AI agents for businesses

From chatbot to agent

The first wave of generative AI in companies consisted of chat: employees ask a question, the AI answers, and the human does the rest. AI agents go one step further. They receive a goal instead of a single question, plan the necessary steps, use tools such as email, calendar, file storage or CRM and deliver a result – for example a finished report, a pre-sorted inbox or a merged Excel file.

This is interesting for mid-sized companies because many processes are too small for an IT project but too frequent for manual work: the weekly status report, gathering documents before a customer meeting, pre-qualifying incoming enquiries. That is exactly where agents deliver their value.

What makes an AI agent

An AI agent typically consists of four building blocks:

• Model: A Large Language Model understands the task, plans and formulates.
• Instructions: A description of role, goal, rules and desired result – carefully formulated according to the principles of prompt engineering.
• Knowledge: Documents and data sources the agent is allowed to access, often via an AI knowledge base.
• Tools: Interfaces to systems through which the agent reads or acts. Open standards such as the Model Context Protocol (MCP) make connecting them easier.

In addition, there is a trigger: an agent starts on request in the chat, at a fixed time or through an event, such as a new email.

Agent, workflow or automation?

Classic automation follows fixed rules: if X, then Y. That is reliable, but fails with unstructured input such as freely worded emails. An agent decides for itself, within its framework, which steps are necessary. In practice, a mix has proven effective: fixed processes in which individual steps are carried out by AI, and clear limits on when a human decides.

Typical fields of application in mid-sized companies

Recurring reports

Every Monday, an agent gathers information from defined sources and creates a summary: open quotes, important appointments for the week, anomalies in exports. The human checks and adds to it instead of starting from scratch.

Email and enquiries

Incoming messages are read, sorted by concern and given draft replies that a person checks and sends. In the back office in particular, this saves time every day.

Preparing meetings

Before a customer meeting, an agent compiles what is known about the customer: recent quotes, open topics, notes from previous conversations. The article AI in sales: 7 use cases shows examples from sales.

Processing documents and data

Agents summarise new documents, extract key figures, compare versions or merge spreadsheets from several sources. With an AI worker on a computer, local files, network drives and programs without an interface can also be reached – more on this in the article AI Worker: delegating tasks to the computer.

Research

An agent searches websites and internal sources on a question, for example about competitors, tenders or suppliers, and delivers a structured overview with sources.

Keeping knowledge up to date

Agents can also help maintain company knowledge: summarise new minutes and file them in the wiki, flag outdated documents for review, or compile a list of missing content from recurring questions. The knowledge base thus benefits from automated routine work, and experts only decide what is adopted.

Control, security and responsibility

The more an agent is allowed to do on its own, the more important clear limits become. Five principles have proven effective:

1. Least privilege: An agent only gets access to the sources and tools it needs for its task – and never more than the responsible person is allowed to see.
2. Approvals for critical steps: Sending, deleting, payments or changes in line-of-business systems should require explicit confirmation as long as the agent has not demonstrably worked reliably.
3. Traceability: Every run should be logged: what was triggered, which steps ran, what result was produced?
4. Visible plan: For multi-step tasks, a plan helps to track progress and intervene in good time.
5. Human responsibility: Agents' results remain proposals until a human approves them. For certain purposes, this is also a requirement of the EU AI Act.

Data protection for agents

Agents often process personal data from emails, calendars and CRM. The same requirements apply as for any use of AI: a legal basis, a data processing agreement, transparency about model providers and data flows, and a permissions concept. In addition, you should document which agents access which systems. The guide GDPR-compliant AI provides details.

Introducing an agent in six steps

1. Select a task: Frequent, clearly defined, with a verifiable result – such as the weekly project report.
2. Define the result: Describe exactly what the result should look like, ideally with an example.
3. Define sources and tools: Which documents, inboxes or systems does the agent need? What may it only read, and what may it also change?
4. Write and test instructions: Formulate role, rules and format. Test with real cases, including difficult ones.
5. Define approvals: Determine which steps need confirmation and who grants it.
6. Monitor operation: Check every run in the first few weeks, collect errors and improve the instructions. Only then is greater autonomy worthwhile.

Measuring the benefit

Agents pay off when the time saved exceeds the costs of setup, model usage and oversight. Record in advance how long the task takes today and how often it occurs. Compare after a few weeks. The article Calculating AI ROI for mid-sized companies offers a calculation scheme; you will find further examples of process automation under AI agents for process automation.

What to look for in an agent platform

• Easy creation: Departments should be able to create agents with instructions, documents and tools without programming.
• Triggers: Start on request, on a schedule and through events.
• Integrations: Connectors to email, calendar, file storage and CRM, plus open interfaces such as MCP.
• Permissions and approvals: Agents follow the organisation's roles; critical actions require confirmation.
• Logs: Every run is traceable.
• Reuse: Processes can be shared and versioned as templates or skills.
• Model choice: A fast or a powerful model, depending on the task.

Example: an agent for the weekly project report

A typical example shows how the steps work together. The project management team at a service provider creates a status report for management every Friday and regularly needs two hours for it.

1. Goal: A one-page report with progress, risks and decisions required for each project, every Friday at 9 a.m.
2. Sources: The project folder with minutes and plans, the project management inbox for the current week and the task management system.
3. Tools: Read access to the folder, inbox and tasks; creating a document in the shared folder.
4. Instructions: Fixed structure, a maximum of five points per project, risks backed up with a source, no assumptions – missing information marked as an open question.
5. Approval: The agent only creates a draft; it is sent only after review by the project management team.
6. Oversight: In the first four weeks, the project management team compares the draft with its own assessment and improves the instructions.

Result: in the best case, two hours of gathering and writing become twenty minutes of checking and adding. Only your own measurement shows how much time is actually freed up.

Limits of AI agents

Agents are powerful, but not infallible. Plan for these limits:

• Errors propagate: A misunderstood first step can distort all subsequent ones. Intermediate results and visible plans help to intervene early.
• Unclear tasks produce unclear results: The more precisely goal, format and rules are described, the more reliably the agent works.
• Hallucinations: Agents too can invent facts if sources are missing. Require evidence and have gaps marked.
• Costs: Multi-step tasks consume more model capacity than a single question. Choose the model to suit the task and keep an eye on usage.
• Systems without an interface: Some line-of-business applications do not offer an API. Agents that can operate a computer help here – with particularly careful approvals.

Roles and responsibilities

Every productive agent needs a responsible person from the department. They define the result, check the runs and decide on changes. IT is responsible for access, integrations and permissions; data protection reviews which data the agent processes. Record these responsibilities in a short overview – it is also part of your documentation for the works council (Betriebsrat) and supervisory authorities.

AI agents with AIPE

In AIPE, teams create their own agents with instructions, documents and tools. In Agent Studio, you design automations as a flow, test and publish them and track every run. Routines execute prompts on a schedule – recurring according to a timetable or once at a set time. Processes can be captured as runbooks, and Claude-compatible skills can be imported, shared and versioned. Agents use activated tools from connectors such as Gmail, Google Calendar, Google Drive or HubSpot as well as configured MCP servers – with the organisation's permissions.

For tasks on the computer there is the AIPE Worker: a Mac or Windows PC takes over tasks involving files, terminal, browser and programs. In "Ask first" mode, you confirm risky actions; with "Free hand", AIPE works without individual confirmations. Every multi-step task starts with a visible plan, and approvals reach you in the chat and on your smartphone.