For industry and medium-sized companies

AI agents forinternal workflows

They search approved sources, prepare work steps and provide traceable results. Changes only happen after human approval.

Schmidt AI workshop on transformer models and AI agents
  1. 01Approved data
  2. 02Agent and tools
  3. 03Human approval
  4. 04Verifiable result

Suitable tasks

When an AI agent is appropriate

An agent is a good fit when information is spread across several systems and the next useful step depends on the current result.

If the sequence, inputs and exceptions are fully known, a fixed workflow is usually easier to test and operate.

01

Internal knowledge search

The agent searches approved documents, wikis and tickets and supports answers with sources.

02

Prepare quotes and emails

Drafts use previous quotes, style guidelines and the current case. A person reviews and sends them.

03

Processes with approvals

The agent researches and prepares work in permitted systems. Effective changes require approval.

Controlled pilot

From use case to limited pilot project

The first version reads and evaluates but does not change production data. Extensions follow only after a traceable evaluation.

  1. 01

    Define the task and success criteria

    A limited use case and real completed cases define what a useful result looks like.

  2. 02

    Limit data and permissions

    Sources, user roles, permitted tools and approval points are defined before technical implementation.

  3. 03

    Test in read-only mode

    The first version searches, compares and prepares drafts without changing production data.

  4. 04

    Evaluate and decide

    A fixed test set evaluates quality, sources and failure cases and provides the basis for the next step.

Pilot boundaries

Define access first

Data, tools and room for action are defined before implementation.

Security and operation

Data access

Only approved sources and roles are connected.

Effective actions

Changes and external actions require human approval.

Resource limits

Runtime, cost and tool calls have fixed limits.

Traceability

Inputs, actions and results are logged and tested.

Foundations and our own practice

How we assess and test agent systems

AI Mux coordinating several coding agents in one interface
Our own development

AI Mux for coding agents

Our development log shows how we coordinate several agents and make requests for human attention visible.

Read development log
Foundations

What is agentic AI?

How agents choose tools, when a fixed workflow is the better fit and which limits are needed before use.

Read article
Project assessment

AI readiness in industry

A practical assessment of process, data, integration, risks and later operation.

Read the readiness check

Assess an AI use case

We clarify whether an agent is appropriate, which data it needs and how a limited pilot can be evaluated.

Gießen office: Flutgraben 4, 35390 Gießen
Visits by appointment, Mon-Fri, 09:00-18:00
Registered office and postal address: Auf der Grube 9, 35041 Marburg
Carsten Schmidt

Carsten Schmidt

CEO