From Manual Routines to Intelligent AI Workflows
A process from almost every company
An invoice arrives by email, usually as a PDF attachment. Someone opens it, reads out the invoice number, amount, tax rate and supplier and types the details into a system. Then it gets checked whether the amount matches the purchase order. After that the invoice goes to accounting, often as a forwarded mail with a short comment.
This routine takes maybe six minutes per invoice. At two hundred invoices a month that is twenty hours in which nobody decides anything and data merely moves from one place to another. Routines like this can largely be automated today without a major project.
What the same routine looks like automated
Automated, the process consists of a few clearly separated steps.
- The inbox is watched. As soon as a mail with an attachment lands in a specific mailbox, the flow starts.
- The relevant data is extracted. An AI model recognises invoice number, date, amount, tax rate and supplier, even when every document is laid out differently.
- The details are checked. Is there a matching purchase order, does the amount fit, is the supplier known, has this invoice number been processed before.
- A human takes a look. The team gets a card or a message with the recognised values and the original document next to them and confirms or corrects.
- The invoice is handed over. Once approved it moves into accounting with clean data, filed and logged in an auditable way.
The fourth step is the most important one. Automation does not mean nobody looks any more. It means people only decide and no longer retype. Six minutes of work turn into twenty seconds of review.
The tools for this
Automation platforms exist for exactly these routines. You assemble the flow graphically, step by step, while the platform takes care of triggers, retries on errors and the connections to the systems involved. Three of them are relevant for mid-sized companies.
n8n is our favourite for serious automation. You can host it yourself, so the data stays in house or on a server in Germany. Changes are versioned and traceable, so you can see who changed which step and when. Permissions can be controlled in detail, so accounting sees its flows and sales sees theirs. And when a step cannot be expressed in the platform, you can drop in your own code at that point instead of abandoning the whole flow.
Zapier has the largest catalogue of ready-made connections to well-known cloud services. When the job is to couple two widespread tools quickly, you are done in fifteen minutes. With branching logic and larger volumes of data, however, it gets confusing and expensive quickly.
Make sits in between. The presentation is very visual and more complex branching and data transformation can be modelled well. Like Zapier it runs purely as a cloud service, so the data leaves your own house.
Why the graphical view is worth more than it looks
The biggest advantage of these platforms is not the technology, it is the visibility. A flow is a chain of boxes with connecting lines. You can put it on the screen in a meeting and everyone in the room understands what happens.
You point at the first box and say, this is where the mail comes in. At the second, this is where an AI model decides whether it is an enquiry, a complaint or an invoice. At the third, this is where the lead is created or updated in the CRM. At the fourth, this is where the AI reads the quote value out of the PDF and routes anything above a certain sum straight to the head of sales. You can have that conversation with your accounting team, with sales, with the management. Nobody needs to be able to read code for it.
The second advantage follows from that. The business departments can contribute. Whoever runs the process every day knows the edge cases better than any developer. In a visible flow they can name them and often adjust them themselves. Small changes then need neither a ticket nor a meeting.
The third advantage is speed. A first working flow takes an afternoon, not a quarter. You see immediately whether the idea holds up and can then extend or drop it without having invested much.
Where AI genuinely helps in a flow
AI is not an end in itself in these workflows, it is a single step among many. It takes on the tasks that fixed rules were never good enough for.
Reading unstructured documents. Every invoice and every quote looks different and a model still recognises the values reliably.
Sorting things. Incoming mails are classified by topic, urgency and responsibility and land directly with the right team.
Summarising. A long mail thread turns into a note in the CRM that sales reads in ten seconds before the next call.
Preparing replies. A draft is created automatically, a human checks it and sends it.
It gets interesting once not every step is hard-wired any more. An AI agent is given a goal and a set of tools and decides for itself which steps to take in which order. It can look something up in the ERP, add a missing detail, ask a colleague and carry on afterwards. Models like Claude have become good enough for this to work in daily practice. For processes with many edge cases that is real progress.
Two things belong with it. An agent needs clear boundaries, meaning only the tools and permissions it actually requires. And anything that moves money or goes outside the company gets a human approval. That does not slow the process down, because the preparatory work is already done.
Where the platforms reach their limits
As good as these tools are, there are places where they cannot go any further.
A system has no usable interface. Older industry software in particular and applications running on your own network appear in no catalogue of ready-made connections.
The logic gets too complex. Pricing calculations with many dependencies, planning and optimisation tasks or analyses across large data volumes can be built as a chain of boxes, but they are hard to maintain and impossible to test properly.
The volume gets too large. Platforms charge per operation. What is cheap at two hundred invoices a month becomes a cost factor at two hundred thousand operations.
Particularly sensitive data or hard response times are involved. Or there is no real test environment for a flow the business depends on.
None of this is an argument against automation platforms. It is simply the point at which one single step has to be built differently.
This is exactly where custom software development starts
When a platform cannot go further at one point, the whole flow does not have to be replaced. It is enough to build that one piece yourself and connect it cleanly.
For example, we write a small dedicated API that runs in a Docker container and provides the connection to a system for which no ready-made building block exists. The workflow calls it like any other step and notices nothing of the complexity behind it.
We build a dedicated application or interface wherever staff work every day and a generic form is not good enough, for instance for the review and approval in the example above.
We implement complex calculations and algorithms as a service of their own, in Python or Go, with automated tests and versioning you can rely on.
And we take over steps that run at high volume, because a dedicated service often costs a fraction per operation.
The result is a division of labour that has proven itself in practice. The platform stays the visible backbone in which the process is comprehensible for everyone. The difficult pieces sit in dedicated services connected through webhooks and APIs. You keep the overview and the speed of the platform and still have no ceiling on what is technically possible.
What a sensible start looks like
Begin with a single process that occurs often, is clearly described and costs time today. Incoming invoices, quote requests, onboarding of new employees or maintaining leads in the CRM are good candidates.
Measure the duration of the routine and its frequency beforehand. Without those numbers the benefit of the automation stays a mere claim later on.
Build the first version deliberately with a human approval at the decisive point. Once it is visible after a few weeks that the recognition runs reliably, you can drop the approval for unambiguous cases and only have the exceptions presented to you.
And appoint someone responsible for the flow. Automated processes need maintenance, because systems, formats and responsibilities change. A flow nobody feels responsible for will fail at some point and nobody notices in time.
Conclusion
Automation platforms like n8n, Zapier and Make have lowered the barrier to process automation considerably. Routines become visible, business departments can have a say and first results appear in days rather than months. Together with AI models and agents, tasks can now be handed over that were still manual work two years ago.
The limit of these tools is real, but it is not an obstacle. It is the place where custom software development comes in. The actual value emerges from the combination, meaning a visible backbone for the process and tailored building blocks everywhere the standard is not enough.
Is one of your processes eating up too much time? Talk to us. We look at the routine and tell you honestly which part a platform can solve and which part needs software of its own. Then we build both.
