AI Workflows

Large companies have been working this way for a while. Now it’s the Mittelstand’s turn.

We connect your mailbox, ERP, accounting and company knowledge to AI assistants that prepare orders, enquiries and invoices. Your team reviews, corrects and decides, and gets more orders done in the same time.

Try the demo Free initial consultation

Not a vision of the future

At banks, insurers and industrial groups, AI assistants already prepare enquiries, orders and documents today. We have built such systems there ourselves and put them into production (references). What used to need dedicated IT departments and large budgets is now possible with open standards and off-the-shelf AI models, for businesses with 10 or 200 employees.

More orders, same team

Once the searching and gathering goes away, every employee handles more cases per day. You grow without having to hire for every additional order.

Faster to the customer

Quotes, order confirmations and replies go out while the customer is still waiting. Whoever replies first and properly often wins the order.

Less slips through

The assistant checks everything, every time: previous order, stock, contract, prices. Discrepancies show up before they reach the customer or accounting.

Knowledge stays in the business

How your team handles a case is written down in the work instructions. New colleagues get up to speed faster, and nothing is lost when someone retires.

Connected, not scattered

Almost every application now comes with its own AI. But each one only sees its own slice. What makes the difference is how well your systems are connected.

AI as an island

The ERP has a chatbot, the mailbox a writing assistant, the document archive a search. A person still has to piece the case together, copying and pasting between five windows.

AI as a colleague

An assistant reads the e-mail, finds the case in the ERP, checks stock, contract and prices, and presents a finished proposal. Your team reviews, corrects and approves.

Try it yourself

Three typical cases from everyday office work. Pick one and send the request to the assistant. You can change its proposal before approving it. Nothing is created or sent until then.

AI workspace
Inbox
From Date

The workspace is a mock-up; companies, names and numbers are made up. In a real setup, the grey lines are queries to your systems.

Who does what

The assistant takes care of searching and cross-checking. Your team steps in where needed: correcting quantities, changing texts, adding what the assistant cannot know. Wherever something becomes binding, a person decides.

  1. Assistant

    Read input

    E-mails and attachments, including scanned documents

  2. Assistant

    Match

    Customer, supplier, case, contract

  3. Assistant

    Cross-check

    Stock, prices, dates, terms

  4. Assistant

    Propose

    Result and open points, each with its source

  5. Person

    Review & step in

    Correct, add, reject or approve

  6. Assistant

    Execute

    Create order, ticket or document, reply as draft

  7. Person

    Complete

    Send to customer, book, pay

Assistant works on its own, read-only or as a draftYour team reviews, corrects and decides

Corrections are not lost: what your team changes often goes into the work instructions and, as a test case, into the test set.

Workflows that pay off

Good candidates are cases that come up often, follow fixed rules and currently involve a lot of searching. A few examples:

Order intake

Input
Order by e-mail or PDF
Assistant
finds the previous order, checks items, stock and prices
You
approve the order

Customer enquiries

Input
Question about a contract, product or appointment
Assistant
replies from contracts, manuals and calendar, with source
You
send the reply

Supplier invoices

Input
Supplier invoice by e-mail
Assistant
checks against purchase order and goods receipt, proposes posting
You
book and pay as usual

Delivery status

Input
“Where is my delivery?”
Assistant
looks up order, production and shipping, drafts a reply
You
send the reply

Quotes

Input
Request with bill of materials or specification
Assistant
finds similar quotes, proposes line items and prices
You
finalise the costing and send

Checks against requirements

Input
Specification, contract or technical enquiry
Assistant
checks against standards, guidelines and your own rules, with source
You
assess the result

Four building blocks

Every workflow is made of the same four parts. None of it is secret knowledge: you can read, understand and shape them.

Interface (MCP)

The Model Context Protocol is an open standard through which an assistant reads data and triggers actions. Built once, an interface works with many user interfaces and AI models.

find_customer · load_case · draft_order

Skill

A work instruction for the assistant, written as if for a new colleague. Your team can read and change it.

“Never silently replace discontinued items; flag the successor instead.”

Agent

An assistant for one specific task. It knows its skills and only has the tools it needs for them.

Order intake: reads mailbox, CRM, ERP; only creates drafts

Harness

The framework around it: permissions, approvals, logging and test set. It keeps the assistant under control.

create_order → only after approval

Interfaces by Entropy

We build and run MCP interfaces to your systems: on your server, in a German cloud or with us. They do not belong to a single workflow or to a particular AI vendor. This is the part that stays.

ready to use

Knowledge base: EntropyZero

Manuals, contracts, standards and guidelines, searchable for every assistant. Answers always come with document and page.

More about EntropyZero →
for your systems

Mailbox

Read e-mails and attachments, create drafts. Microsoft 365, Exchange or IMAP.

for your systems

ERP & inventory

Read orders, items, stock and purchase orders. Creating only as a draft or after approval.

for your systems

CRM & tickets

Customers, contacts, contracts and the history of a case.

for your systems

Accounting

Hand over documents with a posting proposal, for example to DATEV.

for your systems

Document storage

Archive, network drive or Paperless-ngx: find, file and tag documents.

How a workflow comes about

This is how we already work in software development: the workflow is described, connected to the systems, tested with real cases and then continuously improved. The same approach fits your office workflows.

  1. Write down the workflow

    With the colleagues who handle the case today: How do you recognise it? Where do you look things up? What exceptions are there? This becomes the work instruction for the assistant.

  2. Connect the systems

    Each system involved gets an interface, read-only at first. The assistant sees what your team sees, and nothing beyond that.

  3. Set the rules

    What may it do on its own, where does someone have to approve, what gets logged? These rules are built into the technology and cannot be argued away.

  4. Test with real cases

    A set of cases from last year, each with the result you expect. Every change has to pass this test set.

  5. Pilot in daily work

    Two or three people work with it, every case with approval. What stands out goes into the work instruction.

  6. Expand

    Once the results are right, the next workflow follows. It builds on everything that is already there.

Start without risk: the analysis workshop and the first prototype are free. You only pay for the implementation once you are convinced. Interfaces, work instructions and code belong to you.

Controllable and verifiable

An assistant that works in your day-to-day business has to be accountable like a colleague. Three things make sure of that.

Approval at fixed points

The assistant may read on its own. Anything that creates something in the ERP, goes to customers or moves money needs a click from you.

Fully logged

Which e-mail, which case, which source, who approved: everything can be traced later.

Tested with real cases

Before a changed work instruction or a different AI model goes live, it runs against the test set.

Example: test run after a change to the skill “Repeat order from e-mail”37 of 40 passed
Test caseExpectedResult
“same as in March” without order numberCorrect previous order foundpassed
Item since replaced by a successorFlag instead of silent substitutionpassed
Customer with framework agreement and special pricesPrice from the framework agreementpassed
Order as scanned PDF, 6 line itemsAll line items recognised5 of 6

The second workflow is faster

Interfaces, work instructions and rules do not belong to a single workflow. What was built for the first one is available to every further one. What is new is mainly the domain instruction and the test cases.

Stage 1

Repeat orders

Find the previous order, cross-check changes in items, stock and prices, prepare the order for approval.

9 new building blocks
Stage 2

Customer enquiries

Answer questions about contracts and appointments from the knowledge base, with source. Mailbox, CRM and draft replies are already there.

4 new building blocks
Stage 3

Supplier invoices

Check invoices against purchase order and goods receipt, pass the document to accounting with a proposal. At its core, only the DATEV connection is new.

3 new building blocks
Building blockRepeat ordersCustomer enquiriesSupplier invoices
Interfaces
Mailboxnewexistingexisting
CRMnewexisting–
ERP & warehousenew–existing
Knowledge base (EntropyZero)–new–
Tickets & calendar–new–
Accounting (DATEV)––new
Skills
Match case (customer, supplier, order)newexistingexisting
Detect discrepanciesnew–existing
Reply as draftnewexistingexisting
Back statements with sources–new–
Propose posting––new
Harness
Permissions, approvals, loggingnewexistingexisting
Test procedure for the test setnewexistingexisting
Test cases from real casesnewnewnew
To be built943

Every new building block is then also available to the earlier workflows. The knowledge base from stage 2, for example, can later back technical details in repeat orders with sources. The order of the stages follows your priorities.

Open on interface and model

Interfaces and work instructions stay the same, whichever user interface and AI model you end up using. We choose together, based on your requirements for data protection and cost.

User interface

ChatGPT BusinessClaudeLibreChatOpen WebUIGoose

From what you already use, to self-hosted web interfaces, to desktop assistants.

AI model

OpenAIAnthropic ClaudeMistralself-hosted

Chosen per task and replaceable without rebuilding interfaces.

Operations

Your serverGerman cloudAWS / Azure (EU)

On request we take over day-to-day operations, including updates and test runs.

Frequently asked questions

Does the assistant replace employees?

No. It takes the searching and cross-checking off your team’s plate. Decisions, approvals and contact with the outside world stay with your people. They handle more cases in the same time, with fewer errors.

What may the assistant do without approval?

You decide. Typically: reading and creating drafts on its own. Anything binding, such as creating orders, sending e-mails to customers or handing over documents, only after a click from you. These rules are enforced technically, not just agreed.

Do we need new software?

Usually not. We connect the systems you already have. As the user interface, an AI tool you already use or a lean web interface on your server is often enough.

What happens to our data?

We sign a GDPR-compliant data processing agreement. If you wish, interfaces and models run entirely within your own infrastructure. For cloud solutions we only use providers with data centres in Germany or the EU.

What if we want to switch AI models later?

Then you switch. Interfaces and work instructions are independent of the model. Before switching, the new model runs against your test set, so you can see whether it performs at least as well.

Where do we start?

With a workflow that comes up often and can be described well. In the free workshop we pick it together and look at which systems are involved.

Which workflow costs you the most time?

Tell us about it. In a free initial consultation we will tell you honestly whether and how AI can solve it.

Get in touch now

Contact

Write to us — we will reply within one business day.