Solution

Process automation: what someone re-types today, a bot handles.

Read orders from emails, enter them into the business application, post them to a web portal: we automate recurring work between programs, even where there is no interface. Rules and existing interfaces first – AI only where genuine free text calls for it.

The demo: from inbox to web portal

Proof of Concept · Demo · 2:25

Proof of concept with a mock freight forwarding application and freight exchange; the emails are test data. Demo with German user interface, no sound.

What happens in the video

The example is a freight forwarder, but the pattern fits anywhere data travels from emails into programs and portals. Click a timestamp to jump to that point in the video.

  1. Read and triage the inbox

    The bot reads the Outlook inbox. Excel attachments and structured emails are parsed locally by fixed rules – no AI, no cost. Only free-text emails that the local check flags as a likely order go to an AI (Claude API), which decides whether it really is an order and returns the data in structured form. A price enquiry is correctly recognised as “not an order”.

  2. Enter it in the business application

    A bot enters the orders through the user interface of a Windows desktop application, just as a person would. It then checks via SQL that exactly the right data was saved, and it prevents the same order from being created twice.

  3. Post to a web portal without an API

    The bot reads the captured orders via SQL and posts them to a freight exchange using browser automation: login, cookie banner, form, confirmation – including the conversion from kilograms to tonnes.

  4. The portal changes – the bot repairs itself

    After a portal update, every form field has a new label. The bot notices it can no longer find them, has the AI re-map the form once, and then carries on without AI.

Where automation pays off

Wherever someone regularly moves data from one program to the next – following fixed rules, in meaningful volume, with little room for judgement.

  • Order intake from email

    Take orders from emails and attachments into your ERP, inventory system or industry software.

  • Incoming invoices

    Extract invoice data, pre-fill it in your accounting software and queue it for approval.

  • Web portals without an API

    Enter data into customer, supplier or public-authority portals that offer no interface.

  • Syncing applications

    Move master data, statuses and dates between programs that don’t talk to each other.

Logistics, retail and wholesale, administration, order processing: the industry matters less than a process that repeats.

How we work

  1. Map the process

    We look at the process where it happens: which inputs, which programs, which exceptions. Along the way we find out which imports or database access already exist.

  2. Pilot one or two workflows

    We start by automating one or two processes that make a noticeable difference and see how they hold up in a pilot. Then you decide how far to take it – one step at a time.

  3. Run it with monitoring

    Every bot logs what it did. Unclear cases go to a manual review queue instead of being guessed, and AI calls run under a cost cap. We can take over day-to-day operations if you like.

Technology: robust over convenient

  • The most stable interface first

    Existing imports and read-only database access come first. We only automate the user interface where no interface exists. Writes always go through the application itself – never around its business logic.

  • AI where it helps, not everywhere

    Rules and local parsing come first. AI is only asked about genuine free text and for self-healing after portal changes. That keeps costs low – and emails that aren’t orders never leave your premises.

  • Built for operations

    Duplicate protection, a cross-check after every step, manual review for unclear cases and a hard cost cap on AI usage.

  • Tool-agnostic

    Microsoft Power Automate Desktop, Python with Playwright and pywinauto, or another RPA platform: we use what fits your environment and what your team can maintain later. On-premises operation is possible.

  • Designed by people who run systems

    Automating provisioning and operations has been one of our focus areas for years. To us, a bot is a system in production – not a script you start once and forget.

Measured in the demo run

Figures from the proof of concept with mock demo systems – not production numbers. In your environment they depend on the application, the portal and your email volume.

Entry in the desktop application
approx. 3 s per order
Entry in the web portal
approx. 2 s per order
AI cost per free-text email
approx. 1 cent
AI cost for Excel attachments and structured emails
0 cents (parsed locally)
Test against a real inbox
0 of 46 emails sent to the AI unnecessarily
Self-healing after the portal update
10 of 10 fields correctly re-mapped, one-off approx. 1 cent
AI cost of the entire demo run
approx. 3 cents

Frequently asked questions

Do we need an API?

No. If there is an import or read access to the database, we use it, because that is the most stable route. Where neither exists, the bot operates the user interface of the application or web portal the way a person would.

What happens when a portal changes?

The bot notices it can no longer find fields, has the AI re-map the form once, and then continues without AI. If a change cannot be mapped unambiguously, the case goes to manual review instead of being entered incorrectly.

What data goes to the AI?

Only free-text emails that the local check has already flagged as a likely order, and – after a portal change – the field labels of the form. Excel attachments, structured emails and anything that isn’t an order are processed locally and never leave your premises.

Is this GDPR-compliant – and can it run without cloud AI?

Data minimisation is built in: only free-text emails that are likely orders go to an AI at all; everything else is processed locally. If a cloud model is used, we work out the provider, data processing agreement and data location with you and your data protection officer. If the data must not leave your premises at all, a local language model (LLM) can run on your own hardware instead – provided you have a sufficiently powerful GPU. The process stays the same; whether a local model is accurate enough for your emails is something we test in the pilot.

Power Automate or Python?

Both work. Power Automate Desktop is the natural choice if you are already deep in Microsoft and your team prefers low-code; Python with Playwright and pywinauto gives more control over complex processes. The proof of concept in the video is built in Python. We decide together what stays maintainable for you.

How do we get started?

With a conversation about one specific process. That becomes a pilot with one or two workflows – and only once it works in everyday use do we expand.

Which process is still being re-typed at your company?

Describe one process to us – we’ll tell you honestly whether and how it can be automated.