AI-supported process automationHand off recurring tasks, keep control

Staff transfer information by hand, sort inboxes, compile reports or handle the same cases again and again? With AI-supported automation, a workflow takes over reading, classifying and preparing. Anything binding waits for your approval.

Book a first callSee examples
1 process
to start, in parallel operation
4 steps
from workflow to running automation
1 approval
before every binding change
0 €
first call, about thirty minutes
A lever engaging a gear, symbol of automated processes
The problem

The work is not hard. It is just always the same.

Classic automation fails at unstructured input: free-text emails, PDFs, attachments. That is exactly where language models come in. Four patterns I keep seeing in companies.

01

Retyping and transferring

Orders, invoices and enquiries arrive as email or PDF and are transferred by hand into ERP, CRM or spreadsheets. Error-prone and tedious.

02

Inboxes as queues

Shared inboxes are reviewed, sorted and distributed in the morning. Urgent items wait behind unimportant ones because nobody looks in earlier.

03

Reports at month end

Gathering figures from several systems, commenting, formatting. Two days, every month, under time pressure.

04

Rules are not enough

Classic workflows need structured fields. As soon as free text comes into play, automation ends and a person takes over again.

Typical situation

Does this sound familiar?

An overflowing inbox tray with one envelope guided along a light line into three sorting trays
AI-generated
  1. An employee spends half the morning assigning emails from the shared inbox to the right colleagues.
  2. Incoming invoices are printed, retyped and matched against orders.
  3. The monthly report is copy and paste from four systems plus a comment that sounds similar every month.
  4. Customer enquiries are answered only after days because master data and history have to be gathered first.
  5. There is an automation solution, but it breaks as soon as an email is worded differently than expected.
  6. The department has a list of twenty ideas and does not know which one pays off.
The solution

Workflows that can read

AI process automation combines a workflow platform with a language model. The workflow picks up the input, for example an email with an attachment. The model reads it, extracts the relevant fields (customer number, case type, quantity, date), classifies the case and rates urgency. Then the workflow takes over again: create the record, notify colleagues, file the case.

The difference from classic automation: the system copes with free text, varied wording and attachments. The difference from a mere chatbot: it works in the background, connected to your systems, following fixed rules.

The approval step is decisive. Everything that reads, sorts and prepares, the workflow may do itself. Everything binding, postings, replies to customers, changes to orders, waits for a person. We define that line together and only move it once trust has been earned.

The result

How a case runs through the workflow

Reading, classifying and preparing is done by the workflow with the language model. Everything binding waits for a person. Tap a step to see rule and detail.

Example: order by email, anonymised
Workflow and AI
Person
What happens

Customer number, items, quantity, requested date from text and attachment.

Rule

Language model reads, operated locally.

How it works

From workflow to running automation

1
Map the workflow

We map the process as it runs today: input, steps, handovers, exceptions. That usually makes clear which step is really worth it.

2
Define rules and limits

What may the workflow do itself, what waits for approval, what goes straight to a person? Confidence thresholds and exceptions are defined.

3
Build and test with real cases

The workflow first runs in parallel to the existing process. Every decision is logged and compared with the colleagues' result.

4
Take over and extend

After the parallel run, the workflow takes over. Approval steps are relaxed only if the hit rate justifies it. Further processes follow the same pattern.

Business benefit

What automation with AI changes

  • Less manual effort

    Reading, classifying, transferring and preparing run in the background. Staff review and decide instead of retyping.

  • Fewer transcription errors

    Fields are pulled from the original, not retyped. Deviations are flagged instead of overlooked.

  • Clear priorities

    Urgent items are on top because the system has rated every input before a person sees it.

  • Higher process quality

    Every step is logged. You see what the workflow decided and why, and can adjust.

  • Faster external response

    Enquiries, invoices and complaints are prepared before colleagues are at their desks. Replies go out earlier.

  • Control stays with people

    Binding steps wait for approval. You decide how much you leave to the system and can change that at any time.

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

Book a first call
Examples

Three typical automations

Anonymised scenarios from order processing, controlling and sales.

Read the inbox, classify and hand over to ERP and CRM
AI-generated
Order processing

Read the inbox, classify and hand over to ERP and CRM

Starting point
Orders, complaints and shipping notices arrive as free-text emails with attachments. Staff read, retype and create cases in ERP and CRM by hand.
Solution
A language model reads the message, extracts customer number, case type, quantity and date and classifies the case. A workflow creates the record. Anything binding waits for approval.
Possible benefit
Less manual entry, fewer transcription errors, clear priorities in the inbox. Control stays with the staff.
Possible technology
n8n, local language model for extraction and classification, ERP and CRM APIs
Compile recurring reports from several sources
AI-generated
Controlling and management

Compile recurring reports from several sources

Starting point
Every month, figures from ERP, spreadsheets and CRM are gathered, commented and turned into a presentation. Two days of work, often under time pressure at month end.
Solution
A workflow pulls the data together automatically, and a language model drafts the notable deviations. The professional judgement stays with the controller.
Possible benefit
The report is ready as a draft on the first working day. The time saved goes into analysis instead of collection.
Possible technology
n8n, database and spreadsheet connectors, local language model
Pre-qualify enquiries and prepare quotation drafts
AI-generated
Sales

Pre-qualify enquiries and prepare quotation drafts

Starting point
Enquiries from website, email and phone land unsorted in sales. For every quotation, master data, previous quotes and price lists are gathered by hand.
Solution
Incoming enquiries are captured in a structured way, checked against existing customer data and prepared as a quotation draft with matching text blocks and price references.
Possible benefit
Faster response time, consistent quotations and more time for the actual customer conversation.
Possible technology
n8n, CRM API, RAG on quotation history, local language model
See all use cases
Technology

What I build automations with

A workflow platform, a language model and your interfaces. The stack runs locally or in a European environment, depending on data class.

  • n8nWorkflow automation: picks up input, controls the steps, talks to your systems. Open source, self-hostable.
  • Local language modelReads free text, extracts fields, classifies and rates. Via Ollama on your hardware.
  • Confidence and rulesEvery model output carries a confidence. Below the threshold the case goes to a person.
  • ERP, CRM and email APIsConnection to the systems where the case ends up: SAP, Dynamics, HubSpot, Exchange or your in-house system.
  • Document extractionPDF and image processing for invoices, delivery notes and attachments, including text recognition.
  • Log and approvalEvery step is stored. Approvals run via email, Teams or a simple interface.

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

Frequently asked

What companies ask about AI automation

Which processes are suitable for AI automation?

Recurring workflows with unstructured input: inboxes, invoices, enquiries, reports, data transfer between systems. Well suited is what occurs often, follows clear rules and currently fails at free text.

What is the difference from classic process automation or RPA?

Classic automation needs structured fields and fixed sequences. With a language model, the workflow can read, classify and rate free text. Both complement each other: the model understands, the workflow acts.

What is the difference from AI agents?

A workflow follows a fixed sequence we define beforehand. An agent decides itself which steps to take in which order. For most processes in mid-sized companies the fixed workflow is the better entry. Agents come in when workflows are too variable for fixed rules.

How reliable is the language model?

For extraction and classification of typical business cases, good models score high, but not one hundred percent. That is why we work with confidence thresholds: uncertain cases go to a person. We measure the rate in the parallel run.

What happens with errors?

Binding steps wait for approval, so an error becomes visible before it takes effect. Every decision is logged and traceable. Errors turn into rules or examples that improve the system.

Does our data have to go to the cloud?

No. Workflow platform and language model run on your hardware or in a European environment. For processes with customer or contract data I recommend local operation.

How long until the first automation?

A first workflow, for example for a shared inbox, is ready in a few weeks including the parallel run. Connections to ERP and CRM drive the effort more than the AI itself.

What does AI process automation cost?

The effort depends on process scope, interfaces and operating path. The entry with one process is deliberately small. In the first call I estimate with you which process pays off first and what it costs.

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.

What I have learned about new technologies from 30 years of transformation practice

Technologies change quickly. Organisations learn slowly. A personal reflection on patterns that repeat with each technology wave.

Next step

Analyse automation potential

Bring a process that costs you time every day. In the first call we map it and clarify whether and how it can be automated with AI.

Book a first callContinue to AI agents

Related solutions

Enterprise knowledge with AI

Find company knowledge faster

AI assessment and use-case workshop

Find the right AI use case

AI agents for business

Hand multi-step workflows to AI