---
title: "Deploying an AI Agent in Your Company: How to Succeed with the Rollout"
description: "Introducing AI agents in your company: prepare the data foundation, choose a pilot use case, define approvals, bring the team along – step by step."
type: "wissen"
product: "agent"
slug: "ki-agent-im-unternehmen-einsetzen"
source_language: "de"
target_languages: ["de", "en", "es", "pl", "tr"]
published: "2026-06-10"
status: "publish"
faq_json: [{"q":"How do I introduce an AI agent in my company?","a":"In four steps: organize the data foundation, choose a clearly defined pilot use case, define guidelines and approvals, then evaluate with the team and expand step by step."}, {"q":"Which use case is suitable as a pilot?","a":"Recurring tasks with a clear pattern and low risk – such as checking documents for deadlines or regular summaries. Quick benefit, manageable consequences."}, {"q":"What is the most common hurdle during the rollout?","a":"Scattered data: when contacts, tasks, and documents are in separate tools, the agent lacks context. The data foundation comes before the agent."}, {"q":"How do I ease the team's concerns about AI agents?","a":"Through transparency and the approval principle: the agent prepares, humans decide on everything with external impact. This makes it an assistant, not a replacement."}, {"q":"What does getting started cost?","a":"At webRichtung, account and users are free; you only pay for usage via Credits (1 Credit = 1 Euro net). The pilot needs no project budget."}]
language: "en"
source_id: "wissen/ki-agent-im-unternehmen-einsetzen"
source_hash: "31e8295108d8052773479c7eb85b939b1526f29de6825027e8efd0d27b918746"
---

Deploying an AI agent in your company works best in four steps: first organize the data foundation, then start with a clearly defined pilot use case, properly define guidelines and approvals – and only then expand step by step. Companies that do it the other way around and start with grand ambitions on unstructured data mainly produce disappointment.

## Step 1: The Data Foundation Comes Before the Agent

An agent is only as good as the context it can access. If contacts are in one tool, tasks in a second, and documents in a third, it can't establish connections. The first step of the rollout is therefore unspectacular: bring the relevant data into one place and structure it. The article [Dissolving Data Silos](/en/wissen/datensilos-aufloesen.html) explains why scattered systems are the core problem.

## Step 2: Choose a Pilot Use Case

For the start, choose a task with three characteristics:

- **Recurring:** The benefit adds up with every run.
- **Clearly defined:** It's clear what a good result looks like.
- **Low risk:** Mistakes are annoying but not costly – so initially no tasks with direct external impact.

Proven candidates: checking documents for deadlines, creating regular summaries, preparing tasks from cases. The article [AI Agents: 7 Examples](/en/wissen/ki-agent-beispiele.html) provides concrete ideas.

## Step 3: Define Guidelines and Approvals

Now the agent gets "set up" – in two senses:

- **Guidelines:** Describe not only what it should do, but under what conditions. Example: only create deadlines when date, obligation, and source are clearly documented; report uncertain cases instead of creating them.
- **Approvals:** Determine what the agent may do independently and where a human decides. As a principle, it has proven effective: anything with external impact – emails, calls, master data changes – goes through an approval.

In [webRichtung agent](https://www.webrichtung.de/module/agent/), this is exactly the architecture: you activate automations individually, with adjustments like a minimum confidence level; whatever awaits your decision is collected in the inquiries section – you review the suggestion together with its context and approve or reject it.

## Step 4: Bring the Team Along and Expand

The biggest hurdle is rarely the technology, but acceptance. Three things help:

1. **Transparency:** Show the team what the agent does and does not do – the approval principle makes it an assistant, not a replacement.
2. **Evaluation:** After four to six weeks, look at the pilot together: what did it handle, where did it go wrong, which guidelines are missing?
3. **Gradual expansion:** Only once the pilot holds up do further use cases get added – with the same rules.

## Low Barrier to Entry

For the pilot, you don't need a project budget: at webRichtung, account and users cost €0, you only pay for actual usage via Credits (1 Credit = 1 Euro net). This lets you test the benefit of an AI agent in real operation before larger decisions are due – and the risk of the rollout stays as small as the pilot itself.