AI agents with clear permissionsAI agents that prepare cases and trigger nothing without you

Cases run across several systems and people, and every handover costs time and creates follow-up questions? An AI agent reads, checks, pulls data from ERP and CRM and presents a complete case for decision. You decide how many permissions it gets.

Book a first callTo the whitepaper
4 permission levels
read, prepare, change, trigger
13 questions
from the whitepaper as a checklist
1 approval
before anything binding
0 €
first call, about thirty minutes
Threads of light connecting stacks of documents under a magnifier, symbol of agents that read and check cases
The problem

One case, five systems, three handovers

Complaints, invoices, supplier notices: the effort rarely sits in one step but in the chain in between. Four patterns that agents address.

01

Every handover costs a day

Intake, review, forwarding, follow-up. A case that contains an hour of work needs a week of lead time.

02

Incomplete cases

The colleague at the next stage has to ask first because order number, delivery note or history are missing. The case goes back.

03

Fixed workflows are not enough

The sequence depends on the case: sometimes a look into the ERP is enough, sometimes three follow-ups are needed. Too variable for rigid rules.

04

Fear of losing control

A system that decides and acts on its own rightly causes concern. The question is not whether, but how narrowly the permissions are set.

Typical situation

Does this sound familiar?

Three glass panels in a row, a glowing sphere travelling along a light thread from station to station
AI-generated
  1. A complaint is taken by email, transferred into a form, matched with the order and passed on to quality assurance. Every stage waits for the previous one.
  2. Procurement learns about delivery date changes only when production asks.
  3. Incoming invoices with missing order numbers wander between accounting and the department for days.
  4. You have started with process automation, but the exceptions are more frequent than the standard case.
  5. A vendor promises autonomous agents, and nobody in-house can assess what they would actually be allowed to do.
  6. Management wants agents, the works council wants to know who decides in the end.
The solution

An agent is not a chat. It is a chain with handovers.

An AI agent receives a task, for example "Prepare this complaint for decision", and works through it in steps: read the message, pull order and delivery data from the ERP, check deadline and warranty status, assemble the case with a recommendation. Which steps it takes in which order, it decides based on the case. That distinguishes it from a fixed workflow.

The permission level is decisive. I work with four levels: read (view data), prepare (assemble drafts and cases), change (create or modify records) and trigger (postings, replies, orders). Most agents in mid-sized companies start at read and prepare. Everything binding waits for approval. An agent does not remove control, it moves it to the point where decisions are made.

The agent runs on a workflow platform with a local language model, connected to your systems. Every step is logged: what was read, what was decided, with what confidence. That is the basis for trust, improvement and evidence.

Permission levels

How much may the agent do?

Four levels from read to trigger. Most agents in mid-sized companies start at the first two; everything binding waits for approval.

Tap a level: what may the agent do there?
Level 2

Prepare

May
Compile cases, write drafts, formulate a recommendation, state confidence.
Approval
The finished case waits for a person's decision.
Example: complaint
Deadline and warranty checked, case presented with the recommendation of goodwill and a free spare part.
How it works

From workflow to agent in four steps

1
Map the workflow and handovers

We map the case with all systems, roles and exceptions. The thirteen questions from the whitepaper are the checklist.

2
Define permissions and limits

Which level does the agent get, what waits for approval, where does it stop and hand over to a person? This is written down.

3
Build and test in parallel

The agent processes real cases, its proposals are compared with the colleagues' decisions. Confidence and hit rate are measured.

4
Take over and extend permissions

After a proven parallel run, the agent takes over the preparation. Higher permissions only when the numbers justify it, and always revocable.

Business benefit

What agents change in daily work

  • Shorter lead time

    The chain of intake, review and assembly runs in minutes instead of days. People decide instead of collecting.

  • Complete cases

    Order data, history and evidence are there at first contact. Follow-ups and returns decrease.

  • Early visibility

    Deviations, deadlines and risks are reported before they become a problem. You act instead of reacting.

  • Control stays defined

    Four permission levels, written down, adjustable at any time. Works council and data protection know who decides.

  • Traceable decisions

    Every step is logged. You see why the agent prepared a case the way it did and can correct it.

  • Scalable without new positions

    As volume grows, the agent works along. Staff concentrate on the cases that need judgement.

Recognise your situation? In a free first call we check whether this route fits you.

Book a first call
Examples

Three agents with narrow permissions

Anonymised scenarios from customer service, procurement and accounting. All three start at read and prepare.

Record complaints, check them and present them for decision
AI-generated
Customer service and quality

Record complaints, check them and present them for decision

Starting point
Complaints arrive by email and phone, are transferred into a form, matched against order data and passed on to quality assurance. Every handover costs a day.
Solution
An agent reads the complaint, pulls order and delivery records from the ERP, checks deadlines and warranty status and presents a prepared case with a recommendation. A person decides.
Possible benefit
Shorter lead time, complete cases on first contact with quality assurance, a traceable log of every step.
Possible technology
n8n as orchestration, local language model, ERP API, approval step before every change
Monitor supplier notices and flag the need for action
AI-generated
Procurement

Monitor supplier notices and flag the need for action

Starting point
Delivery date changes, price adjustments and quality notices from suppliers arrive via portals and emails. Procurement often notices deviations only when production asks.
Solution
An agent collects notices, compares them with orders and requirements and raises alerts by urgency. Read and prepare only, no independent changes to orders.
Possible benefit
Deviations become visible early, procurement acts instead of reacting. The agent's permissions stay deliberately narrow.
Possible technology
n8n, portal and email integration, local language model, ERP read access
Check incoming invoices and prepare them for approval
AI-generated
Accounting

Check incoming invoices and prepare them for approval

Starting point
Invoices arrive as PDFs by email, are printed or retyped, matched with order and delivery note and forwarded for approval. Missing order numbers cause follow-up questions.
Solution
An agent extracts the invoice data, matches it with orders, detects deviations and presents the invoice with a verification note for approval.
Possible benefit
Less manual entry, faster approvals and documented verification steps. The approval itself stays with the responsible people.
Possible technology
n8n, document extraction with a local language model, ERP API
See all use cases
Technology

What I build agents with

An orchestration, a language model, your interfaces and a log. No black box, every component replaceable.

  • n8nOrchestration of steps, connection to systems, approvals and exceptions. Open source, self-hostable.
  • Local language modelReads, extracts, classifies, rates and formulates the recommendation. Via Ollama on your hardware.
  • Tools and APIsRead and write access to ERP, CRM, email, file storage, each limited to what is necessary.
  • Permission and approval logicFour levels, confidence thresholds, stop rules. Approvals via email, Teams or interface.
  • LogEvery step stored with input, decision and confidence. Basis for control and evidence.
  • Knowledge base (optional)RAG on policies, contracts or product knowledge so the agent decides by your rules.

I build, test and implement. The technology is the last step, not the first.

Frequently asked

What companies ask about AI agents

What is an AI agent?

A system that works through a task in several steps, uses tools along the way (databases, interfaces, documents) and decides itself which step makes sense next. The difference from a chat: it acts instead of just answering. The difference from a workflow: it does not follow a fixed sequence.

When do AI agents pay off in mid-sized companies?

When a case occurs often, runs across several systems and the steps depend on the case. Complaints, invoice verification, supplier notices and enquiries are typical candidates. Not worth it: rare cases, cases with a high share of judgement and workflows a fixed process solves just as well.

How much may an agent decide?

As much as you give it. I work with four permission levels: read, prepare, change, trigger. Most agents start with the first two. Higher levels come only after measurable reliability and remain revocable.

What happens if the agent makes a mistake?

At read and prepare the mistake becomes visible before it takes effect because a person approves. At higher levels, confidence thresholds and stop rules apply. Every step is logged and traceable.

What does an AI agent cost?

The effort lies less in the language model than in process mapping, interfaces and the permission concept. A first agent with narrow permissions is a project of a few weeks. Concrete figures emerge in the first call once case and systems are clear.

Do our systems need to be changed?

Usually not. The agent uses existing interfaces of ERP, CRM and email. Where interfaces are missing, there are interim paths such as exports or email handovers that are enough for a start.

Does an agent fall under the EU AI Act?

Agents that prepare internal cases are usually not high-risk systems. If the agent touches HR decisions, credit decisions or safety, that looks different. We check this before the start based on your case.

What is the best way to start?

With the whitepaper: thirteen questions to check whether a case is ready for an agent. Then a first call about the case that costs the most time.

Read more in the knowledge area

What are AI agents and when do they pay off?

What AI agents are, how they differ from chatbots and workflows, when they pay off in mid-sized companies, which permissions they should get and what you need to clarify beforehand. With example and FAQ.

Which AI trends really matter for the mid-market in 2026?

Not every AI trend matters for every mid-market company. A prioritisation of what counts now, what should be watched, and what can wait.

Next step

Check a case for an agent

Download the whitepaper (10 pages, no form) and check your case with the thirteen questions. Or we go through it together in the first call.

Book a first callDownload the whitepaper

Related solutions

AI process automation

Automate manual processes

Local AI for business

Use AI with sensitive data

AI assessment and use-case workshop

Find the right AI use case