cannot look anything up on its own – the capabilities and knowledge of an agent, for example, make this possible. The model still produces the answer itself: the statistically most probable one. Plausible is not the same as correct.
• Not fully reproducible. The same question can lead to different answers.
• Fluently worded, regardless of correctness. Answers sound convincing through their wording. There is no reliable signal for when the model is uncertain.

Opportunities and Risks

AI helps with wording, takes over recurring work, and opens up company knowledge. At the same time, risks arise that need to be kept in mind.

What AI brings:

• Drafting, summarizing, and revising texts – for example e-mails, reports, and responses
• Automating operational processes so that they run without manual processing
• Opening up company knowledge so that answers are created from own documents
• Handing over recurring tasks so that time remains for other work

What can go wrong:

• Hallucination – an answer that sounds convincing but is wrong or invented
• Bias – systematically one-sided results that stem from one-sided patterns in the training data
• Blind trust – errors that are adopted unchecked
• Data protection – sensitive data that is processed in third-party AI services outside the company's control

People Decide, Agents Support

Agent results are suggestions, not decisions. Agents prepare drafts, summaries, and decision support – the decisions are made by people.

AI outputs can be wrong and still sound right at the same time. An agent that processes business data can invent information. The greatest danger is not the error itself, but adopting it unchecked: Undetected errors end up in decisions, responses to customers, and results.

AI outputs are therefore a suggestion – like a first draft: The AI takes over the bulk of the work; the responsibility for the result remains with the person using the output. Check everything that flows into a decision, a response to a customer, or a business data record against the actual data before use.

Agents do not always answer the same question identically. If answers do not fit, use (“Try again...”) or formulate the request more concretely (see Reviewing and Refining Answers).

When Agents Wait for a Person

incoreon AI combines self-directed action by agents with human control. Two things determine what an agent executes on its own and where it waits for a person: the agent's instruction and the tasks of the automation.

Automations run without a manual start: “Inbox” starts the agent for every incoming e-mail of a mailbox, “Schedule” starts it at set times (see Automation and Workflows).

The results of an automation reach the responsible people as to-dos or information items: To-dos wait for processing or a decision, information items keep everyone up to date on everything new. The tasks of the automation determine who receives to-dos and information items (see ToDos and Information).

If a workflow needs a decision or processing, it waits for the responsible person. Only the processing by the person continues the workflow: For a decision to-do, the chosen option goes back to the agent. For a process to-do, the chat history of the processing goes back to the agent. In technical terms, this principle is called human in the loop: The agent only continues once a person has decided or processed something (see How a Workflow Runs).

It is recommended to know the own configuration: What does the platform process on its own – and where does it wait for an explicit approval? Those who know this keep control over the autonomous actions of the agents (see Defining Tasks for Automation and Agent Components in Detail).

Ways to Stay in Control at a Glance

Agents work in the chat on request or run automatically – in both cases, their work can be controlled. The most important options:

• Pausing automations – every automation can be paused and resumed individually; inactive automations do not run (see Managing Automations).
• Setting capabilities and tools per agent – capabilities can be switched on or off per agent. Custom capabilities are connected via MCP servers; MCP is a standardized connector for additional functions and external systems – comparable to a USB port. Such a capability includes various tools, each an individual function. Every tool can be activated or deactivated (see Configuring Capabilities and Guide: Adding a Custom MCP Server).
• Own MCP servers only with permission – custom MCP servers can only be added to an agent once the user group that manages the agent has the corresponding permission. Global admins grant this permission with the “Allow your own MCP servers” option (see Allowing your own MCP servers).
• Setting a cost limit per agent – when an agent reaches its monthly cost limit, incoreon stops the agent's work; the interface shows the reason (see Behavior When the Cost Limit Is Reached).

Legal Framework

The EU AI Act – the regulation of the European Union on artificial intelligence – sets the legal framework for everyone who provides or uses AI systems. It prohibits practices that harm or disadvantage people – for example deceptive manipulation or the rating and disadvantaging of people by their social behavior. incoreon does not support prohibited practices. Which practices are prohibited in detail is stated in the full text of the regulation.

These prohibitions apply to all agents – including self-created ones.

The EU AI Act also provides that companies using AI systems promote their employees' competence in working with AI. This documentation supports this: It explains what AI can do, where its limits lie, and how results are checked. On request, incoreon supports companies in training their employees.

Help and Feedback

Questions and problems are covered in the Troubleshooting and Support chapter – with an error ID, the support speeds up processing (see Error Messages and Their Meaning). Feedback and reports on the effects of the AI are received via the existing support channels.