# The AI Department: Staffing Roles in Your Org Chart with AI

URL: https://andreasindorf.com/en/ai-department-guide/
Author: Andreas Indorf · Published: 2026-11-07 · Updated: 2026-11-07 · German original: https://andreasindorf.com/ki-department-leitfaden/

> How an AI department works: staffing roles with AI specialists, steering autonomy across four levels and keeping human approval on every critical decision.

Most companies introduce AI as a tool: a chat licence here, a copilot there, a pilot project somewhere else. The result is a patchwork in which nobody can say exactly what the AI actually does, who is responsible for it and what it delivers. The AI department starts somewhere else: not with the tool, but with the org chart. This guide explains how the operating model works, how it is controlled and how to introduce it.

**In short:** An AI department staffs individual roles in the org chart fully or partly with AI specialists that work according to real job profiles, from payroll clerk to dispatcher. Your team steers them in a cockpit, sets the autonomy level for each task and approves every critical decision. The introduction starts with an org chart analysis, delivered after two weeks; the first agents are typically working after around six weeks.

## Why "AI as a tool" stalls in many companies

According to the [Federal Statistical Office of Germany (Destatis)](https://www.destatis.de/EN/Themes/Economic-Sectors-Enterprises/Enterprises/ICT-Enterprises-ICT-Sector/Tables/icte-new-1-enterprises-artifical-intelligence.html), 26 percent of companies in Germany used AI in 2025; among companies with 250 or more employees, the figure was 57 percent. Companies that had considered AI but were not yet using it most often cited [lack of knowledge (72 percent)](https://www.destatis.de/EN/Themes/Economic-Sectors-Enterprises/Enterprises/ICT-Enterprises-ICT-Sector/Tables/icte-new-2-enterprices-against-artifical-intelligence.html), followed by uncertainty about the legal implications (62 percent) and data protection and privacy concerns (60 percent).

These three hurdles cannot be solved with yet another tool. They require clarity on three questions:

1. **What exactly should the AI do?** Not "something with AI", but a defined role with tasks, inputs and outputs.
2. **Who is responsible?** Every AI activity needs a human who owns it from a business perspective.
3. **How is it controlled?** Who approves what, how is it logged, how do you stop the AI?

The AI department answers these questions through the structure every company already has: the org chart.

## What an AI department is and what it is not

An AI department is an organisational unit in which AI specialists take over certain roles or parts of roles. Each specialist corresponds to a job profile, such as accounting, dispatching, customer service or marketing, and has clearly described tasks, tools and limits.

It is explicitly **not** a plan to replace people with machines. In practice, the AI specialist usually takes over the repetitive parts of a role, while the human steers, reviews, handles exceptions and carries the responsibility. Many roles do not become smaller as a result, but more demanding.

| Approach | Starting point | Responsibility | Control | Typical result |
|---|---|---|---|---|
| AI as a tool | individual software licences | with the individual user | hardly any, individual | scattered effects, hard to measure |
| AI project | one use case with a budget | project lead until go-live | within the project, unclear afterwards | pilot that often fails to scale |
| Centre of excellence | central team of experts | with the CoE | standards and advice | know-how, but distance from daily business |
| AI department | roles in the org chart | with the business unit | cockpit, autonomy levels, approvals | measurable relief per role |

A centre of excellence and an AI department are not mutually exclusive. The CoE can set standards while the AI department does the operational work. More on this in the article [Centre of excellence for AI (in German)](https://andreasindorf.com/center-of-excellence-ki/) and, for how it fits into the overall organisation, under [AI organisational structure (in German)](https://andreasindorf.com/ki-organisationsstruktur/).

## The building blocks of the operating model

### Specialists by job profile

Instead of one general-purpose assistant, the AI department works with specialised agents. What distinguishes such agents technically from a chatbot and how they use tools is covered in detail in the guide [AI agents in business](https://andreasindorf.com/en/ai-agents-in-business/). Each one knows its tasks, its tools and its limits, just like a job description. This makes results verifiable and errors attributable.

### A cockpit for control

All specialists are steered from one place: assigning tasks, reviewing results, granting approvals, viewing metrics. Without a central cockpit, you end up with exactly the patchwork you wanted to avoid.

### Four autonomy levels

How much a specialist may decide on its own is set per task, not across the board. The levels range from pure suggestion, through a prepared action with approval and execution with sampling, to supervised autonomous operation. New tasks start at the bottom and move up only once quality has been right over a longer period. The levels are explained in detail in the article [AI agents in business](https://andreasindorf.com/en/ai-agents-in-business/).

### Approval gates

At certain points the process stops and a human decides: before customer communication, before payments, before contractual commitments, before changes to production systems. The principle: an approval gate is always held by a human. The OWASP project explicitly names human approval of high-impact actions as protection against the risk of [Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/), meaning too much functionality, too many permissions and too much autonomy.

### Kill switch

Every specialist and the department as a whole must be able to be stopped immediately. This is not just common sense: for high-risk systems, [Article 14](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14) of the EU AI Act requires that humans can safely interrupt the system through a "stop" button or a similar procedure.

### AI Act log

Every action, every approval and every rejection is logged. When in doubt, the log answers the questions of customers, auditors, data protection officers and the Betriebsrat (works council): who decided what, when, and on what basis?

### Portals on a single data foundation

An AI department does not operate in a vacuum. Employees, customers and service providers need access, for example through a team, customer, service provider and onboarding portal. What matters is that all portals work on the same data foundation, so the AI specialists are not dealing with contradictory versions. Where specialists access manuals, contracts or policies, their answers should come with source references from verified documents; how to achieve this in a GDPR-compliant way is explained in the guide [Knowledge AI with RAG in business](https://andreasindorf.com/en/enterprise-rag-knowledge-ai/).

## Which roles are suitable first

Not every role is a good starting point. Four criteria have proven useful:

- **Volume:** The role handles many similar transactions.
- **Clear rules:** There are traceable criteria for a good result.
- **Data access:** The necessary information is available digitally.
- **Limited damage:** An error can be detected and corrected before it becomes expensive.

Good candidates are often found in accounting, dispatching, customer service, inside sales and marketing. Poor starting points are roles with rare, highly individual decisions or with major consequences for individuals, such as recruiting. How to assess candidates systematically is described in the article [AI readiness and potential analysis (in German)](https://andreasindorf.com/ki-readiness-assessment/).

## The introduction: from analysis to the first agent

### Weeks 1 to 2: org chart analysis

We go through your org chart role by role: which tasks arise, how often, with what effort, on what data? The result is a prioritised list of the roles with the greatest leverage, plus quick wins and a roadmap for the next six months. The analysis is delivered after two weeks.

### From week 3: configuring and connecting specialists

The AI specialists are set up for the prioritised roles: tasks, tools, permissions, autonomy levels and approval gates. In parallel, the necessary systems are connected, such as ERP, CRM, mailbox or document storage.

### From around week 6: first agents in operation

The first specialists work in the cockpit; your team reviews and approves. In this phase, measuring is what counts: lead time, error rate, share of approvals without corrections. These figures determine whether a task is moved up a level.

### Afterwards: expand and promote

Only once the first roles are running stably are more added. Proven tasks move up in autonomy; new ones start at the bottom again. How to translate this into longer-term planning is shown in the article [Creating an AI roadmap (in German)](https://andreasindorf.com/ki-unternehmensstrategie/).

## People, works council and literacy

An AI department changes work, and the people affected feel it first. Three points are therefore mandatory:

**Involve people early.** The employees in the affected roles know best where the routine lies and where the exceptions lurk. They belong in the analysis, not just in the training afterwards. Guidance on this in the article [AI change management (in German)](https://andreasindorf.com/ki-change-management/).

**Involve the works council.** As soon as AI systems can evaluate the performance or behaviour of employees, the Betriebsrat's (works council's) co-determination rights under German law are engaged. The log and the clear approval rules of an AI department make these conversations considerably easier. More on this under [AI and the works council (in German)](https://andreasindorf.com/ki-betriebsrat/).

**Build literacy.** Under [Article 4 of the EU AI Act](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4), deployers of AI systems must take measures to support the AI literacy of their staff, tailored to prior knowledge and context of use; following the Digital Omnibus Regulation (EU) 2026/1744, they are not required to guarantee a specific level of literacy. Anyone who steers and approves AI specialists needs to understand where their limits lie.

## Typical mistakes when building one

**Too many roles at once.** Anyone who wants to fill a dozen roles straight away loses track of quality and accountability. Two or three roles that deliver measurable relief within a few weeks, and serve as internal proof, are a better choice.

**Approvals as a formality.** If approvals are just clicked through, they are worthless. Plan realistic review time and measure how often approvals lead to corrections. If this figure falls permanently, that is a signal for a higher autonomy level, not for inattentive nodding through. Which metrics distinguish the two and how to prevent automation bias is shown in the article on [approval gates and human in the loop](https://andreasindorf.com/en/human-in-the-loop-ai-agents/).

**Missing business ownership.** An AI specialist without a responsible person in the business unit becomes a black box. IT runs the technology, but the business unit is responsible for the results.

**No baseline measurement.** Without figures from before the start, you cannot show later what the AI department has achieved. Lead times and error rates should therefore be part of the org chart analysis. A methodological framework for risk management is the voluntary [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework).

## From my own products

I first built the AI department for my own company: built AI-first since 2021, today run roughly 80 % by AI agents. This gave rise to the company cockpit: 51 AI specialists by job profile in 11 categories, four autonomy levels, a kill switch and an AI Act log for every action. The team, customer, service provider and onboarding portals run on a shared data foundation. I set the most important rule early on and have never changed it: an approval gate is always held by a human. Because an operating model is only as good as its reliability, the entire demo path runs as an automated Playwright end-to-end test. At a mid-sized logistics company, an AI system for route optimisation and warehouse logistics delivered a 40 percent efficiency gain. Effects like these do not come from the model alone, but from clear roles and clean approvals.

If you want to find out which roles in your org chart are suitable first, you will find the process of the org chart analysis under [AI department](https://andreasindorf.com/en/services/ai-department/).

## Frequently asked questions

### What is an AI department?

An AI department is an operating model in which individual roles in the org chart are staffed fully or partly with AI specialists. The specialists work according to real job profiles and are steered by your team in a cockpit. Critical decisions are always approved by a human.

### Does an AI department replace employees?

That is not the goal. In practice, AI specialists mainly take over the repetitive parts of a role, while people steer, review, resolve exceptions and carry responsibility. What this means for positions and tasks should be discussed early and openly with employees and the works council.

### How long does the introduction take?

The org chart analysis with prioritised roles, quick wins and a roadmap is delivered after two weeks. The first AI specialists are typically working in the cockpit after around six weeks. After that, the department is expanded step by step.

### How do you prevent the AI from making wrong decisions?

Through three mechanisms: autonomy levels that define for each task how much the AI may decide on its own, approval gates before every action that is hard to reverse, and a kill switch that stops any specialist immediately. Every action is also logged, so errors can be traced.

### Is an AI department compatible with the EU AI Act?

Yes, the model is designed for it. Human approvals, a kill switch and a complete log reflect the core ideas of the EU AI Act on human oversight and traceability. Which specific obligations apply depends on the area of use, especially if roles fall into high-risk areas such as HR.